KR101895873B1 - Method and apparatus for fabric inspection - Google Patents

Abstract

The present invention relates to a method for fabric inspection which undergoes a process of acquiring a surface image and then preprocessing the surface image, extracts features from the image to identify defective elements by the extracted features, and extracts the features and determines the detective elements by a value determined in a process of testing a sample case and a value obtained by a convoluted neural network. Also, the present invention relates to an apparatus for fabric inspection. The apparatus for detecting a defect of fabric manufactured by a continuous rolling process comprises: a light source to emit a light to the fabric; a CCD camera arranged on a side area of the fabric to photograph a surface of the fabric to which the light is emitted to supply a photographed image; an image processor to preprocess the image from the CCD camera, and detect a defect on the surface of the fabric from the preprocessed image; and a convolution neural network to digitize the detected defect to upgrade the defect to determination data for feature extraction and defective elements.

Description

[0001] METHOD AND APPARATUS FOR FABRIC INSPECTION [0002]

BACKGROUND OF THE INVENTION 1. Field of the Invention [0001] The present invention relates to a method and apparatus for testing fabric fabrics, and more particularly, to a method and apparatus for automatically detecting surface defects of fabric fabrics.

Fabric inspection process is a process to judge whether a fabric is bad or not, and it is a process that occupies a very important position in the textile and sewing industry. Especially, as the market demand for high value added fabrics such as functional and technical textiles has increased recently, it is important to select defective fabrics in the section before the fabrication of the fabric and input defective parts in the fabric, . The primary purpose of this inspection process is to prevent defects in the finished product by putting it into the production process except for the defective part during the cutting process. Failure to select the defects prior to the input of the production process will result in great economic loss to both the dealer and the producer. According to the study, the price drop of the product due to the defective fabric is 55 ~ 65% compared to the normal product (if the normal value is 10,000 won, it is 3,500 ~ 4,500 won for the defective product) If you consider non-monetary damages such as cost, decline in brand value, customer experience, etc., the weight of fabric inspection in the textile and sewing industry is big.

At present, most of the defects are detected by the naked eye by using a detection unit by sampling a part of the fabric. When a person detects a defect, there is a case where the defect can not be found due to fatigue, concentration loss, carelessness, etc. Even if a skilled worker performs work according to a subjective standard, There is a possibility that the standard is inconsistent. In addition, since it has to depend on the naked eye, it is a task that is easily tired due to the nature of work to be concentrated for a long time under bright light and is difficult to maintain for a long time. As a result, it is hard to find a test manpower in a professional dredging company because it is hard to work and the work environment is not good. Therefore, it is a reality that a sewing factory is less accurate when it comes to inspecting a dummy workforce.

As the interest in automated inspection technology grows, researches on improved system than manual inspection method in terms of accuracy, stability, and inspection speed are being conducted. This automation technology can reduce labor costs and improve the quality of finished products. However, the development of an automated inspection system remains a problem because of high hardware costs, software development costs, and the need for a large amount of computing power. In addition, there is still a limit to detect defects by combining various kinds of fabrics and products other than fabrics.

Prior Art Document: Korean Patent Registration No. 10-1453522

Accordingly, a problem to be solved by the present invention is to provide a cloth fabric inspection method and apparatus using a Convolution Neural Network (CNN) in order to detect defects of fabric fabrics, which have better accuracy, stability and inspection speed And also to provide a surface inspection method and apparatus capable of combining various types of fabric fabrics and fabrics other than printed fabrics or semiconductor fabrics.

In order to accomplish the above object, in the present invention, after acquiring an image of a surface, a preprocessing process is performed, and a feature is extracted from an image, Provides a method for testing fabric fabrics by means of a value determined during testing the case and a value obtained by a convolutional neural network (CNN).

According to an embodiment of the present invention, the image acquisition of the surface is performed by irradiating a selected area of the surface with a light source to transmit the light through the selected sample area or capturing the reflected light to image the fabric. to provide.

