CN107153822A - A kind of smart mask method of the semi-automatic image based on deep learning - Google Patents
A kind of smart mask method of the semi-automatic image based on deep learning Download PDFInfo
- Publication number
- CN107153822A CN107153822A CN201710355628.7A CN201710355628A CN107153822A CN 107153822 A CN107153822 A CN 107153822A CN 201710355628 A CN201710355628 A CN 201710355628A CN 107153822 A CN107153822 A CN 107153822A
- Authority
- CN
- China
- Prior art keywords
- mark
- image
- deep learning
- classification
- semi
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24133—Distances to prototypes
- G06F18/24137—Distances to cluster centroïds
- G06F18/2414—Smoothing the distance, e.g. radial basis function networks [RBFN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a kind of smart mask method of semi-automatic image based on deep learning, its main flow is：Image is marked in advance using deep neural network model, again to the pre- automatic processing of mark figure, much noise point is removed, finally transfers to mark librarian use particular tool rapidly to find and correct the defect in pre- mark figure, so as to reach the purpose of reduction label time.It was verified that this method significantly reduces mark burden, the purer manual method of processing efficiency improves more than 3 times.
Description
Technical field
It is more particularly to a kind of based on the semi-automatic of deep neural network the invention belongs to intelligent transportation and field of image recognition
Image essence mask method.
Background technology
Existing pixel-level image mask method be all it is pure be accomplished manually, start from scratch and original image be labeled, this
Need to expend substantial amounts of manpower and financial resources.Well-known traffic scene data set Cityscapes uses purely manual marking software
LabelMe carries out Pixel-level mark, and extremely inefficient, the process average such as mark, examination ＆ verification, amendment of every image needs cost 1.5
Hour, although therefore it is costly, Cityscapes has also only produced 5000 essence mark view data, and its quantity reaches far away
The standard of actual use.Similar data sets such as KITTI, CamVid etc. are also built with completely manual mode, are faced with
The predicament that efficiency is low, cost is high.
The content of the invention
The technical problems to be solved by the invention：The deficiencies in the prior art are overcome to be based on deep neural network there is provided one kind
The smart mask method of semi-automatic image, solve the problem of smart annotating efficiency of pixel-level image is extremely low, cost is high, significantly reduce
Mark burden, the purer manual method of processing efficiency improves more than 3 times.
Technical scheme to be solved by this invention：
1st, the smart mask method of a kind of semi-automatic image based on deep learning, it is characterised in that comprise the following steps：
(1) image is marked in advance using based on deep learning image, semantic parted pattern；
(2) pre- mark figure is processed automatically, removes noise spot；
(3) algorithm of target detection based on deep learning is run to original image, obtains target classification and positional information, root
The target classification and positional information obtained according to detection, has in the range of target location and enters one to removing the pre- mark figure after noise spot
Step amendment segmentation errors；
(4) fine processing is carried out using correction software, in image stacking, profile is highlighted, the translucent backman of segmentation figure
Tool rapidly finds and corrects the defect in pre- mark figure, obtains final mark figure.
Wherein, in the step (2), noise spot is removed using sliding window variable element filter algorithm, is implemented as follows：
(1) n gray value is equably chosen between 0-255 as the definition of class label, wherein n is classification number；
(2) corrosion expansion process is carried out to pre- mark figure.The final classification of final each pixel is given by：
C=argmini|Pnew-P[i]|
Wherein, c is class number, and Pnew represents the gray value after corrosion expansion, and P [i] represents the gray scale of i-th of classification
Value.
Wherein, a series of auxiliary mark personnel are given quickly discovery and the tool set of corrective pitting in the step (4),
Its general workflow is as follows：
(1) original image is laminated with pre- mark image；
(2) pre- mark image is set to translucent, its transparency is adjustable；
(3) to be automatically performed profile according to the semantic region provided in pre- mark image highlighted；
(4) mark personnel complete semantic tagger using each edges of regions of instrument amendment such as paintbrush, magic rods.
The method that the present invention is applied to has：
(1) image, semantic parted pattern SegNet
SegNet is that the one kind proposed based on full volume machine neural network (FCN) encodes-decoded structure end to end, encodes net
Original image is carried out the operation such as convolution and maximum pond by network, forms 1x1xh character representation, then at this character representation
Penalize, the semantic segmentation figure consistent with artwork size is decoded into by upper storage reservoir and deconvolution.
