CN109685756A

CN109685756A - Image feature automatic identifier, system and method

Info

Publication number: CN109685756A
Application number: CN201710970382.4A
Authority: CN
Inventors: 黄哲瑄; 黄玺轩; 张书修
Original assignee: Leda Creative Technology Co Ltd
Current assignee: Leda Creative Technology Co Ltd
Priority date: 2017-10-16
Filing date: 2017-10-16
Publication date: 2019-04-26

Abstract

The present invention relates to a kind of image feature automatic identifiers, system and method.The image feature automatic identification method includes the following steps: to carry out image processing program to the multiple sample images for being respectively provided with different image features, and respectively overlaps processing result image caused by image processing program and standard picture to generate multiple sample images of amplification；The multiple sample images for expanding forward and backward are supplied to deep learning system, to establish the automatic identification algorithm of image feature comprising there is the appreciative standard for image feature；The testing image of object to be detected is captured, testing image is analyzed with image feature automatic identification algorithm, and judges whether testing image has image feature according to appreciative standard.Effectively improve processing procedure with image feature automatic identification technique and reduce flaw, saves a large amount of time and cost.

Description

Image feature automatic identifier, system and method

Technical field

The present invention relates to a kind of image feature automatic identifiers, system and method, will be used for depth more particularly to a kind of Image feature automatic identifier, the system and method that the sample image of degree learning system is first expanded.

Background technique

As machine learning (Machine Learning) technology reaches its maturity, no matter in image identification, speech recognition Or the various aspects such as natural language processing, there is more diversified application.

Wherein, the combination of image identification technology and machine learning techniques is especially outstanding, in addition to distinguishing in basic handwriting Except the application such as knowledge, Object identifying and human face recognition, the skill of deep learning (Deep Learning) and image identification integration Art also combines automatic optics inspection (Automated Optical Inspection, abbreviation AOI) system often, and is applied Product quality management control in process of producing product.

Machine learning algorithm is that a kind of automatically analyze from data obtains rule, and assimilated equations carry out in advance unknown data The algorithm of survey.It will be apparent that in order to ensure the algorithm come out produced by deep learning can correctly differentiate unknown data, it is necessary to A large amount of sample data is provided in the training process (Training) of deep learning especially largely has the sample of label (Label) Product data can sufficiently and correctly learn the key to differentiation so as to the model (model) of deep learning.

However, the collection not a duck soup of a large amount of samples, especially when deep learning system is used in process of producing product In quality control, the label of Learning Identification is usually wanted to be the unwanted visual characteristic of product, if require the producer elder generation output of product Extremely large amount of flaw product, and after hand picking goes out a large amount of flaw sample and is marked, it can effectively apply depth Learning system assists to generate the algorithm that can recognize unwanted visual characteristic automatically, and then the algorithm by recognizing unwanted visual characteristic automatically, returns Head come assist the producer carry out quality control and improve processing procedure, this is not obviously an ideal scheme.

Summary of the invention

Technical problem to be solved by the present invention lies in provide a kind of image feature in view of the deficiencies of the prior art and distinguish automatically Identification device, system and method can pass through the quantity of massive amplification sample image, effectively promote training under deep learning system The data diversity that Cheng Suoneng is touched, so can promote efficiency that deep learning system is recognized automatically applied to image feature with Correctness.

In order to solve the above technical problems, it is special to be to provide a kind of image for a wherein technical solution of the present invention Levy automatic identifier comprising a storage element, a processing unit and an image acquisition unit, wherein the storage is single Member stores a database, and the database stores one first image feature classification group and an at least standard picture, institute It states the first image feature classification group and stores multiple first sample images, each described first sample image is respectively provided with not The first same image feature；The processing unit is connect with the storage element signal；The image acquisition unit and the place Cell signal connection is managed, with the testing image for capturing an object to be detected, the processing unit is read in the database Multiple first sample images, multiple first sample images carry out image processing programs to generate multiple images respectively Processing result, and multiple described image processing results are overlapped with multiple standard pictures respectively, to generate amplification respectively Multiple first sample images, the processing unit executes according to multiple first sample images before amplification and after amplification The training program of one deep learning system, to establish the automatic identification algorithm of an image feature, the image feature recognizes calculation automatically Method includes one first appreciative standard for first image feature, and the processing unit takes from the image acquisition unit The testing image is obtained, the processing unit executes the automatic identification algorithm of image feature and divides the testing image Analysis, and the processing unit judges whether the testing image has first image special according to first appreciative standard Sign.

In order to solve the above technical problems, it is special to be to provide a kind of image for an other technical solution of the present invention Levy automatic identification system comprising a server-side and a test side, wherein the server-side includes a storage element and one Processing unit, the storage element store a database, the database store one first image feature classification group with And an at least standard picture, the first image feature classification group store multiple first sample images, each described A sample image is respectively provided with the first different image features；The processing unit is connect with the storage element signal.Wherein, The processing unit reads multiple first sample images in the database, and carries out an image processing program respectively, And processing result image caused by described image processing routine is respectively overlapped with the standard picture to generate respectively Multiple first sample images of amplification；The processing unit is executed according to multiple first sample images before amplification and after amplification The training program of one deep learning system, to establish the automatic identification algorithm of an image feature, the image feature recognizes calculation automatically Method includes one first appreciative standard for first image feature.The test side is connect with the servo end signal, And the automatic identification algorithm of image feature can be received by the server-side, the test side include an image acquisition module and One processing module, a testing image of the image acquisition module to capture an object to be detected；The processing module with it is described The connection of image acquisition module signal, to obtain the testing image from the image acquisition module, the processing module executes institute The automatic identification algorithm of image feature is stated to analyze the testing image, and according to first appreciative standard judgement it is described to Whether altimetric image has first image feature.

In order to solve the above technical problems, it is special to be to provide a kind of image for an other technical solution of the present invention Levy automatic identification method comprising the following steps: to multiple first sample images for being respectively provided with the first different image features Carry out an image processing program, and respectively by processing result image caused by described image processing routine with do not have described the One standard picture of one image feature is overlapped to generate multiple first sample images of amplification；Before expanding and after amplification Multiple first sample images be supplied to a deep learning system, to establish the automatic identification algorithm of an image feature, the image The automatic identification algorithm of feature includes one first appreciative standard for first image feature；Capture the one of an object to be detected Testing image analyzes the testing image with the automatic identification algorithm of the image feature, and according to first identification Standard judges whether the testing image has first image feature.