According to an embodiment of the present invention, the feature extraction in the image is performed by determining a target spot in the arrays in the captured image, calculating the displacement in the target spot plane (XY), the width of the target spot, , And applying one or more values among the intensities of the target spots.

According to an embodiment of the present invention, the determination of the defective element may be performed by applying a predetermined value in the course of testing the sample case, or by upgrading the data by the CNN, And a value to be accumulated in this manner is applied.

According to another aspect of the present invention, there is provided an apparatus for detecting a defect in a fabric produced through a continuous rolling process, the apparatus comprising: a light source for emitting light to the fabric; An image processor for preprocessing an image from the CCD camera and detecting a defect occurring on the surface of the fabric from the preprocessed image, And a convolutional neural network for upgrading the feature extraction data and the defective element determination data.

According to an embodiment of the present invention, the light source includes a DC light source for generating DC light as a direct current (DC) light source for irradiating a predetermined light with respect to time, and a DC light source for transmitting DC light generated in the DC light source There is provided an apparatus for inspection of fabric fabrics including an optical fiber cable composed of a plurality of strands and a light distribution device for evenly distributing a plurality of strands of the optical fiber cable.

According to an embodiment of the present invention, the CCD camera provides a detection device for a fabric fabric positioned on a horizontal line coinciding with the center of the fabric and movable as desired.

According to an embodiment of the present invention, there is provided a cloth fabric inspection apparatus in which pattern light sources are additionally arranged. In this case, the pattern light source may be a conventional one, but when it is possible to detect a minute defect as in the present invention, it is preferable to apply a pattern light source with a better pattern. In the present invention, a liquid crystal panel including a liquid crystal sheet, a polarizing film and a color filter, and a plurality of liquid crystal panels provided on the rear surface of the liquid crystal panel and having optical characteristics controlled by independent electrical signals, Emitting quantum dot light sources.

According to an embodiment of the present invention, there is provided a device for inspecting a fabric fabric having a display tape for indicating the position of the defective element.

The fabric tailoring method and the inspection apparatus of the present invention have the following effects.

1. As factory automation, processes that require skilled workers can be automated and smart factoryized.

2. Failure to detect defective fabric in the preliminary inspection process will increase the likelihood that defects will continue to the finished product. According to the present invention, such a weakness can be improved. In addition, it is possible to detect defects such that a human can not detect them with the naked eye, and it can be verified several times with a naked eye or a patterned light source.

3. With the application of AI, continuous data accumulation can be performed and better defect detection can be performed.

4. It can be applied not only to fabric but also to whole industries such as printed matter, semiconductor, electronic equipment, construction materials.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a flow chart briefly showing a method of checking the fabric of the present invention.
Fig. 2 is a schematic view briefly showing the configurations of a fabric inspection apparatus of the present invention.

Hereinafter, the method and apparatus of the present invention will be described in detail with reference to a textile fabric (which can be applied to printed matter, semiconductor, electronic equipment, construction materials, etc.) for convenience of explanation.

First, in the present invention, the image of the surface of a target product is acquired and then subjected to a preprocessing process.

The light source is irradiated onto the selected area of the transferred original surface to transmit the selected sample area or capture the reflected light to store the image in the memory. At this time, the light source can selectively use halogen, LED, high frequency fluorescent lamp and commonly used illumination, and it is possible to selectively use the illumination angle of the image to be sharp according to the kind and texture of the fabric (other printed matter, semiconductor, electronic equipment, Since the absorption and reflection angles of the light source are different, it is possible to set the angle at which the image is kept sharp while slowly changing the irradiation angle at the initial setting by inputting the fabric tail. Temporarily stored images are pre-processed to remove noise and distortion. Median filtering method is used to remove fine noise, and histogram equalization method is used to adjust unstable lighting part. Through this process, the image that removes the obstacles to defect detection is used in the next step.