(2) sliding window becomes ginseng filter algorithm
First, n gray value is equably chosen between 0-255 as the definition of class label, and (n is categorical measure, such as
Target in traffic scene image is divided into 12 classes by SegNet).Then, corrosion expansion process is carried out to segmentation figure.It is last each
The final classification of pixel is given by：
C=argmini|Pnew-P[i]|
Wherein, c is class number, and Pnew represents the gray value after corrosion expansion, and P [i] represents the gray scale of i-th of classification
Value.The effect for corrode expansion to full figure using same parameter is not fully up to expectations, because in the number of different noise regions
There is larger difference in amount, so should be handled using different corrosion expansion parameters different situations.
(3) target detection model Faster R-CNN
Faster R-CNN are based on depth convolutional neural networks, are main flow target detection frameworks instantly.Mainly by region
Network (RPN), Pooling layers of ROI, classification branching networks and box Recurrent networks are elected to constitute.Faster R-CNN are real first
The model training of target detection end to end and identification based on deep learning are showed, accuracy rate is higher, and speed has reached quasi real time.
(4) manual examination and verification correction software
The artwork of road travel scene also has three below subject matter after being handled by automatic marking part：
There is noise spot, objective contour inaccurate and there are undefined category regions.For these three problems, a set of software is developed auxiliary
Work of helping others further audits amendment.
Developing instrument mainly provides both sides function, and one is to provide effective comparison function, enables mark auditor
Enough convenient contrast artworks and preprocessing figure, there is the place of segmentation errors in quick discovery and positioning；Two are to provide and easily repair
Positive instrument, makes mark auditor after discovery and locating segmentation errors, can be with most easy operation amendment mistake.Mark is examined
Core personnel utilize the two class major functions that instrument is provided, and thick mark knot is corrected for the particular problem occurred in thick annotation results
Really.
The advantage of the present invention compared with prior art is：
(1) present invention proposes a series of practical Method and kit fors to aid in mark personnel to complete complicated Pixel-level figure
As essence mark.The existing image, semantic parted pattern SegNet based on deep learning has been used to mark image in advance.By
In model Shortcomings itself, substantial amounts of noise spot is generated in pre- mark image, and these noise spots are distributed not in figure
, a kind of sliding window variable element filter algorithm is then proposed, correct segmentation is had substantially no effect on while noise spot is removed
Region.So far, the pre- mark figure that an accuracy rate is higher, globality is stronger has been obtained, a set of practical instrument is provided again
Collection so that mark personnel can be quickly handled pre- mark figure, completes mark task.
(2) in data set Cityscapes process of construction, the essence mark of an image is completed, audited, corrected and waited
Journey, average cost 1.5 hours, although therefore it is costly, Cityscapes has also only produced 5000 essence mark picture numbers
According to its data volume reaches far away the standard of actual use.Using the series methods of the present invention, this time can foreshorten to 20 minutes
Left and right.
Brief description of the drawings
Fig. 1 is the implementation process figure of the inventive method；
Fig. 2 is SegNet network structures；
Fig. 3 is the typical problem in preprocessing figure；
Fig. 4 chooses misrecognition noise spot, and noise is revised as into correct pixel value；
Fig. 5 is the part design sketch marked in advance.
Embodiment
As shown in figure 1, the smart mask method of semi-automatic image of the present invention based on deep learning, is realized by following steps：
1.SegNet is slightly marked
The SegNet that the present invention is used is that Cambridge proposes to aim to solve the problem that the image of automatic Pilot or intelligent robot
Semantic segmentation depth network, open source code, based on caffe frameworks.SegNet network structures are as shown in Fig. 2 Input schemes for input
Piece, the image that Output is split for output, different colours represent different classification.It is a symmetrical network, by centre
Pooling layers are used as segmentation with upsampling layers, and high dimensional feature is extracted by convolution, and picture is diminished by pooling,
Again by deconvolution and upsampling, by deconvolution so that feature is reappeared after image classification, upsampling makes figure
It is big as becoming, finally by Softmax, export the maximum of different classifications.
2. the merger of variable element outlier and corrosion based on sliding window expand
The segmentation figure of SegNet outputs represents 12 object classifications, such as road surface, road sign, people's row with 12 kinds of different colors
Road, building and cycling personnel etc., it can be found that the uneven situation of a large amount of outliers and edge, this is very not from segmentation figure
Beneficial to artificial further processing, so devising a change ginseng filter algorithm based on sliding window.
First, 12 gray values are equably chosen between 0-255 as the definition of class label.Then, to segmentation figure
Carry out corrosion expansion process.The final classification of last each pixel is given by：
C=argmini|Pnew-P[i]|
Wherein, c is class number, PnewThe gray value after corrosion expansion is represented, P [i] represents the gray value of i-th of classification,
The effect for corrode expansion to full figure using same parameter is not fully up to expectations, because the quantity in different noise regions is deposited
In larger difference, so should be handled using different corrosion expansion parameters different situations.
