CN110189351A - A kind of scratch image data amplification method based on production confrontation network - Google Patents
A kind of scratch image data amplification method based on production confrontation network Download PDFInfo
- Publication number
- CN110189351A CN110189351A CN201910303496.2A CN201910303496A CN110189351A CN 110189351 A CN110189351 A CN 110189351A CN 201910303496 A CN201910303496 A CN 201910303496A CN 110189351 A CN110189351 A CN 110189351A
- Authority
- CN
- China
- Prior art keywords
- scratch
- image
- network
- scratch image
- image data
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/50—Image enhancement or restoration by the use of more than one image, e.g. averaging, subtraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/13—Edge detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
Abstract
The invention discloses a kind of scratch image data amplification methods based on production confrontation network, comprising the following steps: step S1 constructs scratch image data set;Step S2, scratch image preprocessing;Step S3, building production fight network;Step S4, dual training;Step S5 exports scratch image amplification data collection.Beneficial effects of the present invention are mainly manifested in: producing a large amount of scratch image datas using network is generated, it can learn the validity feature of true scratch image using differentiation network, by to generation network and differentiation network implementation dual training, the probability distribution statistical rule difference for generating image and true picture can effectively be reduced, to not only meet the defect detecting system based on deep learning to the quantity demand of training sample, but also meet its quality requirement to training sample.
Description
Technical field
The present invention relates to a kind of image data amplification methods, and in particular to a kind of scratch figure based on production confrontation network
As data amplification method, belong to computer vision field.
Background technique
In recent years, have benefited from the training sample of magnanimity and the calculation power of growth, deep learning is in theoretical research and multiple applications
Breakthrough, especially computer vision field are achieved in field.Deep learning simulate human brain nerve connection and
The low-level feature of high dimensional data is combined by signal processing mode, is gradually taken out by way of stratification more complicated
Character representation, but the parameter of flood tide is also brought simultaneously.In order to avoid there is over-fitting in model, need great amount of samples to model into
Row training, however, the cost collected and mark sample is very high and time-consuming.Existing image data amplification technique mainly includes
Flip horizontal, cutting and adjustment contrast etc., however, these conventional methods are very limited to the amplification quantity of image data, it is difficult
To meet the training demand of the application system (for example, defect detecting system) based on deep learning.In addition, surface defect image phase
Quantity is less for natural image, and acquisition and mark are more difficult.
Summary of the invention
The present invention is directed to the limitation of the prior art, proposes that a kind of scratch image data based on production confrontation network expands
Increasing method.
The present invention uses following technical scheme, the scratch image data amplification method packet based on production confrontation network
Include following steps:
Step S1 constructs scratch image data set;
Step S2, scratch image preprocessing;
Step S3, building production fight network;
Step S4, dual training;
Step S5 exports scratch image amplification data collection.
According to the above technical scheme, in step sl, construct scratch image data set the following steps are included:
Step S1.1 acquires product surface image by CCD camera, chooses the wherein image with scratch and is used as training
Image, the image of remaining no marking are used as the background image of subsequent composograph;
Step S1.2 carries out handmarking to scoring position, and tag image is carried out binary conversion treatment.
According to the above technical scheme, in step s 2, scratch image preprocessing the following steps are included:
Scratch image is converted to grayscale image by step S2.1;
Step S2.2, scratch positioning;
Step S2.3 cuts scored area image;
Step S2.4 zooms in and out the scratch image after cutting;
Step S2.5 carries out Random Level overturning to the scratch image after scaling.
According to the above technical scheme, in step S2.2, scratch positioning the following steps are included:
S2.2.1: contours extract is carried out to binaryzation tag image, gained profile is approximate region where scratch;
S2.2.2: Threshold segmentation is carried out in contoured interior, edge detection is carried out to scratch.
According to the above technical scheme, in step S2.3, scored area image is cut the following steps are included:
S2.3.1: traversal scratching edge all pixels point finds minimum abscissa, ordinate and the maximum of these pixels
Abscissa, ordinate, thus obtain one it is horizontal, scratch can be completely included including minimum rectangle;
S2.3.2: using rectangular centre as image cropping center, the image containing scratch having a size of 120 × 120 is cut out;
S2.3.3: the image after centering idea is cut carries out multiple random cropping, and random cropping is having a size of 100 × 100.