Next, in the present invention, the features are extracted from the obtained image, and the defective elements are grasped by the extracted features.

The feature extraction in the obtained image is performed by determining the target spot in the arrays in the captured image and calculating the displacement in the target spot plane XY, the width of the target spot, the length of the target spot, Contrast, color, etc.). The determination of the presence or absence of a defective element of the extracted feature is initially performed by comparing the measured value with a value determined in the course of testing the sample case and then compared with a value obtained by continually upgrading the data determined to be defective by the CNN And detects whether or not it exists. That is, if the defect element is detected, the defect data corresponding to the defect element is collected and applied to the criterion of the defect element determination. The existing automatic inspection system is mainly used statistical classification method. However, these statistical methods have a disadvantage in that human subjective judgment plays an important role in defining the characteristics, and therefore, how the characteristics are defined properly and in a balanced manner greatly affects the result.

In the present invention, a CNN (Convolution Neural Network) is used to extract features corresponding to various samples. This technology can be applied to the QC process of the entire manufacturing industry that produces an automated flow as well as a printing industry in which the product is produced in the form of a sheet or a roll like a fabric. The application of CNN is infinitely expanding the capability of detecting defects because it involves data from other textile companies or other industrial sites. The application of CNN can be applied to the classification of materials in fabrics or other industries It is also easy to manage the data of each product, one by one, of each type of fabric. It can be applied to various industries such as printing, semiconductors, electronics, construction as well as sewing and textile industry by securing a technology to enable computer to classify by real-time analysis by applying deep learning such as CNN.

For example, but not limited to, data that can be read by a graphics processing unit ("GPU"), and a computer implemented method for training a convolutional neural network to identify an image using one or more GPU run programs, Wherein the at least one texture comprises at least one layer comprising a texture having two-dimensional addressing and at least one texture comprising a plurality of patches, Performing at least one of the GPU run programs in a GPU to perform a forward path in a circuit neural network, the run step being indicative of a neural network variable having two-dimensional or more addressing in which the circuit neural network is smoothed to two-dimensional addressing, ≪ / RTI > And performing at least one of the GPU run programs in a GPU to execute a reverse path in a circuit neural network, the execute step comprising performing a convolution operation on the patch, Executing one or more of the GPU run programs in the GPU to modify the line network neural network and modifying the forward path backward path and graphical data by performing one or more of the GPU run programs until the line neural network is trained May include repeating steps.

As an example, the address of a value appearing in a texture with two-dimensional addressing can be determined by linear combination of the x and y offset coordinates of the upper left corner of the output texture. The GPU executable program may be written in the form of one or more pixel shader programs. The graphical data may be adjusted using a slope descent. The one or more programs may use a functional equation to determine a gradient. Equations can be combined and simplified logarithmically to reduce calls to the pixel shader program.

Where the neural network may include one or more fully connected layers and one or more GPU run programs may include one or more GPU run programs specific to one or more fully connected layers using separate formulas for fully connected layers . The neural network may include one or more transition layers and may include one or more GPU run programs specific to one or more transition layers using separate formulas for one or more GPU run program migration layers.

One or more GPU run programs include one or more totals, and one or more totals are each divided into multiple paths. The circuit neural network can perform handwriting recognition.

The present invention may further comprise generating one or more computer readable media containing data describing the circuit network trained through this preliminary process.

Further, for example, in one or more computer-readable storage media including instructions for executing a method of training a circuit neural network to a graphics card when executed on a graphics card, the plurality of textures may include at least a portion of the texture, Receiving a plurality of textures that at least partially represent a quadrangle circuit kernel of the neural network representing a quadrangle circuit kernel having the two dimensional or more addressing and a two dimensional addressing designation step; Calculating a plurality of forward paths of neural networks of different input data including directional line kernel lines and subsampling, calculating a reverse path of the neural network using a slope function for each of the plurality of forward paths, And modifying the information contained in the quadrangle line kernel in the plurality of textures to influence the neural network training according to the result of the slope function.