3. and the combination of target detection
To original image operational objective detection algorithm Faster R-CNN, after the rectangle frame for obtaining mark object, to rectangle
Denoising, Nogata are filtered in the range of frame, and conventional method is pre-processed in a balanced way.
Meanwhile, there is the output of target detection network, can aid in correcting the pre- mark figure that SegNet networks are provided.Specifically
Ground, if scene is understood in the output figure of network, except to rectangle frame in addition to region occur in that same category of object, then
It is considered as wrong classification, is regarded as background process.So do is highly significant, such as in automatic Pilot scene, if nothing
People drives the vehicle that front side has misrecognition, then it no longer will move ahead or make non-essential avoidance decision-making.
In addition, to reach this effect, the recall rate of target detection network should be improved as much as possible.Therefore, should be by
The threshold value of Faster RCNN final outputs is turned down as far as possible, and experiment shows, is adjusted to 0.6 proper.
4. manual examination and verification correction software
The artwork of road travel scene also has three below subject matter after being handled by automatic marking part：
There is noise spot, objective contour inaccurate and there are undefined category regions.As shown in figure 3, sight that can be apparent in figure
Observe, road surface part has the noise spot of misrecognition；The profile of target vehicle is not exactly accurate compared with artwork；Billboard portion
Divide due to not making definition in network, so also and unidentified.
For these three subject matters in thick annotation results, following instrument indirect labor is designed and developed and has further audited
Amendment.
Developing instrument mainly provides both sides function, and one is to provide effective comparison function, enables mark auditor
Enough convenient contrast artworks and preprocessing figure, there is the place of segmentation errors in quick discovery and positioning；Two are to provide and easily repair
Positive instrument, makes mark auditor after discovery and locating segmentation errors, can be with most easy operation amendment mistake.Mark is examined
Core personnel utilize the two class major functions that instrument is provided, and thick mark knot is corrected for the particular problem occurred in thick annotation results
Really.
Using the major function of above-mentioned developing instrument collection, done for the particular problem occurred in thick annotation results such as Fig. 4
Translucentization is operated.Artwork and preprocessing figure, in different figure layers, are presented simultaneously with different transparencies.In order to quickly send out
Now with the particular location of locating segmentation problem, it is necessary to set up two figure layers, while artwork and corresponding preprocessing image is presented, and
Transparency is adjusted, makes mark auditor while seeing two figures, intuitively the difference of two figures of contrast, quick positioning question institute
.
For there is noise spot in preprocessing figure：The noise spot part can be chosen using magic wand menu, it is determined that
The corresponding correct classification of the noise spot, by the instrument that changes colour, correct classification correspondence is revised as by the pixel value for misidentifying noise spot
Pixel value.Or one piece of region comprising noise spot is chosen using magic wand menu, selection carries out the corrosion of different operator sizes
Expansive working, removes noise spot.
For the inaccurate situation of objective contour in preprocessing figure：Choose the part of this in segmentation figure overall using magic wand menu
Region, extracts region contour, by being contrasted with artwork, the inaccurate place of quick positioning profile；Using adding and subtracting instrument, according to
The accurate modification region profile of artwork, complies with artwork actual conditions.During amendment, the back of the body is revised as in the part that will be deleted
The pixel value of scape target, will increased part be revised as the corresponding pixel value of the regional aim.
For there is the situation of classification target undefined in semantic segmentation network：, can according to residing region actual conditions
So that by marking Category criteria of the auditor according to formulation, the pixel value corresponding to new definition classification and classification utilizes magic rod work
Tool and drawing pencil describe objects' contour, choose target corresponding region and are revised as the actual pixel value of definition.
Finally the segmentation result audited after the completion of correcting is preserved, final segmentation result is used as.
By above-mentioned primary operational, noise spot problem present in preprocessing figure, the inaccurate problem of profile is present undefined
New category problem be solved, the Accurate Segmentation result finally obtained is used directly for grinding for automatic Pilot algorithm
Studying carefully can be used for the further training that scene understands network.
5. target detection amendment is tested
First, on the premise of at utmost recall rate is ensured, using training/checking collection of KITTI data sets, instruct respectively
Practice YOLO, SSD and Faster-RCNN, its result as shown in table 1 (only make rough contrast herein, only contrast three primary categories,
And do not repartition easy, moderate and hard for each classification).
Then, participated in using the Faster R-CNN trained on KITTI in scene understanding task, according to advance reality
Existing algorithm flow, finds after the processing of this step, and significant changes, which do not occur, for the mIoU of segmentation result (changes ± 0.1
Between), but really can correct some obvious misclassification situations.