According to the above technical scheme, in step s3, production confrontation network includes a generation network G and a differentiation
Network D.
According to the above technical scheme, in step s 4, dual training refers to alternately excellent using stochastic gradient descent method
Change objective function (1) and objective function (2):
Wherein, LDIt is the loss function for differentiating network D,It isFrom the expectation of the probability of authentic specimen
Value,X~Pr, x is true scratch image,It is the scratch image generated by generation network G, PrAnd PgIt respectively indicates true
Real scratch image data set and the probability distribution for generating scratch image data set,It isFrom the probability of authentic specimen,
D (x) is the probability that x derives from authentic specimen, Ex[D (x)] is desired value of the x from the probability of authentic specimen, and λ isIn LDIn weight coefficient,It is the interpolation of true scratch image and generation image,ε is to obey equally distributed random number, and ε~U [0,1], U are to be uniformly distributed,It is the desired value of gradient penalty term,It isGradient,It isFrom authentic specimen
Probability, LGIt is the loss function for generating network G.
According to the above technical scheme, in step s 5, scratch image amplification data collection is by the trained generation net of step S4
Network G is generated.
Scratch image data amplification method disclosed by the invention based on production confrontation network, the beneficial effect is that:
A large amount of scratch image datas are produced using network is generated, the validity feature of true scratch image can be learnt using differentiation network,
By can effectively reduce the probability distribution for generating image and true picture to generating network and differentiating network implementation dual training
Statistical law difference, to not only meet the defect detecting system based on deep learning to the quantity demand of training sample, but also full
Its quality requirement to training sample of foot.Since method disclosed by the invention only needs a small amount of true scratch image that can generate greatly
The artificial synthesized scratch image of amount and true scratch image equal quality, therefore greatly reduce the application system based on deep learning
The demand united to true scratch image data effectively saves collection and marks the cost of sample.
Detailed description of the invention
Fig. 1 is the method for the present invention flow chart;
Fig. 2 a- Fig. 2 c is for trained true scratch image;
Fig. 3 a- Fig. 3 c is the tag image of binaryzation;
Fig. 4 a- Fig. 4 c is scratch positioning schematic diagram;
Fig. 5 is the scratch image after multiple image successively overturns, cuts and scales;
Fig. 6 (a) is the scratch image generated by the method for the present invention;
Fig. 6 (b) is true scratch image.
Specific embodiment
In order to more clearly illustrate technical solution of the present invention, below in conjunction with the attached drawing in the present invention to skill of the invention
Art scheme is described further.Obviously, content described in this specification embodiment is only the way of realization to inventive concept
Enumerate, protection scope of the present invention should not be construed as being limited to the specific forms stated in the embodiments, protection of the invention
Range also and in those skilled in the art conceive according to the present invention it is conceivable that equivalent technologies mean.
Embodiment 1
As shown in Figure 1, a kind of scratch image data amplification method based on production confrontation network is present embodiments provided,
The following steps are included:
Step S1 constructs scratch image data set;
Step S2, scratch image preprocessing;
Step S3, building production fight network;
Step S4, dual training;
Step S5 exports scratch image amplification data collection.
Specifically, in step sl, construct scratch image data set the following steps are included:
Step S1.1 acquires product surface image by CCD camera, chooses the wherein image with scratch and is used as training
Image, as shown in Fig. 2 a- Fig. 2 c, the image of remaining no marking is used as the background image of subsequent composograph;
Step S1.2 carries out handmarking to scoring position, and tag image is carried out binary conversion treatment, as Fig. 3 a- schemes
Shown in 3c.
Further, in step s 2, scratch image preprocessing the following steps are included:
Scratch image is converted to grayscale image by step S2.1;
Step S2.2, scratch positioning, as shown in Fig. 4 a- Fig. 4 c;
Step S2.3 cuts scored area image;
Step S2.4 zooms in and out the scratch image after cutting, and the scratch image that size is 64 × 64 is obtained after scaling,
As shown in Fig. 5;
Step S2.5 carries out Random Level overturning to the scratch image after scaling.
Further, in step S2.2, scratch positioning the following steps are included:
S2.2.1: contours extract is carried out to binaryzation tag image, gained profile is approximate region where scratch;
S2.2.2: Threshold segmentation is carried out in contoured interior, edge detection is carried out to scratch.