Here the neural network may be trained to recognize handwriting. The plurality of textures may at least partially represent the line kernel, and the line kernel may operate on the input data representing the handwriting. The plurality of textures represent at least partially fully connected neural network levels and transition levels.

On the other hand, it is preferable to display the far end section in which the defective element is found by an electric signal, a display tape, or the like.

Hereinafter, preferred embodiments applicable to fabric fabrics among the devices to which the inspection method of the present invention can be applied will be described. 2 is a schematic view showing a flow of an example of the fabric inspection apparatus of the present invention.

First, the inspection apparatus of the present invention is provided with light sources 201 and 202 for irradiating light to a far end.

The light source can be selectively used with halogen, LED, high frequency fluorescent lamp and commonly used illumination. It is preferable to use a direct current (DC) light source that irradiates a constant light with time, if possible. An optical fiber cable consisting of a DC light source for generating DC light, a plurality of strands for transmitting the DC light generated by the DC light source, and a light distribution device for evenly distributing a plurality of strands of the optical fiber cable Is more preferable. Since a direct current (DC) light source irradiates a constant light without fluctuating time unlike an alternating current (AC) light source, it is possible to prevent a phenomenon in which an obtained image is streaked, This is possible.

Next, the inspection apparatus of the present invention is provided with a CCD camera 300 which is disposed in a lateral area of the raw material 100 and picks up the surface of the raw material irradiated with the light and supplies the sensed image. The CCD camera 300 is disposed in a side area of the raw fabric 100 and picks up a side surface of a raw material irradiated with light. In addition, the CCD camera 300 is preferably arranged so as not to capture the reflected light reflected by the light source, and may be operated by adjusting the angle. At this time, the angle of the light can be adjusted, and a desired image can be obtained by combining them. The CCD camera 300 may be a line CCD camera 300 and the light source 201 and the CCD camera 300 may be integrated. When the light source 201 and the CCD camera 300 are integrally formed, it is possible to prevent the position irradiated by the light source due to the external influence from fluctuating. The CCD camera 300 is preferably positioned on a horizontal line that coincides with the center of the raw fabric 100, and is movable when desired.

Next, the inspection apparatus of the present invention includes an image processor for pre-processing an image from the CCD camera and detecting a defect occurring on the surface of the raw material from the preprocessed image. The image processor is used for pre-processing to remove noise and distortion from a raw image acquired at high resolution. To remove fine noise, an intermediate value filtering method is used. To adjust an unstable illumination part, a histogram smoothing method is used. Through this process, the image that removes the obstacles to defect detection is used in the next step. The features extracted from the preprocessed image are extracted and images similar to the defective elements of the fabric are extracted based on the generated feature vectors. If the definition of the feature is correct, the success rate that can successfully distinguish the defect becomes higher. This feature is determined during the testing of the sample case and is considered bad when it exceeds the predefined thresholds.

Next, the inspection apparatus of the present invention comprises a convolutional neural network for converting the detected defects into data and upgrading them to feature extraction and defective factor determination data. The definition of the predefined reference value is inevitably subjective, and there is a possibility that the foreign matter on the fabric, shading due to illumination, and stain may be recognized as defective. To reduce this possibility, the present invention applies a convolutional neural network. Convolutional neural networks can be applied to learn, and as the data accumulates, the possibility can be reduced.

For example, in the present invention, a graphics processing device configured to perform a method of training a handwriting recognition circuit neural network may be configured such that the line neural network includes one or more layers, at least a portion of which includes a plurality of quadrature line kernel patches, The processing device may be one having a data storage device configured to store one or more graphics textures representing at least a portion of the graphics texture having a two-dimensional addressing and a square line kernel patch with addressing. A plurality of pixel shader units configured by cell shader programming, wherein the graphic texture is flattened in two-dimensional addressing describing a quadrature line kernel patch of the handwriting recognition neural network, a neural network iterative forward path and a backward path in handwriting input data Wherein the path comprises performing a convolution operation on a square line kernel patch; The results can be stored in multiple graphical textures and the line kernel patches of multiple texture squares can be modified to result in forward and backward paths to train the neural network.