The beneficial effects of the present invention are image feature automatic identifiers, system provided by technical solution of the present invention And method, can by " to be respectively provided with different images feature multiple sample images carry out image processing program, and with do not have There is the standard picture of the image feature to be overlapped to generate multiple sample images of amplification " and " before expanding and expand Multiple sample images after increasing are supplied to deep learning system " technical characteristic, to promote training under deep learning system The data diversity that Cheng Suoneng is touched, so can promote efficiency that deep learning system is recognized automatically applied to image feature with Correctness, and the stage of extremely large amount of flaw sample is arrived in not yet accumulation, it will be able to it is special to generate image using deep learning system Automatic identification algorithm is levied, in the early stage of new product processing procedure, can effectively be improved with image feature automatic identification technique Processing procedure reduces flaw, saves a large amount of time and cost.

For that can be further understood that feature and technology contents of the invention, please refer to the following detailed descriptions related to the present invention With attached drawing, however provided attached drawing is merely provided for reference and description, is not intended to limit the present invention.

Detailed description of the invention

Fig. 1 is the image feature automatic identifier functional block diagram of first embodiment of the invention.

Fig. 2A is that the image processing program of first embodiment through the invention adjusts the image shape of the first sample image Adjust result schematic diagram.

Fig. 2 B is that the image processing program of first embodiment through the invention adjusts the image gray scale of the second sample image Adjust result schematic diagram.

Fig. 3 is the image feature automatic identification system functional block diagram of second embodiment of the invention.

Fig. 4 is the broad flow diagram of the image feature automatic identification method of third embodiment of the invention.

Fig. 5 is that the image feature automatic identification method of fourth embodiment of the invention executes the flow chart of Pre-testing program.

Specific embodiment

Be below illustrated by specific specific example it is presently disclosed it is related " image feature automatic identifier, The embodiment of system and method ", those skilled in the art can understand advantages of the present invention by content disclosed in this specification With effect.The present invention can be implemented or be applied by other different specific embodiments, the various details in this specification Various modifications and change can be carried out without departing from the spirit of the present invention based on different viewpoints and application.In addition, of the invention is attached Figure is only simple schematically illustrate, not according to the description of actual size, states in advance.The following embodiments and the accompanying drawings will further specifically Bright the relevant technologies content of the invention, but the protection scope that disclosure of that is not intended to limit the invention.

First embodiment

It please refers to shown in Fig. 1, Fig. 2A and Fig. 2 B, Fig. 1 is that the image feature of first embodiment of the invention recognizes dress automatically Set functional block diagram.Fig. 2A is that the image processing program of first embodiment through the invention adjusts the image shape of the first sample image The adjustment result schematic diagram of shape.Fig. 2 B is that the image processing program of first embodiment through the invention adjusts the second sample image The adjustment result schematic diagram of image gray scale.By in above-mentioned figure it is found that first embodiment of the invention to provide a kind of image feature automatic Device for identifying 1 comprising storage element 11, processing unit 12 and image acquisition unit 13.

Refering to Figure 1, storage element 11 stores database 111, database 111 stores the first image feature class Other group G1, the second image feature classification group G2 and an at least standard picture P0.First image feature classification group G1 storage There are multiple first sample image P11~P1N；Second image feature classification group G2 stores multiple second sample image P21 ~P2N.Processing unit 12 is connect with 11 signal of storage element；Image acquisition unit 13 is connect with 12 signal of processing unit, with In the testing image for capturing object (attached drawing is not shown) to be detected.

For convenience of explanation, for below by way of the Defect Detection of contact lenses, illustrate the shadow of first embodiment of the invention As the function mode of feature automatic identifier 1.Please arrange in pairs or groups the functional block diagram of Fig. 1, together refering to shown in Fig. 2A and Fig. 2 B. In the first embodiment of the present invention, the storage element 11 of device 1 regards as qualified contact lenses after storing artificial screening Image (i.e. standard picture P0) and multiple and different flaw contact lens images (i.e. first sample image P11~P1N and second Sample image P21~P2N), the contact lens image of aforementioned qualification and multiple and different flaw contact lens images are all stored up There are in database 111.

In the present embodiment, multiple first sample image P11~P1N are respectively provided with different first image feature D1, more A second sample image P21~P2N is respectively provided with the second different image feature D2.More specifically, working as specific contact lenses When with bubble, contact lenses sample image obtained is shot, compared to qualified contact lens image, can have " has The unwanted visual characteristic (calling " bubble artifacts " in the following text) of bubble ", in the present embodiment just by the unwanted visual characteristic of this " bubble artifacts " as One image feature D1, and the sample image for having the first image feature D1 be the first sample image P11 in the present embodiment~ P1N.According to the flaw contact lenses taken, the first image feature D1 on each first sample image P11~P1N is (i.e. Bubble) it can all be slightly different.On the other hand, when specific contact lenses are during production, in edge generation when demoulding Flaw then shoots contact lenses sample image obtained, compared to qualified contact lens image, can have " edge The unwanted visual characteristic (calling " demoulding flaw " in the following text) of generation flaw ", in the present embodiment just does the unwanted visual characteristic of this " demoulding flaw " For the second image feature D2, and the sample image for having the second image feature D2 is the second sample image in the present embodiment P21~P2N.The second image feature according to the flaw contact lenses taken, on each second sample image P21~P2N D2 (i.e. " demoulding flaw ") also can be slightly different.

In order to allow deep learning system that can correctly pick out different unwanted visual characteristics, it is necessary to suitably by unwanted visual characteristic Classified and is marked.In addition, importantly, sufficient amount of sample image must be provided for deep learning system It practises.In the present embodiment, the producer of contact lenses may first in a manual manner (can certainly be in AOI detection device Under auxiliary), circle selects sample image defective, and according to the type of flaw, will categorizedly have the first image feature First sample image P11~P1N of D1 is stored into the first image feature classification group G1 of database 111, will have second Second sample image P21~P2N of image feature D2 is stored into the second image feature classification group G2 of database 111.It lifts The first sample image P11~P1N data that respectively can have the first image feature D1 for 100 for example and 100 tools There is second sample image P21~P2N data storage of the second image feature D2 into database 111.

However, recognizing automatically to allow the algorithm of subsequent generation that can accurately carry out image feature, therefore, the present invention is first Sample image is expanded.In the present embodiment, multiple sample images in the first reading database 111 of the meeting of processing unit 12, And image processing program is carried out to multiple sample images, to generate multiple images processing result respectively.Image processing program packet Include image shape adjustment programme, image contrast's adjustment programme, image gray scale adjustment programme and photo dazzle color temperature adjustment programme it One or more of combination.Processing unit 12 carries out multiple images processing result with multiple standard picture P0 respectively Overlapping, to generate multiple sample images of amplification respectively.

Specifically for example, it please refers to shown in Fig. 2A, in the present embodiment, processing unit 12 can first reading database The first sample image P11 in 111, and the first sample image P11 is stretched or deformed by image shape adjustment programme, To generate multiple images processing result, and multiple images processing result is overlapped with standard picture P0 respectively, to generate respectively Multiple first sample image P111~P113 of amplification.It more specifically, is the first image spy for being directed to the first sample image P11 Sign D1 (i.e. on image by " bubble artifacts " part that circle is selected) is stretched or is deformed, and by deformed processing result image It is overlapped respectively with standard picture P0.