Further, in step S2.3, scored area image is cut the following steps are included:
S2.3.1: traversal scratching edge all pixels point finds minimum abscissa, ordinate and the maximum of these pixels
Abscissa, ordinate, to obtain horizontal and including scratch capable of being completely included a minimum rectangle;
S2.3.2: using rectangular centre as image cropping center, the image containing scratch having a size of 120 × 120 is cut out;
S2.3.3: the image after centering idea is cut carries out 8 random croppings, and random cropping is having a size of 100 × 100.
Further, in step s3, production confrontation network includes a generation network G and a differentiation network D, net
Network framework details difference is as shown in Table 1 and Table 2, wherein Conv indicates that convolutional layer, Res indicate that residual block, BN indicate batch regularization
Layer, ReLU indicate that the linear unit activating function of amendment, Tanh indicate that tangent activation primitive, fc indicate full articulamentum, and convolution kernel is such as
[3 × 3] indicate that convolution kernel size is 3 × 3, output 512 × 4 × 4 indicates that output channel number is 512, and output characteristic pattern size is 4
×4。
The framework details of the generation network G of table 1.
The framework details of the differentiation of table 2. network D
Further, in step s 4, dual training refers to utilizing stochastic gradient descent method alternately optimization aim letter
Number (1) and objective function (2):
Wherein, LDIt is the loss function for differentiating network D,It isFrom the expectation of the probability of authentic specimen
Value,X~Pr, x is true scratch image,It is the scratch image generated by generation network G, PrAnd PgIt respectively indicates true
Real scratch image data set and the probability distribution for generating scratch image data set,It isFrom the probability of authentic specimen,
D (x) is the probability that x derives from authentic specimen, Ex[D (x)] is desired value of the x from the probability of authentic specimen, and λ isIn LDIn weight coefficient,It is the interpolation of true scratch image and generation image,ε is to obey equally distributed random number, and ε~U [0,1], U are to be uniformly distributed,It is the desired value of gradient penalty term,It isGradient,It isFrom authentic specimen
Probability, LGIt is the loss function for generating network G.
Further, in step s 5, scratch image amplification data collection is generated by the trained generation network G of step S4,
As shown in Fig. 6 (a), only needed on a small quantity really by can be seen that the present invention with the true scratch image comparison as shown in Fig. 6 (b)
Scratch image can generate a large amount of and true scratch image equal quality artificial synthesized scratch image, so as to substantially reduce
Demand based on the application system of deep learning to true scratch image data, effectively save collection and mark sample at
This.
Claims (8)
1. a kind of scratch image data amplification method based on production confrontation network, which comprises the following steps:
Step S1 constructs scratch image data set;
Step S2, scratch image preprocessing;
Step S3, building production fight network;
Step S4, dual training;
Step S5 exports scratch image amplification data collection.
2. the method as described in claim 1, which is characterized in that in the step S1, building scratch image data set include with
Lower step:
Step S1.1 acquires product surface image by CCD camera, chooses the wherein image with scratch and is used as training figure
Picture, the image of remaining no marking are used as the background image of subsequent composograph;
Step S1.2 carries out handmarking to scratch approximate location, and tag image is carried out binary conversion treatment.
3. the method as described in claim 1, which is characterized in that in the step S2, scratch image preprocessing includes following step
It is rapid:
Scratch image is converted to grayscale image by step S2.1;
Step S2.2, scratch positioning;
Step S2.3, scratch image cropping;
Step S2.4 zooms in and out the scratch image after cutting;
Step S2.5 carries out Random Level overturning to the scratch image after scaling.
4. method as claimed in claim 3, which is characterized in that in the step S2.2, scratch positioning the following steps are included:
S2.2.1: contours extract is carried out to binaryzation tag image, gained profile is approximate region where scratch;
S2.2.2: Threshold segmentation is carried out in contoured interior, edge detection is carried out to scratch.
5. method as claimed in claim 3, which is characterized in that in the step S2.3, scratch image cropping includes following step
It is rapid:
S2.3.1: traversal scratching edge all pixels point finds the minimum abscissa, ordinate and maximum horizontal seat of these pixels
Mark, ordinate, thus obtain one it is horizontal, scratch can be completely included including minimum rectangle;
S2.3.2: using rectangular centre as image cropping center, the image containing scratch is cut out;
S2.3.3: the image after centering idea is cut carries out multiple random cropping.