Such a handwriting recognition neural network may include at least partially one line level and one fully connected level. The one or more graphics textures must be configured to include a simplified picture such that all processing performed by the graphics processing unit requires computation for a portion of the pixel shader unit. The pixel shader unit may be configured such that the sum of the forward path and the backward path of the neural network is divided into several smaller layers.

A computing environment employing deep running, such as CNN, may have additional functionality (not shown) such as, for example, storage, one or more input devices, one or more output devices, and one or more communication connections. Interconnecting mechanisms (not shown), including buses and controllers, can interconnect the computing environment. Typically, operating system software (not shown) provides a computing environment (providing an operating environment for other software in execution) and coordinates the operation of computing environment components. The storage may be removable or non-removable and may include magnetic disks, magnetic tape, cassettes, CD-ROMs, CD-RWs, DVDs, or any other medium that can be used and can be used to store information have. The storage stores instructions for software implementing the neural network training techniques. The input device (s) may be a keyboard, a mouse, a pen, or a touch input device, such as a trackball, a voice input device, a scanning device, or other device that provides input to a computing environment. In the case of audio, it may be a sound card or similar device that accepts a CD-ROM reader audio input that provides audio samples to the input device (s) in analog or digital form, or in a computing environment. The output device (s) may be a patterned light source, a printer, a speaker, a CD writer, or other device that provides output to a computing environment. The communication connection (s) enables communication via a communication medium to another computing entity. The communication medium conveys information such as computer-executable instructions, compressed audio or video information, or other data within the modulated data signal. A modulated data signal is a signal having one or more characteristics set or encoded to encode information in the signal. By way of example, communication media may be embodied in electrical, optical, RF, infrared, acoustic, or other carrier wave, including wired or wireless technologies.

Digital media processing techniques may be described herein in the general context of computer readable media. Computer readable media are any available media that can be accessed within a computing environment. As an example, in a computing environment, computer-readable media includes memory, storage devices, communication media, and the like.

In the present disclosure, neural network training techniques that may be described in the general context of computer-executable instructions, such as those included in program modules, are executed in a computing environment on a target real or virtual processor. Generally, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The functionality of the program modules may be combined as desired in various embodiments or may be partitioned between program modules. Computer-executable instructions for the program modules may be executed within a local or distributed computing environment.

The terms "apply", "adjust", "create", "determine" are used to describe computer operations in an expression computing environment. These terms are high-level abstracted terms for operations performed by a computer, and should not be confused with operations performed by a person. Actual computer operations corresponding to these terms are implementation dependent.

More specifically, it is difficult to detect defects of a fabric when it is not defective in a single color fabric. However, when a good image is input and learned, And determines whether the image is a pass or fail for a newly input image. In other words, it is possible to detect defects simply by displaying and learning the image without implementation and program for the type or characteristic of defective as in the conventional method. After learning and verification, you can call it in the form of a library, which makes it possible to develop and validate inspection software that takes several months in a half day. Further, if the learning continues, better performance can be achieved.