Separately please refer to shown in Fig. 2 B, in the present embodiment, processing unit 12 can also in first reading database 111 second Sample image P21, and by image gray scale adjustment programme to the second image feature D2 of the second sample image P21 (or only for Upper " demoulding flaw " part at contact lenses edge of image) grayscale adjustment is carried out, to generate multiple images processing result, and Multiple images processing result is overlapped with standard picture P0 respectively, to generate multiple second sample images of amplification respectively P211 and P212.

Although the first sample image P11 is stretched or deformed with image shape adjustment programme respectively above, and lead to It crosses image gray scale adjustment programme and grayscale adjustment, but the classification of image processing program and first is carried out to the second sample image P21 There is no fixed corresponding relationships for the classification of image feature D1 or the second image feature D2.In addition, aforementioned image shape adjusts journey It is between sequence, image contrast's adjustment programme, image gray scale adjustment programme and photo dazzle color temperature adjustment programme nor mutually exclusive Relationship, but can be combined with each other collocation and carry out image adjustment, or even other image regulating methods are added jointly to sample drawing As being adjusted, to increase the diversity of processing result image.It accordingly, just can be by original respective only 100 the first samples Product image P11~P1N and second sample image P21~P2N data massive amplification (for example, be expanded to 100,000 stroke counts respectively According to).

Special one is needed to be mentioned that, in the present embodiment, database 111 only stores a standard picture P0, but originally Invention is not limited thereto, the specific practice, can also store a standard respectively in each image feature classification group Image P0, and the operations such as image procossing and overlapping are executed respectively.

Next, processing unit 12 is according to multiple first sample image P11~P1N and more before amplification and after amplification A second sample image P21~P2N executes the training program of deep learning system, to establish the automatic identification algorithm of image feature. Specifically, used deep learning system can be Caffe, Theano, TensorFlow or Lasagne, Keras very To the frame of DSSTNE etc., the present invention is not limited specifically and is carried out using which kind of frame.Pass through image caused by foregoing manner The automatic identification algorithm of feature includes the first identification for the first image feature D1 (being in the present embodiment " bubble artifacts ") Standard and the second appreciative standard for being directed to the second image feature D2 (being in the present embodiment " demoulding flaw ").

After the completion of the automatic identification algorithm of image feature is established, detectable substance progress just can be treated with the device of the invention 1 Detection.Specifically, image acquisition unit 13 (such as pick-up lens) captures the testing image of object to be detected, then, processing unit 12 self imaging acquisition units 13 obtain testing image, and execute the automatic identification algorithm of image feature and analyze testing image. Processing unit 12 judges whether testing image has the first image feature D1 according to the first appreciative standard, and is marked according to the second identification Standard judges whether testing image has the second image feature D2.

It so far, is exactly the function mode on 1 basis of image feature automatic identifier of the invention.In the present embodiment, lead to Massive amplification the first sample image P11~P1N and second sample image P21~P2N is crossed, so that in the instruction of deep learning system Practice in program, the automatic identification algorithm of more accurately image feature can be established out.

The discovery and foundation of flaw classification

Although it should be strongly noted that doing example explanation immediately above with " bubble artifacts " and " demoulding flaw ", originally Invention is not limited to set a variety of flaw types at the very start to be detected, and most can also start only to set a kind of flaw Defect feature detects, and during the automatic identification algorithm of subsequent execution image feature analyzes testing image, ability root Other the unwanted visual characteristic classifications to be tested are screened and established according to testing result.

As an example it is assumed that only storing the first image feature classification group G1 in original database 111, and generated The automatic identification algorithm of image feature, which is only set, recognizes the first image feature D1 of first sample image P11~P1N The first appreciative standard.However, during being tested by the automatic identification algorithm of image feature to multiple testing images, Processing unit 12 judges that multiple testing images do not meet standard picture P0, but all has the second image feature D2, at this point, processing is single Member 12 can establish the second image feature D2 classification group G2 in database 111, and have the second image special for multiple respectively The testing image of sign D2 is respectively recorded as multiple second sample image P21, and is stored in the second image feature D2 classification group G2 In.By this mode, help to find and establish new flaw classification, to lack for unknown on product more initiatively Be trapped into row improves in real time.

Accuracy verifying and improvement

By process presented hereinbefore, it has been enough to enable deep learning system of the invention using very huge and more The sample image of sample is trained, and can generate the judgement more accurately automatic identification algorithm of image feature.However, aforementioned expansion Generated sample image after increasing, naturally it is also possible to carry out image processing program again, further amplification becomes a greater amount of samples Product image, and may further improve the precision of the automatic identification algorithm identification of image feature.It must in order to be confirmed whether to have Sample image is expanded again, the device of the invention 1 can also be in the automatic identification algorithm pair of practice image feature Before testing image is tested, the accuracy that the generated automatic identification algorithm of image feature carries out identification operation is verified in advance.

The specific practice is by processing unit 12 before the training program for the deep learning system that executes, first by multiple the The first sample image of a part (such as 2 one-tenth) P11 in a sample image P11 is selected as the first verifying image, and with remaining The training program of multiple (such as remaining 8 one-tenth) first sample image P12~P1N execution deep learning system.Also, in depth After learning system establishes the automatic identification algorithm of image feature, processing unit 12 executes the automatic identification algorithm of image feature to previously choosing The first verifying selected is verified with image.That is, being marked according to the first identification in the automatic identification algorithm of image feature Standard, judges whether the first verifying image has the first image feature D1, to confirm in the automatic identification algorithm of image feature The correctness of one appreciative standard.If the first verifying image can not be picked out with the first image feature D1, or will not have The image (such as standard picture P0) of first image feature D1 is recognized as having the first image feature D1, then it represents that identification result is wrong Accidentally；Conversely, indicating that identification result is correct.

The user of device 1 can preset the first accuracy threshold value with through the processing unit 12, further, it is also possible to according to Need to adjust the quantity of the first verifying image ratio shared in multiple first sample image P11.It is practical when carrying out, it can be with The quantity of first verifying image is set in 3% to 50% in all the first sample image P11.Executing image feature certainly Dynamic identification algorithm, and whether have first according to the first appreciative standard judgment criteria image P0 and multiple first verifying images After image feature D1, its identification result is recorded correctly or incorrectly, in other words, the first appreciative standard pair is recorded according to judging result The first identification accuracy of first image feature D1.