Performances of table 1YOLO, Faster R-CNN and SSD on KITTI train/validation, evaluation index is
MAP, experiment GPU are NVIDIA TITAN X (12GB).It is seen that, Faster R-CNN best performance (is not considering fortune
In the case of the row time).
6. the experiment of thick annotation results combining target detection
The segmentation result obtained by SegNet networks is directly calculated into Average Accuracy, result 1 is obtained；
Primary segmentation result is obtained by SegNet networks, corrosion expansion is further carried out and goes outlier to handle, then
Combining target detects web results, to same image, if outside the rectangle frame of given definite object, occur in that same again
The object of classification, then be regarded as the classification of mistake, this partial pixel be revised as to the value of background pixel, result 2 is finally obtained,
Calculate Average Accuracy.The accuracy rate of all categories such as table 2 that three operations are obtained.
Table 2 marks experimental result in advance
| Class | SegNet | 3.5K dataset training | Our method |
| Building | 88 | 73.8 | 78.7 |
| Tree | 87.3 | 90.7 | 92.1 |
| Sky | 92.3 | 90.1 | 93.8 |
| Car | 80 | 83 | 86.8 |
| Sign-symbol | 29.5 | 83.9 | 86.4 |
| Road | 97.6 | 95.21 | 96.3 |
| Pedestrian | 57.2 | 86.8 | 90.2 |
| Fence | 49.4 | 68 | 70.1 |
| Column-Pole | 27.8 | 74.6 | 80.2 |
| Side-walk | 84.8 | 95.3 | 95.4 |
| Bicyclist | 30.7 | 53 | 59 |
| Class avg. | 65.9 | 81.3 | 86.2 |
| Global avg. | 88.6 | 86.8 | 90.9 |
| Mean I/U | 50.2 | 69.1 | 70.5 |
Fig. 5 is the part design sketch marked in advance, is respectively from left to right：Original graph, SegNet results, after semi-automatic processing
Result, legitimate reading.
7. final efficiency comparative
In order to contrast artificial mark with set forth herein semi-automatic label technology efficiency, design following simple experiment checking.
Prepare two groups of equal images of quantity (each 100).One of which according to set forth herein semi-automatic label technology
Flow, carries out semantic segmentation with SegNet first, in conjunction with target detection and traditional images processing method, by the thick mark drawn
As a result artificial further check and correction is delivered, final result is obtained.Another set directly delivers artificial mark.Because of the dry of people
In advance, the accuracy rate of two kinds of notation methods in theory is all 100%, so only comparing label time.Experimental result is as follows：
The annotating efficiency of table 3 is contrasted
In a word, the present invention is marked in advance using deep neural network model to image, then to the pre- automatic processing of mark figure,
Much noise point is removed, finally transfers to mark librarian use particular tool rapidly to find and correct the defect in pre- mark figure,
So as to reach the purpose of reduction label time.It was verified that the present invention significantly reduces mark burden, processing efficiency is purer artificial
Method improves more than 3 times.
Claims (3)
1. a kind of smart mask method of the semi-automatic image based on deep learning, it is characterised in that comprise the following steps：
(1) image is marked in advance using based on deep learning image, semantic parted pattern；
(2) pre- mark figure is processed automatically, removes noise spot；
(3) algorithm of target detection based on deep learning is run to original image, target classification and positional information is obtained, according to inspection
The target classification and positional information measured, has in the range of target location and is further repaiied to removing the pre- mark figure after noise spot
Positive segmentation errors；
(4) fine processing is carried out using correction software, in image stacking, profile is highlighted, the translucent aid of segmentation figure is fast
Find fastly and correct the defect in pre- mark figure, obtain final mark figure.
2. the smart mask method of a kind of semi-automatic image based on deep learning according to claim 1, it is characterised in that：Institute
State in step (2), noise spot is removed using sliding window variable element filter algorithm, is implemented as follows：
(1) n gray value is equably chosen between 0-255 as the definition of class label, wherein n is classification number；
(2) corrosion expansion process is carried out to pre- mark figure, the final classification of final each pixel is given by：
C=argmini|Pnew-P[i]|
Wherein, c is class number, PnewThe gray value after corrosion expansion is represented, P [i] represents the gray value of i-th of classification.