6. the method as described in claim 1, which is characterized in that in the step S3, it includes a life that production, which fights network,
At network G and a differentiation network D.
7. the method as described in claim 1, which is characterized in that in the step S4, dual training refers to utilizing boarding steps
Spend descent method alternately optimization object function (1) and objective function (2):
Wherein, LDIt is the loss function for differentiating network D,It isFrom the desired value of the probability of authentic specimen,X~Pr, x is true scratch image,It is the scratch image generated by generation network G, PrAnd PgIt respectively indicates true
Scratch image data set and the probability distribution for generating scratch image data set,It isFrom the probability of authentic specimen, D
It (x) is probability of the x from authentic specimen, Ex[D (x)] is desired value of the x from the probability of authentic specimen, and λ isIn LDIn weight coefficient,It is the interpolation of true scratch image and generation image,ε is to obey equally distributed random number, and ε~U [0,1], U are to be uniformly distributed,It is the desired value of gradient penalty term,It isGradient,It isFrom authentic specimen
Probability, LGIt is the loss function for generating network G.
8. the method as described in claim 1, which is characterized in that in the step S5, scratch image amplification data collection is by step
The trained generation network G of S4 generates.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910303496.2A CN110189351A (en) | 2019-04-16 | 2019-04-16 | A kind of scratch image data amplification method based on production confrontation network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910303496.2A CN110189351A (en) | 2019-04-16 | 2019-04-16 | A kind of scratch image data amplification method based on production confrontation network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110189351A true CN110189351A (en) | 2019-08-30 |
Family
ID=67714637
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910303496.2A Pending CN110189351A (en) | 2019-04-16 | 2019-04-16 | A kind of scratch image data amplification method based on production confrontation network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110189351A (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110188824A (en) * | 2019-05-31 | 2019-08-30 | 重庆大学 | A kind of small sample plant disease recognition methods and system |
CN111008692A (en) * | 2019-11-08 | 2020-04-14 | 国网天津市电力公司 | Method and device for generating multi-energy metering characteristic data based on improved generation countermeasure network |
CN112116601A (en) * | 2020-08-18 | 2020-12-22 | 河南大学 | Compressive sensing sampling reconstruction method and system based on linear sampling network and generation countermeasure residual error network |
CN112462001A (en) * | 2020-11-17 | 2021-03-09 | 吉林大学 | Gas sensor array model calibration method for data amplification based on condition generation countermeasure network |
CN113159047A (en) * | 2021-04-16 | 2021-07-23 | 南通大学 | Power transformation equipment infrared image temperature value identification method based on CGAN image amplification |
CN114266768A (en) * | 2022-03-01 | 2022-04-01 | 聚时科技(江苏)有限公司 | Method for generating surface scratch defect image |
CN115270614A (en) * | 2022-07-18 | 2022-11-01 | 郑州轻工业大学 | Visual generation method for multi-physical-field digital twin organisms of muddy water circulation system |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107016406A (en) * | 2017-02-24 | 2017-08-04 | 中国科学院合肥物质科学研究院 | The pest and disease damage image generating method of network is resisted based on production |
CN108711138A (en) * | 2018-06-06 | 2018-10-26 | 北京印刷学院 | A kind of gray scale picture colorization method based on generation confrontation network |
CN108960409A (en) * | 2018-06-13 | 2018-12-07 | 南昌黑鲨科技有限公司 | Labeled data generation method, equipment and computer readable storage medium |
CN109101976A (en) * | 2018-07-10 | 2018-12-28 | 温州大学 | A kind of detection method of arc extinguishing grid pieces surface defect |
US20190057520A1 (en) * | 2017-08-18 | 2019-02-21 | Synapse Technology Corporation | Generating Synthetic Image Data |
-
2019
- 2019-04-16 CN CN201910303496.2A patent/CN110189351A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107016406A (en) * | 2017-02-24 | 2017-08-04 | 中国科学院合肥物质科学研究院 | The pest and disease damage image generating method of network is resisted based on production |
US20190057520A1 (en) * | 2017-08-18 | 2019-02-21 | Synapse Technology Corporation | Generating Synthetic Image Data |
CN108711138A (en) * | 2018-06-06 | 2018-10-26 | 北京印刷学院 | A kind of gray scale picture colorization method based on generation confrontation network |