Next, the first identification accuracy is compared to each other with preset first accuracy threshold value.It is distinguished first When knowing accuracy lower than the first accuracy threshold value, expression is supplied to the sample image that deep learning system is trained may be simultaneously Not enough, therefore, in order to execute further training, processing unit 12 by multiple first sample image P11 again into Row image processing program, with the further multiple first sample image P11 for generating amplification.Certainly, expand again again to verify Sample image after increasing is provided to deep learning system and is trained, and after the first appreciative standard of amendment, if has been able to The automatic identification algorithm of sufficiently accurate image feature is generated, therefore, is still to carry out proving program above-mentioned.Specifically, processing For unit 12 in multiple first sample image P11 after expanding again, selecting 3% to 50% again is the first verifying image, And remaining multiple first sample image P11 is again supplied to deep learning system, to correct the first appreciative standard, and to repair Whether judgment criteria image P0 and multiple first verifying images have the first image special to the first appreciative standard after just again D1 is levied, and obtains the first identification accuracy again.If will be returned to that is, the first accuracy threshold value cannot be reached Image processing program further expands sample image.Until the first identification accuracy is being higher than (including just reaching) first just When true rate threshold value, the automatic identification algorithm of image feature is just applied to the detection of object to be detected, and by processing unit 12 The automatic identification algorithm of image feature is executed, with according to modified first appreciative standard final in the automatic identification algorithm of image feature, Judge capture from the testing image of object to be detected whether there is the first image feature D1.In order to avoid setting excessively high at the beginning One accuracy threshold value leads to that standard is not achieved always, and image processing program is caused constantly to expand sample image, can also be with Set other stop conditions, but itself and non-present invention institute and the emphasis emphasized, in addition how actual conditions set is not gone to live in the household of one's in-laws on getting married herein It states.

Pre-testing

On the other hand, it is also necessary to a problem of consideration is, if the first image is special in the sample image being initially provided of The diversity for levying D1 is just insufficient, then the automatic identification algorithm of image feature for being likely to influence subsequent foundation is being identified When correctness, even if carrying out repeated amplification to sample image, it is also possible to because diversity at the beginning is too low to carry out Efficient training program.In order to solve the problems, such as this one, in the present embodiment, device 1 is before executing image processing program, moreover it is possible to Processing unit 12 is enough set to first carry out a Pre-testing program in advance.

Specifically, multiple first sample image P11~P1N (are not yet carried out any image procossing journey by processing unit 12 Sequence) in 3% to 50% be selected as the first preliminary examination image, and by remaining multiple first sample drawing in the first sample image P11 It include one first preliminary examination standard inspection for the first image feature D1 to establish as P11~P1N is supplied to deep learning system It is quasi-.It is enough due to only tentatively to confirm whether first acquired sample image P11~P1N has herein during one Diversity, therefore can be carried out using the more simplified deep learning system of framework.It similarly, also will 12 setting through the processing unit Door is defined as the first Pre-testing threshold value herein.Due to the diversity only for first sample image P11~P1N A preliminary judgement is done, therefore, the first Pre-testing threshold value need not set too high.It is quasi- compared to being previously used for ensuring finally judging First accuracy threshold value of true rate, the numerical requirements of the first Pre-testing threshold value herein are lower, therefore, take at the same time In the case of Pre-testing and accuracy verifying, the first Pre-testing threshold value can be lower than the first accuracy threshold value.Processing unit Whether 12 have the first image feature according to the first Pre-testing standard judgment criteria image P0 and multiple first verifying images D1, and the first Pre-testing standard is recorded to the first Pre-testing accuracy of the first image feature D1 according to judging result.

In aforementioned Pre-testing program, if the first Pre-testing accuracy, which has, reaches the first Pre-testing threshold value, the is indicated The diversity of a sample image P11~P1N, which has, reaches desired standard, therefore can continue main flow of the invention, In other words, image processing program can then be executed；Conversely, if the first Pre-testing accuracy is lower than the first Pre-testing door When value, indicate first sample image P11~P1N diversity be in fact it is inadequate, just should terminate that program at this time, not hold The subsequent image processing program of row, and should first acquire more the first sample image P11 with the first image feature D1 of difference ~P1N, to guarantee to carry out deep learning system the effective training of essence.

Second embodiment

It please refers to shown in Fig. 3, Fig. 3 is 2 functional block diagram of image feature automatic identification system of second embodiment of the invention. By upper figure it is found that the second embodiment of the present invention provides a kind of image feature automatic identification system 2 comprising server-side 21, inspection End 22 and sample image feed end 23 are surveyed, in the present embodiment, server-side 21 is the offer of the automatic identification algorithm of image feature Person, test side 22 are connect with 21 signal of server-side, and can receive the automatic identification algorithm of image feature by server-side 21, so that detection End 22 is able to carry out the automatic identification algorithm of image feature, and the testing image for treating detectable substance is detected and analyzed.As for sample Image feed end 23 is also to connect with 21 signal of server-side, and be capable of providing multiple first sample image P11~P1N to server-side 21.On the practical framework of system 2, test side 22 may be same with sample image feed end 23, it is also possible to by distinguishing Setting.

It illustrates, can be seen that in the present embodiment from block diagram shown in Fig. 3, server-side 21, test side 22 And sample image feed end 23 be by network each other signal connect, and can exchange each other signal (such as transmission sample drawing Picture or the automatic identification algorithm of image feature), however, so-called " signal connection " of the invention is not limited to this, by server-side 21, inspection The data storage of end 22 and 23 either end of sample image feed end is surveyed in storage medium such as CD, flash memory or hard disks, then Other any one end are served data to, " signal connection " referred to herein is also complied with, indicates hereby.

Equally using the Defect Detection of contact lenses as example, in simple terms, (the contact lenses production of sample image feed end 23 With proofer) image feature that can recognize " bubble artifacts " (the first image feature D1) automatically in order to obtain recognizes calculation automatically Method, therefore multiple first sample image P11~P1N with the first different image feature D1 are supplied to server-side 21 and (are calculated Method supplier).Server-side 21 generates image by a series of step after receiving multiple first sample image P11~P1N The automatic identification algorithm of feature, and the automatic identification algorithm of image feature is supplied to test side 22 (contact lenses production and inspection Person), so that test side 22 can apply the automatic identification algorithm of image feature in contact lenses produced (object to be detected) In detection.It in other words, can be from the sample image supply for having demand although server-side 21 and the detection without contact lenses The material for obtaining at 23 and generating the automatic identification algorithm of image feature is held, and finished product (the automatic identification algorithm of image feature) is provided To equally having demand and the test side 22 of the automatic identification algorithm of image feature will be executed.