3. the smart mask method of a kind of semi-automatic image based on deep learning according to claim 1, it is characterised in that：Institute
State and give a series of auxiliary mark personnel quickly discovery and the tool set of corrective pitting, its general workflow in step (4)
It is as follows：
(1) original image is laminated with pre- mark image；
(2) pre- mark image is set to translucent, its transparency is adjustable；
(3) to be automatically performed profile according to the semantic region provided in pre- mark image highlighted；
(4) mark personnel complete semantic tagger using each edges of regions of instrument amendment such as paintbrush, magic rods.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710355628.7A CN107153822A (en) | 2017-05-19 | 2017-05-19 | A kind of smart mask method of the semi-automatic image based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710355628.7A CN107153822A (en) | 2017-05-19 | 2017-05-19 | A kind of smart mask method of the semi-automatic image based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107153822A true CN107153822A (en) | 2017-09-12 |
Family
ID=59794324
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710355628.7A Pending CN107153822A (en) | 2017-05-19 | 2017-05-19 | A kind of smart mask method of the semi-automatic image based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107153822A (en) |
Cited By (38)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108022243A (en) * | 2017-11-23 | 2018-05-11 | 浙江清华长三角研究院 | Method for detecting paper in a kind of image based on deep learning |
| CN108564587A (en) * | 2018-03-07 | 2018-09-21 | 浙江大学 | A kind of a wide range of remote sensing image semantic segmentation method based on full convolutional neural networks |
| CN108573279A (en) * | 2018-03-19 | 2018-09-25 | 精锐视觉智能科技（深圳）有限公司 | Image labeling method and terminal device |
| CN108595544A (en) * | 2018-04-09 | 2018-09-28 | 深源恒际科技有限公司 | A kind of document picture classification method |
| CN108830466A (en) * | 2018-05-31 | 2018-11-16 | 长春博立电子科技有限公司 | A kind of image content semanteme marking system and method based on cloud platform |
| CN108875769A (en) * | 2018-01-23 | 2018-11-23 | 北京迈格威科技有限公司 | Data mask method, device and system and storage medium |
| CN108985293A (en) * | 2018-06-22 | 2018-12-11 | 深源恒际科技有限公司 | A kind of image automation mask method and system based on deep learning |
| CN108985186A (en) * | 2018-06-27 | 2018-12-11 | 武汉理工大学 | A kind of unmanned middle pedestrian detection method based on improvement YOLOv2 |
| CN109002840A (en) * | 2018-06-26 | 2018-12-14 | 北京纵目安驰智能科技有限公司 | One kind being based on cascade semantic segmentation method, system, terminal and storage medium |
| CN109190631A (en) * | 2018-08-31 | 2019-01-11 | 阿里巴巴集团控股有限公司 | The target object mask method and device of picture |
| CN109255044A (en) * | 2018-08-31 | 2019-01-22 | 江苏大学 | A kind of image intelligent mask method based on YOLOv3 deep learning network |
| CN109255294A (en) * | 2018-08-02 | 2019-01-22 | 中国地质大学（北京） | A kind of remote sensing image clouds recognition methods based on deep learning |
| CN109284779A (en) * | 2018-09-04 | 2019-01-29 | 中国人民解放军陆军工程大学 | Object detection method based on deep full convolution network |
| CN109377479A (en) * | 2018-09-27 | 2019-02-22 | 中国电子科技集团公司第五十四研究所 | Satellite dish object detection method based on remote sensing image |
| CN109377509A (en) * | 2018-09-26 | 2019-02-22 | 深圳前海达闼云端智能科技有限公司 | Method, apparatus, storage medium and the equipment of image, semantic segmentation mark |
| CN109409248A (en) * | 2018-09-30 | 2019-03-01 | 上海交通大学 | Semanteme marking method, apparatus and system based on deep semantic network |
| CN109446369A (en) * | 2018-09-28 | 2019-03-08 | 武汉中海庭数据技术有限公司 | The exchange method and system of the semi-automatic mark of image |
| CN109670060A (en) * | 2018-12-10 | 2019-04-23 | 北京航天泰坦科技股份有限公司 | A kind of remote sensing image semi-automation mask method based on deep learning |
| CN109685870A (en) * | 2018-11-21 | 2019-04-26 | 北京慧流科技有限公司 | Information labeling method and device, tagging equipment and storage medium |
| CN109902765A (en) * | 2019-03-22 | 2019-06-18 | 北京滴普科技有限公司 | A kind of intelligent cloud labeling method for supporting artificial intelligence |