Next, the inspection apparatus of the present invention may selectively arrange the patterned light source 400. The fabric inspection according to the present invention can output a visual output using the patterned light source 400. However, it may be necessary to perform a better display because it can catch minute defects due to automatic approach and AI application. The light source of this pattern may be a general one in a general work, but it is more preferable to use a light source capable of improving the contrast ratio while having a high resolution. Therefore, in the present invention, a liquid crystal panel including a liquid crystal sheet, a polarizing film, and a color filter, and a plurality of quantum dot light emitting sources provided on the rear surface of the liquid crystal panel and each having an optical characteristic controlled by an independent electrical signal Use a light source. In this case, the optical characteristics of the divided regions can be independently controlled in the divided regions, and the optical characteristics of the divided regions are controlled according to the optical characteristics to be realized by the plurality of pixels of the liquid crystal panel matched to the divided regions, There is an advantage that the improvement can be implemented. The quantum dot light emitting source is formed at a position matched with a plurality of neighboring pixels of the liquid crystal panel and the electrical signal applied to the quantum dot light emitting source can be interlocked with the optical characteristics of the pixels of the liquid crystal panel at the matched position More preferable. Further, the brightness of the quantum dot light emitting source can be controlled to be proportional to the maximum brightness among a plurality of liquid crystal panels matched with the quantum dot light emitting source. It is more preferable that the quantum dot luminescent layer included in the quantum dot light emitting source is a mixture of quantum dot materials having different optical properties so that white light is realized.

Next, the inspection apparatus of the present invention is provided with a display tape for indicating the position of the defective element. If the position of the defective element is judged, the defective section of the fabric is indicated by the marking device. If a final defect is detected while the fabric is transferred, an interval in which the defect exists by the electrical signal is displayed, and this is indicated using a display tape.

While the present invention has been described in connection with what is presently considered to be practical exemplary embodiments, it is to be understood that the invention is not limited to the disclosed embodiments, but, on the contrary, . Therefore, the embodiments described in the present invention are not intended to limit the scope of the present invention but to limit the scope of the present invention. The scope of protection of the present invention should be construed according to the claims, and all technical ideas within the scope of equivalents should be construed as falling within the scope of the present invention.

100: Fabric
201: front light
202: back light
300: CCD camera
400: pattern light source

Claims (9)

In this paper, we propose a new method to extract the features of the image from the images and to extract the feature from the image. Is based on the value obtained by the convolutional neural network (CNN)
Wherein shading due to illumination is excluded from the defective element through application of the CNN. The method according to claim 1,
The acquisition of the image of the surface,
And irradiating the selected region of the surface with a light source to transmit the reflected light to the sample region where the surface is selected, and capturing the reflected light to image. The method according to claim 1,
The feature extraction in the image may include:
The target spot is determined in the arrays in the captured image and one or more of the displacement in the target spot plane (XY), the width of the target spot, the length of the target spot, and the intensity of the target spot are applied. A method of inspection of a fabric. The method according to claim 1,
In the defective element, the determination of the exclusion of shading due to illumination may be made,
The data is upgraded by the convolutional neural network (CNN) while applying the predetermined value in the process of testing the sample case or the inspection is performed by applying the values, and the value to be accumulated is applied. How to check the fabric. An apparatus for detecting defects in a fabric produced through a continuous rolling process, comprising: a light source for irradiating light to the fabric; A CCD camera disposed in a side region of the far end to pick up an image of the surface of the raw material irradiated with the light and to supply the sensed image; An image processor for preprocessing an image from the CCD camera and detecting a defect occurring on the surface of the fabric from the preprocessed image; And a convolutional neural network for converting the detected defects into data and upgrading the extracted features to feature extraction and determination of defective elements,
Wherein shading due to illumination is excluded from the defective element through application of the CNN. ≪ RTI ID = 0.0 > 8. < / RTI > The method of claim 5,
The light source includes a DC light source for generating DC light as a DC light source for irradiating a constant light according to time, an optical fiber cable composed of a plurality of strands for transmitting the DC light generated by the DC light source, ≪ / RTI > further comprising a light distribution device for evenly distributing a plurality of strands of the fabric fabric. The method of claim 5,
Wherein the CCD camera is located on a horizontal line coinciding with the center of the fabric and is movable as desired. The method of claim 5,
A patterned light source provided on the rear surface of the liquid crystal panel and including a plurality of quantum dot light emitting sources to which optical signals are respectively applied with independent electrical signals, Of the fabric. The method of claim 5,
And a display tape for indicating the position of the defective element.

Images