Specifically, server-side 21 includes storage element 211 and processing unit 212, and storage element 211 stores data Library 2111, database 2111 store the first image feature classification group G1 and at least a standard picture P0, and the first image is special Sign classification group G1 stores multiple first sample image P11~P1N (being provided by sample image feed end 23), each the A sample image P11~P1N is respectively provided with the first different image feature D1.Processing unit 212 and 211 signal of storage element connect Connect, and can multiple first sample image P11~P1N in reading database 2111, and carry out an image processing program respectively. Image processing program includes image shape adjustment programme, image contrast's adjustment programme, image gray scale adjustment programme and image A combination of one or more among colour temperature adjustment programme.Processing unit 212 respectively will be caused by image processing program Processing result image is respectively overlapped with standard picture P0 to generate multiple first sample image P11~P1N of amplification.

Processing unit 212 executes a depth according to multiple first sample image P11~P1N before amplification and after amplification The training program of learning system, to establish the automatic identification algorithm of image feature, the automatic identification algorithm of image feature includes for The first appreciative standard of one image feature D1.

More specifically, in the present embodiment, processing unit 212 is before the training program for executing deep learning system, just At least a sample image in multiple first sample image P11~P1N first can be selected as the first verifying image, and with remaining Multiple first sample image P11~P1N execute the training program of deep learning system.When practical operation, image is used in the first verifying Quantity be it is multiple, and the quantity of the first verifying image accounts for the 3% to 50% of multiple first sample image P11~P1N quantity. In addition, processing unit 212 sets the first accuracy threshold value, and image feature is established in deep learning system and recognizes calculation automatically After method, processing unit 212 executes the automatic identification algorithm of image feature, with according to the first appreciative standard judgment criteria image P0 and Whether multiple first verifying images have the first image feature D1, to confirm the correctness of the first appreciative standard.Processing unit 212, which record the first appreciative standard according to judging result, recognizes accuracy to the first of the first image feature D1, and first is distinguished Accuracy is known compared with aforementioned first accuracy threshold value.

When the first identification accuracy is lower than the first accuracy threshold value, processing unit 212 is by multiple first sample images P11~P1N carries out image processing program again, with the further multiple first sample image P11~P1N for generating amplification, and In multiple first sample image P11~P1N after expanding again, 3% to 50% is selected again as the first verifying image, and Remaining multiple first sample image P11~P1N is again supplied to deep learning system, to correct the first appreciative standard, and With revised first appreciative standard, whether judgment criteria image P0 and multiple first verifying images have the first shadow again As feature D1, and the first identification accuracy is obtained again.Conversely, if the first identification accuracy has reached the first accuracy door The automatic identification algorithm of image feature can then be provided and give test side 22 by bank value.

Test side 22 include image acquisition module 222 and processing module 221, image acquisition module 222 to capture to The testing image of detectable substance.Processing module 221 is connect with 222 signal of image acquisition module, and can be from image acquisition module 222 Obtain testing image.After the reception automatic identification algorithm of image feature of server-side 21, processing module 221 executes image for test side 22 The automatic identification algorithm of feature analyzes testing image, and judges whether testing image has first according to the first appreciative standard Image feature D1.

Special one is needed to be mentioned that, since database 2111 in the present embodiment is to be located at server-side 21, without examining End 22 is surveyed, therefore, in the present embodiment, when the processing module 221 of test side 22 judges that multiple testing images do not meet standard drawing As P0, but when all there is the second image feature D2, test side 22 can respectively by it is multiple have the second image feature D2 to mapping As being respectively recorded as multiple second sample images, and multiple second sample images are supplied to server-side 21.After this, then by The processing unit 212 of server-side 21 establishes the second image feature classification group G2 in database 2111, and by it is multiple received from Second sample image of test side 22 is stored in the second image feature classification group G2.

After completing new classification and establishing, the processing unit 212 of server-side 21 can also further reading database Multiple second sample images in 2111, and image processing program is carried out respectively, and respectively will be caused by image processing program Processing result image is respectively overlapped with standard picture P0 to generate multiple second sample images of amplification.And processing unit 212 execute the training program of deep learning system according to multiple first sample image P11~P1N before amplification and after amplification, To establish the second appreciative standard for being directed to the second image feature D2 in the automatic identification algorithm of image feature.After completion, servo End 21 can provide the automatic identification algorithm of image feature to test side 22 again.When test side 22 receives image feature certainly again It is according to the after dynamic identification algorithm, and when processing module 221 analyzes testing image with the automatic identification algorithm of image feature One appreciative standard judges whether testing image has the first image feature D1, and judges that testing image is according to the second appreciative standard It is no that there is the second image feature D2.

As described in previously, the present invention can make test side 22 that image feature is automatic by above-mentioned technical characteristic The actual use of identification algorithm is as a result, Real-time Feedback allows server-side 21 that can do further analysis to server-side 21.

Another embodiment is the discovery after test side 22 is detected using the automatic identification algorithm of image feature It is undetected (regardless of whether presetting the unwanted visual characteristic to be recognized) there are bad products, and will be corresponded to by the mode of manual record Testing image store into sample image, and be supplied to server-side 21 and analyzed.If server-side 21 finds have after an analysis New flaw classification, the then adjustment of test side 22 of issuing a separate notice, optimization processing procedure, forms a kind of benign interactive process.

It mentions as in the first embodiment, if in the sample image being initially provided of, the first image feature D1's Diversity is just insufficient, then being likely to influence the automatic identification algorithm of image feature of subsequent foundation when being identified just True property.Therefore, in the present embodiment, device 1 is before executing image processing program, additionally it is possible to processing unit 12 be made to first carry out one in advance A Pre-testing program.After server-side 21 receives multiple first sample image P11~P1N from sample image feed end 23, also can Enough Pre-testing programs that first carries out are to confirm whether first received sample image P11~P1N has enough diversity.Tool For body, the processing unit 212 of server-side 21 by 3% to 50% in multiple first sample image P11~P1N (preferably 15% to 25%) it is selected as the first preliminary examination image, and by remaining multiple first sample drawing in first sample image P11~P1N As P11~P1N (i.e. 50% to 97% first sample image P11~P1N, preferably 75% to 85% the first sample image P11~P1N) it is supplied to deep learning system, it include the first Pre-testing standard for the first image feature D1 with foundation.Together Sample, also 212 it can set the first Pre-testing threshold value through the processing unit in the present embodiment.Processing unit 212 is according to depth Whether the first Pre-testing standard judgment criteria image P0 and multiple first verifying images that learning system is established have first Image feature D1, and it is correct to the first Pre-testing of the first image feature D1 according to judging result the first Pre-testing standard of record Rate.After obtaining the first Pre-testing accuracy, processing unit 212 is by the first Pre-testing accuracy and preset first Pre-testing Threshold value compares.

If the first Pre-testing accuracy is lower than the first Pre-testing threshold value, then it represents that sample image feed end 23 provided First sample image P11~P1N diversity is insufficient, at this point, in the present embodiment, subsequent shadow item amplification is not carried out in server-side 21 Program, but more first different sample image P11~P1N are asked for sample image feed end 23, to enrich the first sample The diversity of image P11~P1N.It in the nature of things, still will be again after receiving more first sample image P11~P1N Confirm whether its diversity has reached standard by educating check problem.Conversely, it is pre- to reach first in the first Pre-testing accuracy When examining threshold value, server-side 21 can continue main flow of the invention, in other words, can hold to received Row the first sample image P11~P1N image processing program is to be expanded.