CN108960409A (en) * | 2018-06-13 | 2018-12-07 | 南昌黑鲨科技有限公司 | Labeled data generation method, equipment and computer readable storage medium |
CN109101976A (en) * | 2018-07-10 | 2018-12-28 | 温州大学 | A kind of detection method of arc extinguishing grid pieces surface defect |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110188824A (en) * | 2019-05-31 | 2019-08-30 | 重庆大学 | A kind of small sample plant disease recognition methods and system |
CN110188824B (en) * | 2019-05-31 | 2021-05-14 | 重庆大学 | Small sample plant disease identification method and system |
CN111008692A (en) * | 2019-11-08 | 2020-04-14 | 国网天津市电力公司 | Method and device for generating multi-energy metering characteristic data based on improved generation countermeasure network |
CN112116601A (en) * | 2020-08-18 | 2020-12-22 | 河南大学 | Compressive sensing sampling reconstruction method and system based on linear sampling network and generation countermeasure residual error network |
CN112462001A (en) * | 2020-11-17 | 2021-03-09 | 吉林大学 | Gas sensor array model calibration method for data amplification based on condition generation countermeasure network |
CN112462001B (en) * | 2020-11-17 | 2021-07-23 | 吉林大学 | Gas sensor array model calibration method for data amplification based on condition generation countermeasure network |
CN113159047A (en) * | 2021-04-16 | 2021-07-23 | 南通大学 | Power transformation equipment infrared image temperature value identification method based on CGAN image amplification |
CN113159047B (en) * | 2021-04-16 | 2024-02-06 | 南通大学 | Substation equipment infrared image temperature value identification method based on CGAN image amplification |
CN114266768A (en) * | 2022-03-01 | 2022-04-01 | 聚时科技(江苏)有限公司 | Method for generating surface scratch defect image |
CN115270614A (en) * | 2022-07-18 | 2022-11-01 | 郑州轻工业大学 | Visual generation method for multi-physical-field digital twin organisms of muddy water circulation system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110189351A (en) | A kind of scratch image data amplification method based on production confrontation network | |
CN109360206B (en) | Method for dividing rice ears in field based on deep learning | |
CN105894045B (en) | A kind of model recognizing method of the depth network model based on spatial pyramid pond | |
CN110059758A (en) | A kind of remote sensing image culture pond detection method based on semantic segmentation | |
CN108346159A (en) | A kind of visual target tracking method based on tracking-study-detection | |
CN107273502A (en) | A kind of image geographical marking method learnt based on spatial cognition | |
CN111982910B (en) | Weak supervision machine vision detection method and system based on artificial defect simulation | |
CN111753828A (en) | Natural scene horizontal character detection method based on deep convolutional neural network | |
CN102831604A (en) | Two-dimensional Renyi entropic threshold segmentation method for grayscale images | |
CN112784736A (en) | Multi-mode feature fusion character interaction behavior recognition method | |
CN111127417B (en) | Printing defect detection method based on SIFT feature matching and SSD algorithm improvement | |
CN110322403A (en) | A kind of more supervision Image Super-resolution Reconstruction methods based on generation confrontation network | |
CN110414509A (en) | Stop Ship Detection in harbour based on the segmentation of extra large land and feature pyramid network | |
CN109685030A (en) | A kind of mug rim of a cup defects detection classification method based on convolutional neural networks | |
CN116542932A (en) | Injection molding surface defect detection method based on improved YOLOv5s | |
CN111754538A (en) | Threshold segmentation method for USB surface defect detection | |
CN113657528A (en) | Image feature point extraction method and device, computer terminal and storage medium | |
CN111310754A (en) | Method for segmenting license plate characters | |
Zhong et al. | Identification and depth localization of clustered pod pepper based on improved Faster R-CNN | |
CN108133182B (en) | New energy power generation prediction method and device based on cloud imaging | |
CN110175548A (en) | Remote sensing images building extracting method based on attention mechanism and channel information | |
CN111161227B (en) | Target positioning method and system based on deep neural network | |
CN117079125A (en) | Kiwi fruit pollination flower identification method based on improved YOLOv5 | |
CN114863258B (en) | Method for detecting small target based on visual angle conversion in sea-sky-line scene | |
CN112767387B (en) | Automatic wood image recognition method based on block gradient weighting |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190830 |
|
RJ01 | Rejection of invention patent application after publication |