| CN110110723A (en) * | 2019-05-07 | 2019-08-09 | 艾瑞迈迪科技石家庄有限公司 | A kind of method and device that objective area in image automatically extracts |
| CN110276343A (en) * | 2018-03-14 | 2019-09-24 | 沃尔沃汽车公司 | The method of the segmentation and annotation of image |
| CN110298823A (en) * | 2019-06-17 | 2019-10-01 | 天津大学 | A kind of infrared image auxiliary mask method based on MSER algorithm |
| CN110334772A (en) * | 2019-07-11 | 2019-10-15 | 山东领能电子科技有限公司 | A kind of quick mask method of expansion classification formula data |
| CN110570434A (en) * | 2018-06-06 | 2019-12-13 | 杭州海康威视数字技术股份有限公司 | image segmentation and annotation method and device |
| CN110674807A (en) * | 2019-08-06 | 2020-01-10 | 中国科学院信息工程研究所 | Curved scene character detection method based on semi-supervised and weakly supervised learning |
| CN110866930A (en) * | 2019-11-18 | 2020-03-06 | 北京云聚智慧科技有限公司 | Semantic segmentation auxiliary labeling method and device |
| CN111061901A (en) * | 2019-11-28 | 2020-04-24 | 中国船舶重工集团公司第七0九研究所 | Intelligent image annotation method and system and image annotation quality analysis method |
| CN111444746A (en) * | 2019-01-16 | 2020-07-24 | 北京亮亮视野科技有限公司 | Information labeling method based on neural network model |
| CN111783783A (en) * | 2020-06-18 | 2020-10-16 | 哈尔滨市科佳通用机电股份有限公司 | Annotation system and annotation method for image segmentation |
| CN111985394A (en) * | 2020-08-19 | 2020-11-24 | 东南大学 | Semi-automatic instance labeling method and system for KITTI data set |
| CN112308163A (en) * | 2020-11-09 | 2021-02-02 | 复旦大学 | Brain glioma auxiliary labeling method and device based on deep learning |
| CN112381840A (en) * | 2020-11-27 | 2021-02-19 | 深源恒际科技有限公司 | Method and system for marking vehicle appearance parts in loss assessment video |
| CN112561480A (en) * | 2020-12-16 | 2021-03-26 | 中国平安人寿保险股份有限公司 | Intelligent workflow pushing method, equipment and computer storage medium |
| CN112988733A (en) * | 2021-04-16 | 2021-06-18 | 北京妙医佳健康科技集团有限公司 | Method and device for improving and enhancing data quality |
| CN113111716A (en) * | 2021-03-15 | 2021-07-13 | 中国科学院计算机网络信息中心 | Remote sensing image semi-automatic labeling method and device based on deep learning |
| CN113468350A (en) * | 2020-03-31 | 2021-10-01 | 京东方科技集团股份有限公司 | Image annotation method, device and system |
| US11429472B1 (en) | 2021-03-26 | 2022-08-30 | International Business Machines Corporation | Automated cognitive software application error detection |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105678297A (en) * | 2015-12-29 | 2016-06-15 | 南京大学 | Portrait semantic analysis method and system based on label transfer and LSTM model |
| CN106530305A (en) * | 2016-09-23 | 2017-03-22 | 北京市商汤科技开发有限公司 | Semantic segmentation model training and image segmentation method and device, and calculating equipment |
-
2017
- 2017-05-19 CN CN201710355628.7A patent/CN107153822A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105678297A (en) * | 2015-12-29 | 2016-06-15 | 南京大学 | Portrait semantic analysis method and system based on label transfer and LSTM model |
| CN106530305A (en) * | 2016-09-23 | 2017-03-22 | 北京市商汤科技开发有限公司 | Semantic segmentation model training and image segmentation method and device, and calculating equipment |
Non-Patent Citations (1)
| Title |
|---|
| YUNCHAO WEI ETC,: ""Learning to segment with image-level annotations"", 《PATTERN RECOGNITION》 * |
Cited By (50)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108022243A (en) * | 2017-11-23 | 2018-05-11 | 浙江清华长三角研究院 | Method for detecting paper in a kind of image based on deep learning |
| CN108875769A (en) * | 2018-01-23 | 2018-11-23 | 北京迈格威科技有限公司 | Data mask method, device and system and storage medium |
| CN108564587A (en) * | 2018-03-07 | 2018-09-21 | 浙江大学 | A kind of a wide range of remote sensing image semantic segmentation method based on full convolutional neural networks |
| CN110276343A (en) * | 2018-03-14 | 2019-09-24 | 沃尔沃汽车公司 | The method of the segmentation and annotation of image |
| CN110276343B (en) * | 2018-03-14 | 2023-10-27 | 沃尔沃汽车公司 | Method for segmenting and annotating images |
| CN108573279A (en) * | 2018-03-19 | 2018-09-25 | 精锐视觉智能科技（深圳）有限公司 | Image labeling method and terminal device |
| CN108595544A (en) * | 2018-04-09 | 2018-09-28 | 深源恒际科技有限公司 | A kind of document picture classification method |