3rd embodiment

It please refers to shown in Fig. 4, Fig. 4 is the main flow of the image feature automatic identification method of third embodiment of the invention Figure.Illustrate the main flow of image feature automatic identification method provided by the present invention below by way of Fig. 4.Image of the invention is special Sign automatic identification method mainly includes the following steps:

S100: sample image is obtained；

S102: sample image is expanded by image processing program；

S104: verifying image is chosen in multiple sample images；

S106: remaining sample image is supplied to deep learning system and is trained；

S108: with the verifying automatic identification algorithm of image feature of vision inspections deep learning system output；

S110: judge to recognize whether accuracy reaches accuracy threshold value, if so, entering step S112；(it is lower than) if not, Then return to step S102；

S112: the testing image of object to be detected is captured；

S114: testing image is recognized with the automatic identification algorithm of image feature.

Specifically, image feature automatic identification method of the invention can be directed to sample after obtaining sample image first Image carries out image processing program to be expanded (step S100 and step S102).Its specific practice is to be directed to be respectively provided with Multiple first sample images of the first different image features, carry out image shape adjustment programme, image contrast's adjustment programme, A combination of one or more among image gray scale adjustment programme and photo dazzle color temperature adjustment programme, and respectively will be at image Processing result image caused by reason program is overlapped with the standard picture for not having the first image feature to generate amplification Multiple first sample images.

In order to verify whether the first sample image after amplification has been enough to establish the image for providing accurate test effect Therefore at least a sample image in multiple first sample images is selected as in advance in this stage by the automatic identification algorithm of feature Image (step S104) is used in one verifying.At the same time, the first accuracy threshold value can also be set together.In the present embodiment, The quantity of first verifying image accounts for 3% to 50% (for example, taking 20%) of multiple first sample image quantity.First is correct Rate threshold value can do appropriate adjustment for the serious forgiveness of specific flaw depending on product, for example, for " the gas of contact lenses Steep flaw ", due to belonging to more serious flaw, and detection difficulty is not very high, accuracy that should be more demanding, at this point, First accuracy threshold value can be set as 99% or more.On the other hand, when image feature automatic identification method of the invention is same When be applied to detection the second image feature when, can also for the second image feature and with the second image feature second Sample image selects the second verifying image, and sets the second accuracy threshold value.For " demoulding flaw " of contact lenses, Due to belonging to the more difficult flaw accurately judged, if the second image feature is " demoulding flaw " of contact lenses, at this time it is contemplated that Second accuracy threshold value is set as 90~95% or so.

Next, multiple first sample images (or including second sample image) before amplification and after amplification are supplied to Deep learning system is trained, to establish the automatic identification algorithm (step S106) of image feature.Image feature recognizes calculation automatically Method includes the first appreciative standard for the first image feature.

Due to previously selecting a part in multiple first sample images as the first verifying image, After deep learning system establishes the automatic identification algorithm of image feature, whether the first verifying image is judged according to the first appreciative standard With the first image feature, test (step to the automatic identification algorithm of image feature of deep learning system output accordingly S108).Specific practice is, executes the automatic identification algorithm of image feature, and according to the first appreciative standard judgment criteria image and Whether multiple first verifying images have the first image feature, record the first appreciative standard to the first image according to judging result First identification accuracy of feature then compares the first identification accuracy with preset first accuracy threshold value, Whether it is higher than preset first accuracy threshold value (step S110) with confirmation the first identification accuracy.

If the first identification accuracy is lower than the first accuracy threshold value, then step S102 is returned to, by multiple first samples Image carries out image processing program again, with further multiple first sample images for generating amplification, and sequentially executes step S104 to step S110 selects 3% to 50% again and uses for the first verifying from multiple first sample images after amplification again Image, and remaining multiple first sample image is again supplied to deep learning system, to correct the first appreciative standard, and with Whether judgment criteria image and multiple first verifying images have the first image special to revised first appreciative standard again Sign, and the first identification accuracy is obtained again.It is correct with preset first again after obtaining the first identification accuracy again Rate threshold value compares, until the first identification accuracy reaches the first accuracy threshold value.

After the first identification accuracy reaches the first accuracy threshold value, so that it may which the image feature established is automatic Identification algorithm is applied to detect object to be detected.Specifically, first have to capture the testing image (step of object to be detected S112), next, being analyzed with the automatic identification algorithm of image feature testing image, and judged according to the first appreciative standard Whether testing image has the first image feature (step S114).

It is above exactly the main flow of image feature automatic identification method of the invention.Certainly, image feature of the invention Automatic identification method, additionally it is possible to when judging that multiple testing images do not meet standard picture, but all there is the second image feature, point Multiple testing images multiple second sample images are not recorded as.This part details can refer to first and second embodiment, herein It does not repeat to repeat.In addition, image feature automatic identification method of the invention to be applied to have the second different image features Multiple second sample images when, equally to carry out following procedure: to being respectively provided with multiple the of the second different image features Two sample images carry out image processing program, and respectively by processing result image caused by image processing program with do not have the The standard picture of two image features is overlapped to generate multiple second sample images of amplification；It will be before amplification and after amplification Multiple second sample images are supplied to deep learning system, are directed to the second image to establish in the automatic identification algorithm of image feature One second appreciative standard of feature；The testing image of object to be detected is captured, with the automatic identification algorithm of image feature to testing image It is analyzed；Wherein, judge whether testing image has the first image feature according to the first appreciative standard；Wherein, according to second Appreciative standard judges whether testing image has the second image feature.

Fourth embodiment

In order to ensure the sample image most started has enough diversity, image feature automatic identification method of the invention Can also aforementioned third embodiment step S100 to the centre of step S102, add Pre-testing process.Please refer to Fig. 5 institute Show, Fig. 5 is that the image feature automatic identification method of fourth embodiment of the invention executes the flow chart of Pre-testing program.Of the invention Image feature automatic identification method mainly includes the following steps: when executing Pre-testing program

S100: sample image is obtained；

S116: preliminary examination image is chosen in multiple sample images；

S118: remaining sample image is supplied to deep learning system and is trained；

S120: with preliminary examination vision inspections sample image diversity；

S122: judge to recognize whether accuracy is higher than Pre-testing threshold value, if so, entering step S102；(it is lower than) if not, Then return to step S100；

S102: sample image is expanded by image processing program；

Specifically, image feature automatic identification method of the invention, can be first by multiple first after obtaining sample image 3% to 50% in sample image is selected as the first preliminary examination image (step S100 and step S116).It at the same time, can also be with The first Pre-testing threshold value is set together.As described in previously, the first Pre-testing threshold value is not necessarily like the first accuracy threshold value It is so high, therefore, it can be set as 70% to 90%, certainly, concrete condition optionally can be adjusted voluntarily.

It is instructed next, remaining multiple first sample image in the first sample image are supplied to deep learning system Practice, includes the first Pre-testing standard (step S118) for the first image feature with foundation.

After the completion of abovementioned steps, with preliminary examination vision inspections sample image diversity (step S120).Specifically, it is Whether there is the first image feature according to the first Pre-testing standard judgment criteria image and multiple first preliminary examination images, according to Judging result records the first Pre-testing standard to the first Pre-testing accuracy of the first image feature.

Next, the first Pre-testing accuracy is compared (step S122) with the first Pre-testing threshold value.Wherein, When one Pre-testing accuracy is lower than the first Pre-testing threshold value, indicate that the diversity of the first sample image is insufficient, it is necessary to increase The quantity of the first sample image with the first different image features, therefore step S100 is returned to, obtaining more has difference The first image feature the first sample image；Conversely, when the first Pre-testing accuracy reaches the first Pre-testing threshold value, i.e., The main flow for carrying out image feature automatic identification method of the present invention can be connected, in other words, S102 can be entered step, executes figure As processing routine expands the first sample image.

The beneficial effects of the present invention are image feature automatic identifiers, system provided by technical solution of the present invention And method, can by " to be respectively provided with different images feature multiple sample images carry out image processing program, and with do not have There is the standard picture of image feature to be overlapped to generate multiple sample images of amplification " and " will be before amplification and after amplification Multiple sample images be supplied to deep learning system " technical characteristic, to promote the training process institute under deep learning system The data diversity that can be touched, and then efficiency that deep learning system is recognized automatically applied to image feature and correct can be promoted Property, and the stage of extremely large amount of flaw sample is arrived in not yet accumulation, it will be able to image feature is generated certainly using deep learning system Dynamic identification algorithm effectively can improve processing procedure with image feature automatic identification technique in the early stage of new product processing procedure Flaw is reduced, a large amount of time and cost are saved.

Content disclosed above is only preferred possible embodiments of the invention, not thereby limits to right of the invention and wants The protection scope of book is sought, so all equivalence techniques variations done with description of the invention and accompanying drawing content, are both contained in In the protection scope of claims of the present invention.

Claims

1. a kind of image feature automatic identifier characterized by comprising

One storage element, stores a database, and the database stores one first image feature classification group and extremely A few standard picture, the first image feature classification group store multiple first sample images, each described first sample Product image is respectively provided with the first different image features；

One processing unit is connect with the storage element signal；And

One image acquisition unit is connect with the processing unit signal, with the testing image for capturing an object to be detected:

Wherein, the processing unit reads multiple first sample images in the database, multiple first samples Image carries out image processing program to generate multiple images processing result respectively, and multiple described image processing results respectively with it is more A standard picture is overlapped, to generate multiple first sample images of amplification respectively；

Wherein, the processing unit executes a deep learning according to multiple first sample images before amplification and after amplification The training program of system, to establish the automatic identification algorithm of an image feature, the automatic identification algorithm of image feature includes needle To one first appreciative standard of first image feature；

Wherein, the processing unit obtains the testing image from the image acquisition unit, described in the processing unit executes The automatic identification algorithm of image feature analyzes the testing image, and the processing unit is according to first appreciative standard Judge whether the testing image has first image feature.

2. the apparatus according to claim 1, which is characterized in that described image processing routine includes image shape adjustment journey Sequence, image contrast's adjustment programme, image gray scale adjustment programme and photo dazzle color temperature adjustment programme one or both of with On combination.

3. the apparatus according to claim 1, which is characterized in that the processing unit also further executes following procedure:

Before the training program for the deep learning system that executes, by least one first sample in multiple first sample images Image is selected as one first verifying image, and executes the deep learning system with remaining multiple described first sample image Training program；And

After the deep learning system establishes the automatic identification algorithm of the image feature, judged according to first appreciative standard Whether the first verifying image has first image feature, to confirm the correctness of first appreciative standard.

4. device according to claim 3, which is characterized in that the first verifying quantity of image is multiple, and institute The quantity for stating the first verifying image accounts for the 3% to 50% of multiple first sample image quantity, the processing unit also into One step executes following procedure:

Set one first accuracy threshold value；And

It executes the automatic identification algorithm of the image feature, and the standard picture and is judged according to first appreciative standard more Whether a first verifying image has first image feature, records first appreciative standard according to judging result Accuracy is recognized to the one first of first image feature；

Wherein, the following steps are executed when the first identification accuracy is lower than the first accuracy threshold value:

Multiple first sample images are subjected to described image processing routine again, with further multiple first samples for generating amplification Image；And

In multiple first sample images after expanding again, select 3% to 50% again as the first verifying image, and by its Remaining multiple first sample images are again supplied to the deep learning system, to correct first appreciative standard, and to repair First appreciative standard after just judges whether the standard picture and multiple first verifying images have again First image feature, and the first identification accuracy is obtained again；

Wherein, when the first identification accuracy reaches the first accuracy threshold value, according to first appreciative standard Judge capture from the testing image of the object to be detected whether there is first image feature.

5. the apparatus according to claim 1, which is characterized in that the processing unit also further executes following procedure:

When judging that multiple testing images do not meet the standard picture, but all there is second image feature, described One second image feature classification group of Database, and have multiple described in second image feature to mapping respectively As being respectively recorded as multiple second sample images, and it is stored in the second image feature classification group.

6. device according to claim 5, which is characterized in that the processing unit also further executes following procedure:

Multiple second sample images in the database are read, and carry out described image processing routine respectively, and respectively Processing result image caused by described image processing routine is respectively overlapped with the standard picture to generate amplification Multiple second sample images；

Execute the training program of the deep learning system according to multiple first sample images before amplification and after amplification, with One second appreciative standard for being directed to second image feature is established in the automatic identification algorithm of image feature；And

Judge whether the testing image has first image feature according to first appreciative standard, and according to described Two appreciative standards judge whether the testing image has second image feature.

7. the apparatus according to claim 1, which is characterized in that the processing unit is executing described image processing routine Before, also further execute following procedure:

3% to 50% in multiple first sample images is selected as the first preliminary examination image, and by first sample drawing Remaining multiple first sample image as in are supplied to the deep learning system, include for first image to establish One first Pre-testing standard of feature；

Set one first Pre-testing threshold value；And

Judge whether the standard picture and multiple first verifying images have according to the first Pre-testing standard First image feature records the first Pre-testing standard to the one first of first image feature according to judging result Pre-testing accuracy；

Wherein, it when the first Pre-testing accuracy reaches the first Pre-testing threshold value, executes described image and handles journey Sequence；And

Wherein, when the first Pre-testing accuracy is lower than the first Pre-testing threshold value, terminator.