| CN108830466A (en) * | 2018-05-31 | 2018-11-16 | 长春博立电子科技有限公司 | A kind of image content semanteme marking system and method based on cloud platform |
| CN110570434B (en) * | 2018-06-06 | 2022-09-30 | 杭州海康威视数字技术股份有限公司 | Image segmentation and annotation method and device |
| CN110570434A (en) * | 2018-06-06 | 2019-12-13 | 杭州海康威视数字技术股份有限公司 | image segmentation and annotation method and device |
| CN108985293A (en) * | 2018-06-22 | 2018-12-11 | 深源恒际科技有限公司 | A kind of image automation mask method and system based on deep learning |
| CN109002840A (en) * | 2018-06-26 | 2018-12-14 | 北京纵目安驰智能科技有限公司 | One kind being based on cascade semantic segmentation method, system, terminal and storage medium |
| CN108985186B (en) * | 2018-06-27 | 2022-03-01 | 武汉理工大学 | Improved YOLOv 2-based method for detecting pedestrians in unmanned driving |
| CN108985186A (en) * | 2018-06-27 | 2018-12-11 | 武汉理工大学 | A kind of unmanned middle pedestrian detection method based on improvement YOLOv2 |
| CN109255294A (en) * | 2018-08-02 | 2019-01-22 | 中国地质大学（北京） | A kind of remote sensing image clouds recognition methods based on deep learning |
| CN109255044A (en) * | 2018-08-31 | 2019-01-22 | 江苏大学 | A kind of image intelligent mask method based on YOLOv3 deep learning network |
| CN109190631A (en) * | 2018-08-31 | 2019-01-11 | 阿里巴巴集团控股有限公司 | The target object mask method and device of picture |
| CN109284779A (en) * | 2018-09-04 | 2019-01-29 | 中国人民解放军陆军工程大学 | Object detection method based on deep full convolution network |
| CN109377509A (en) * | 2018-09-26 | 2019-02-22 | 深圳前海达闼云端智能科技有限公司 | Method, apparatus, storage medium and the equipment of image, semantic segmentation mark |
| CN109377479A (en) * | 2018-09-27 | 2019-02-22 | 中国电子科技集团公司第五十四研究所 | Satellite dish object detection method based on remote sensing image |
| CN109377479B (en) * | 2018-09-27 | 2021-10-22 | 中国电子科技集团公司第五十四研究所 | Butterfly satellite antenna target detection method based on remote sensing image |
| CN109446369B (en) * | 2018-09-28 | 2021-10-08 | 武汉中海庭数据技术有限公司 | Interaction method and system for semi-automatic image annotation |
| CN109446369A (en) * | 2018-09-28 | 2019-03-08 | 武汉中海庭数据技术有限公司 | The exchange method and system of the semi-automatic mark of image |
| CN109409248A (en) * | 2018-09-30 | 2019-03-01 | 上海交通大学 | Semanteme marking method, apparatus and system based on deep semantic network |
| CN109685870A (en) * | 2018-11-21 | 2019-04-26 | 北京慧流科技有限公司 | Information labeling method and device, tagging equipment and storage medium |
| CN109685870B (en) * | 2018-11-21 | 2023-10-31 | 北京慧流科技有限公司 | Information labeling method and device, labeling equipment and storage medium |
| CN109670060A (en) * | 2018-12-10 | 2019-04-23 | 北京航天泰坦科技股份有限公司 | A kind of remote sensing image semi-automation mask method based on deep learning |
| CN111444746B (en) * | 2019-01-16 | 2024-01-30 | 北京亮亮视野科技有限公司 | Information labeling method based on neural network model |
| CN111444746A (en) * | 2019-01-16 | 2020-07-24 | 北京亮亮视野科技有限公司 | Information labeling method based on neural network model |
| CN109902765A (en) * | 2019-03-22 | 2019-06-18 | 北京滴普科技有限公司 | A kind of intelligent cloud labeling method for supporting artificial intelligence |
| CN110110723A (en) * | 2019-05-07 | 2019-08-09 | 艾瑞迈迪科技石家庄有限公司 | A kind of method and device that objective area in image automatically extracts |
| CN110298823A (en) * | 2019-06-17 | 2019-10-01 | 天津大学 | A kind of infrared image auxiliary mask method based on MSER algorithm |
| CN110334772A (en) * | 2019-07-11 | 2019-10-15 | 山东领能电子科技有限公司 | A kind of quick mask method of expansion classification formula data |
| CN110674807A (en) * | 2019-08-06 | 2020-01-10 | 中国科学院信息工程研究所 | Curved scene character detection method based on semi-supervised and weakly supervised learning |
| CN110866930A (en) * | 2019-11-18 | 2020-03-06 | 北京云聚智慧科技有限公司 | Semantic segmentation auxiliary labeling method and device |
| CN110866930B (en) * | 2019-11-18 | 2022-04-12 | 北京云聚智慧科技有限公司 | Semantic segmentation auxiliary labeling method and device |
| CN111061901A (en) * | 2019-11-28 | 2020-04-24 | 中国船舶重工集团公司第七0九研究所 | Intelligent image annotation method and system and image annotation quality analysis method |
| CN113468350A (en) * | 2020-03-31 | 2021-10-01 | 京东方科技集团股份有限公司 | Image annotation method, device and system |
| CN111783783B (en) * | 2020-06-18 | 2021-06-04 | 哈尔滨市科佳通用机电股份有限公司 | Annotation system and annotation method for image segmentation |