8. a kind of image feature automatic identification system characterized by comprising

One server-side comprising:

One storage element, stores a database, and the database stores one first image feature classification group and extremely A few standard picture, the first image feature classification group store multiple first sample images, each described first sample Product image is respectively provided with the first different image features；And

One processing unit is connect with the storage element signal；

Wherein, the processing unit reads multiple first sample images in the database, and carries out an image respectively Processing routine, and respectively respectively fold processing result image caused by described image processing routine with the standard picture It closes to generate multiple first sample images of amplification；

Wherein, the processing unit executes a deep learning system according to multiple first sample images before amplification and after amplification Training program, to establish the automatic identification algorithm of an image feature, the automatic identification algorithm of image feature includes for institute State one first appreciative standard of the first image feature；And

One test side is connect with the servo end signal, and can be received the image feature by the server-side and be recognized automatically Algorithm, the test side include:

One image acquisition module, to capture a testing image of an object to be detected；And

One processing module is connect with the image acquisition module signal, described to be measured to obtain from the image acquisition module Image, the processing module execute the automatic identification algorithm of image feature and analyze the testing image, and according to institute It states the first appreciative standard and judges whether the testing image has first image feature.

9. system according to claim 8, which is characterized in that described image processing routine includes image shape adjustment journey Sequence, image contrast's adjustment programme, image gray scale adjustment programme and photo dazzle color temperature adjustment programme one or both of with On combination.

10. system according to claim 8, which is characterized in that the processing unit also further executes following procedure:

Before the training program for the deep learning system that executes, by least a sample image in multiple first sample images It is selected as the first verifying image, and executes the training program of the deep learning system with remaining multiple first sample image； And

11. system according to claim 10, which is characterized in that the quantity of the first verifying image is multiple, and The quantity of the first verifying image accounts for the 3% to 50% of multiple first sample image quantity, and the processing unit is also Further execute following procedure:

Set one first accuracy threshold value；And

Wherein, when the first identification accuracy is lower than the first accuracy threshold value, the following steps are executed:

Wherein, when the first identification accuracy reaches the first accuracy threshold value, the image feature is distinguished automatically Algorithm offer is provided and gives the test side.

12. system according to claim 8, it is characterised in that:

The processing module of the test side is judging that multiple testing images do not meet the standard picture, but all has When one second image feature, multiple testing images with second image feature are respectively recorded as multiple respectively Two sample images, and multiple second sample images are supplied to the server-side；

The processing unit of the server-side is in the one second image feature classification group of Database, and by multiple institutes The second sample image is stated to be stored in the second image feature classification group.

13. system according to claim 12, which is characterized in that the processing unit also further executes following procedure:

Multiple second sample images in the database are read, and carry out described image processing routine respectively, and respectively Processing result image caused by described image processing routine is respectively overlapped with the standard picture to generate amplification Multiple second sample images；And

Execute the training program of the deep learning system according to multiple first sample images before amplification and after amplification, with One second appreciative standard for being directed to second image feature is established in the automatic identification algorithm of image feature.

14. system according to claim 13, which is characterized in that the processing module is recognized automatically with the image feature It is to judge whether the testing image has institute according to first appreciative standard when algorithm analyzes the testing image The first image feature is stated, and judges whether the testing image has second image special according to second appreciative standard Sign.

15. system according to claim 8, which is characterized in that may further comprise:

A sample image feed end is connect with the servo end signal, to provide multiple first sample images to described Server-side；

Wherein, the processing unit of the server-side also further executes following procedure:

Set one first Pre-testing threshold value；And

Wherein, it when the first Pre-testing accuracy is lower than the first Pre-testing threshold value, is supplied from the sample image End receives first sample image more with the first different image features；

Wherein, it when the first Pre-testing accuracy reaches the first Pre-testing threshold value, executes described image and handles journey Sequence.

16. a kind of image feature automatic identification method, characterized in that it comprises the following steps:

One image processing program is carried out to multiple first sample images for being respectively provided with the first different image features, and respectively will Processing result image caused by described image processing routine is carried out with the standard picture for not having first image feature It overlaps to generate multiple first sample images of amplification；

Multiple first sample images before amplification and after amplification are supplied to a deep learning system, to establish an image feature Automatic identification algorithm, the automatic identification algorithm of image feature include the one first identification mark for first image feature It is quasi-；And

A testing image of an object to be detected is captured, the testing image is divided with the image feature automatic identification algorithm Analysis, and judge whether the testing image has first image feature according to first appreciative standard.

17. according to the method for claim 16, which is characterized in that described image processing routine includes image shape adjustment journey Sequence, image contrast's adjustment programme, image gray scale adjustment programme and photo dazzle color temperature adjustment programme one or both of with On combination.

18. according to the method for claim 16, which is characterized in that may further comprise:

Before multiple first sample images are supplied to the deep learning system, by multiple first sample images extremely Few a sample image is selected as the first verifying image, and remaining multiple first sample image in first sample image are mentioned Supply the deep learning system；And

19. according to the method for claim 18, which is characterized in that the quantity of the first verifying image is multiple, and The quantity of the first verifying image accounts for the 3% to 50% of multiple first sample image quantity, and the method is also into one Step includes:

Set one first accuracy threshold value；And

20. according to the method for claim 16, which is characterized in that may further comprise:

When judging that multiple testing images do not meet the standard picture, but all there is second image feature, respectively will Multiple testing images are recorded as multiple second sample images.

21. according to the method for claim 20, which is characterized in that may further comprise:

Described image processing routine is carried out to multiple second sample images for being respectively provided with the second different image features, and respectively By processing result image caused by described image processing routine and without the standard picture of second image feature It is overlapped to generate multiple second sample images of amplification；

Multiple second sample images before amplification and after amplification are supplied to the deep learning system, with special in the image Levy one second appreciative standard established in automatic identification algorithm and be directed to second image feature；And

Capture the testing image of the object to be detected, with the automatic identification algorithm of the image feature to the testing image into Row analysis；Wherein, judge whether the testing image has first image feature according to first appreciative standard；Its In, judge whether the testing image has second image feature according to second appreciative standard.

22. according to the method for claim 16, which is characterized in that before executing described image processing routine, first carry out one Pre-testing program, the Pre-testing program include:

Set one first Pre-testing threshold value；And

Judge whether the standard picture and multiple first preliminary examination images have according to the first Pre-testing standard First image feature records the first Pre-testing standard to the one first of first image feature according to judging result Pre-testing accuracy；

Wherein, when the first Pre-testing accuracy is lower than the first Pre-testing threshold value, increasing has different first The quantity of first sample image of image feature；And