| CN111783783A (en) * | 2020-06-18 | 2020-10-16 | 哈尔滨市科佳通用机电股份有限公司 | Annotation system and annotation method for image segmentation |
| CN111985394B (en) * | 2020-08-19 | 2021-05-28 | 东南大学 | Semi-automatic instance labeling method and system for KITTI data set |
| CN111985394A (en) * | 2020-08-19 | 2020-11-24 | 东南大学 | Semi-automatic instance labeling method and system for KITTI data set |
| CN112308163A (en) * | 2020-11-09 | 2021-02-02 | 复旦大学 | Brain glioma auxiliary labeling method and device based on deep learning |
| CN112381840B (en) * | 2020-11-27 | 2024-07-09 | 深源恒际科技有限公司 | Method and system for marking vehicle appearance parts in loss assessment video |
| CN112381840A (en) * | 2020-11-27 | 2021-02-19 | 深源恒际科技有限公司 | Method and system for marking vehicle appearance parts in loss assessment video |
| CN112561480A (en) * | 2020-12-16 | 2021-03-26 | 中国平安人寿保险股份有限公司 | Intelligent workflow pushing method, equipment and computer storage medium |
| CN113111716A (en) * | 2021-03-15 | 2021-07-13 | 中国科学院计算机网络信息中心 | Remote sensing image semi-automatic labeling method and device based on deep learning |
| CN113111716B (en) * | 2021-03-15 | 2023-06-23 | 中国科学院计算机网络信息中心 | Remote sensing image semiautomatic labeling method and device based on deep learning |
| US11429472B1 (en) | 2021-03-26 | 2022-08-30 | International Business Machines Corporation | Automated cognitive software application error detection |
| CN112988733A (en) * | 2021-04-16 | 2021-06-18 | 北京妙医佳健康科技集团有限公司 | Method and device for improving and enhancing data quality |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107153822A (en) | A kind of smart mask method of the semi-automatic image based on deep learning | |
| US11580647B1 (en) | Global and local binary pattern image crack segmentation method based on robot vision | |
| CN103049763B (en) | Context-constraint-based target identification method | |
| CN110263635B (en) | Marker detection and identification method based on structural forest and PCANet | |
| CN112396128B (en) | Automatic labeling method for railway external environment risk source sample | |
| CN103793708B (en) | A kind of multiple dimensioned car plate precise positioning method based on motion correction | |
| CN112488046B (en) | Lane line extraction method based on high-resolution images of unmanned aerial vehicle | |
| CN106934455B (en) | Remote sensing image optics adapter structure choosing method and system based on CNN | |
| CN113240623B (en) | Pavement disease detection method and device | |
| CN111160328B (en) | Automatic extraction method of traffic marking based on semantic segmentation technology | |
| CN101901343A (en) | Remote sensing image road extracting method based on stereo constraint | |
| CN108537782A (en) | A method of building images match based on contours extract with merge | |
| CN112613097A (en) | BIM rapid modeling method based on computer vision | |
| CN104299009A (en) | Plate number character recognition method based on multi-feature fusion | |
| CN101980317A (en) | Method for predicting traffic flow extracted by improved C-V model-based remote sensing image road network | |
| CN112819748B (en) | Training method and device for strip steel surface defect recognition model | |
| CN111754538B (en) | Threshold segmentation method for USB surface defect detection | |
| CN113409267A (en) | Pavement crack detection and segmentation method based on deep learning | |
| CN115661072A (en) | Disc rake surface defect detection method based on improved fast RCNN algorithm | |
| CN106228136A (en) | Panorama streetscape method for secret protection based on converging channels feature | |
| CN102542543A (en) | Block similarity-based interactive image segmenting method | |
| CN114299247A (en) | Rapid detection and problem troubleshooting method for road traffic sign lines | |
| CN111882573B (en) | Cultivated land block extraction method and system based on high-resolution image data | |
| CN111476890A (en) | Method for repairing moving vehicle in three-dimensional scene reconstruction based on image | |
| CN111768385B (en) | Neural network detection method for USB surface defect detection |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170912 |
|
| WD01 | Invention patent application deemed withdrawn after publication |