CN112085734B - GAN-based image restoration defect detection method - Google Patents
GAN-based image restoration defect detection method Download PDFInfo
- Publication number
- CN112085734B CN112085734B CN202011024273.1A CN202011024273A CN112085734B CN 112085734 B CN112085734 B CN 112085734B CN 202011024273 A CN202011024273 A CN 202011024273A CN 112085734 B CN112085734 B CN 112085734B
- Authority
- CN
- China
- Prior art keywords
- image
- image data
- loss
- defect
- gan
- 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.)
- Active
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 79
- 230000007547 defect Effects 0.000 title claims abstract description 74
- 238000003708 edge detection Methods 0.000 claims abstract description 31
- 238000000034 method Methods 0.000 claims abstract description 24
- 238000012360 testing method Methods 0.000 claims abstract description 15
- 238000007781 pre-processing Methods 0.000 claims abstract description 10
- 238000012549 training Methods 0.000 claims description 39
- 238000001914 filtration Methods 0.000 claims description 15
- 238000004364 calculation method Methods 0.000 claims description 8
- 230000004927 fusion Effects 0.000 claims description 7
- 238000012545 processing Methods 0.000 claims description 6
- 230000003321 amplification Effects 0.000 claims description 5
- 238000003199 nucleic acid amplification method Methods 0.000 claims description 5
- 230000003190 augmentative effect Effects 0.000 claims description 4
- 230000011218 segmentation Effects 0.000 claims description 3
- 238000003709 image segmentation Methods 0.000 claims 1
- 230000002950 deficient Effects 0.000 abstract description 16
- 238000000605 extraction Methods 0.000 abstract description 7
- 238000012805 post-processing Methods 0.000 abstract description 4
- 238000013461 design Methods 0.000 abstract description 3
- 238000011084 recovery Methods 0.000 abstract description 3
- 238000013135 deep learning Methods 0.000 description 4
- 238000011156 evaluation Methods 0.000 description 3
- 238000004519 manufacturing process Methods 0.000 description 3
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000011895 specific detection Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/70—Denoising; Smoothing
-
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/40—Analysis of texture
- G06T7/41—Analysis of texture based on statistical description of texture
- G06T7/44—Analysis of texture based on statistical description of texture using image operators, e.g. filters, edge density metrics or local histograms
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20024—Filtering details
- G06T2207/20032—Median filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20092—Interactive image processing based on input by user
- G06T2207/20104—Interactive definition of region of interest [ROI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20212—Image combination
- G06T2207/20221—Image fusion; Image merging
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30168—Image quality inspection
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Quality & Reliability (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a GAN-based image restoration defect detection method. The method designs an image restoration network based on GAN, realizes a learning task based on a small number of non-defective samples by preprocessing non-defective sample image data, and solves the problem of unbalance of positive and negative samples in an actual industrial application scene; by combining the network with the edge detection Loss, the element recovery capability of the network is enhanced; meanwhile, a post-processing method of region of interest (ROI) extraction and edge removal is adopted, so that the defect detection precision of the network is greatly improved, and the defect detection task of the industrial element product is well realized. The test results on 150 test image data of the industrial component product (139 pieces of which were defective and 11 pieces of which were not defective) confirmed the validity of the present invention for detecting defects on the industrial component product.
Description
Technical Field
The invention belongs to the technical field of image recognition, and particularly relates to a defect detection method for image restoration based on GAN.
Background
In the production process of industrial components, a defect detection link is very critical for ensuring the delivery quality of the components, and industrial component products containing defects need to be accurately detected so as to ensure the delivery yield of the industrial components.
The conventional defect detection method comprises the following steps: image preprocessing, combined with machine learning, etc., while effective in certain applications, still suffer from a number of deficiencies, such as: poor robustness, huge computation, and inability to accurately detect the size and shape of the defect.
In recent years, with the rapid development of the field of deep learning, the method based on deep learning is gradually and widely applied to industrial scenes, and compared with the conventional defect detection method, the defect detection method based on deep learning can obtain a better detection effect, further save a large amount of manpower and material resources, and improve the production efficiency of industrial elements.
However, in an actual industrial application scenario, the number of non-defective samples is far greater than that of defective samples, and the training difficulty of the defect detection network is increased due to the unbalanced problem of positive and negative samples, so that the defect detection network is difficult to be trained sufficiently, and the final defect detection effect is affected.
Disclosure of Invention
Aiming at the problems of the existing defect detection method based on deep learning, the invention provides a defect detection method based on GAN image restoration.
The invention is realized by adopting the following technical scheme:
a defect detection method for GAN-based image repair comprises the following steps:
1) preprocessing the original 1000 defect-free sample training image data, namely sequentially performing image data amplification, image data gray value adjustment and image data noise addition on the original 1000 defect-free sample training image data to generate 4000 training image data for GAN network training;
2) adding edge detection Loss, and performing weighted summation on the edge detection Loss, the generator Loss of the GAN network and the discriminator Loss to serve as the total Loss of the generator of the GAN network for training of the generator;
3) extracting an ROI (region of interest) aiming at an image to be detected, generating an ROI image, carrying out image fusion on the ROI image and a residual image, and then carrying out threshold processing to generate a primary defect detection image;
4) and performing edge detection on the reconstructed image of the GAN network, and removing unnecessary edge regions in the preliminary defect detection image by using the detection result to obtain a final defect detection result.
A further development of the invention is that, in step 1), by analyzing the main defect types of the component, it comprises: the size and color of the defect, the specific form of the defect and the position of the defect in the element are combined with the color characteristics of the element, and the original 1000 pieces of defect-free sample training image data are preprocessed.
The further improvement of the invention is that in the step 1), the specific implementation method is as follows:
firstly, image data is augmented, the image data is respectively subjected to horizontal overturning, vertical overturning and diagonal overturning, filtering is carried out on the basis by adopting 50% probability, so that the image is smoother, and three types of mean filtering, median filtering and Gaussian filtering are randomly selected in a filtering mode to generate 4000 data augmented images;
secondly, adjusting the gray value of the image data, changing the gray value range of 4000 image data after the data is amplified, randomly and integrally adjusting the gray value of the 4000 image data to be 0-80 on the basis of the original gray value, keeping the original gray value of the black part with the gray value being less than 50, and keeping the training image data to be closer to the test image data without changing the black part;
finally, image data is subjected to noise addition, whether noise addition is carried out on the image data is determined according to the probability of 90%, four types of curve noise, linear noise, circular noise and elliptic noise are randomly selected according to the noise type, the gray value of the noise is randomly selected from any gray value within the range of 0-255, and the number and the size of the noise are randomly selected within a preset value range;
and sending each 4000 pieces of image data before and after noise addition as final required training data into the GAN network for training.
The invention is further improved in that, in step 2), the Loss of edge detection is added to the total Loss of the generator of the GAN network, i.e. Sobel edge detection is performed on the input image and the reconstructed image of the GAN network respectively, L2_ Loss between the edge detection results is calculated as the Loss of edge detection, and is weighted and summed with the Loss of the generator of the GAN network and the Loss of the discriminator to serve as the total Loss of the generator of the GAN network for training of the generator, and the calculation formula is as follows:
g_total_loss＝2*g_loss+0.05*d_loss+2*edge_loss。
the further improvement of the invention is that, in the step 3), the ROI is extracted from the image to be detected, the ROI is white, the region not to be interested is black, and the image fusion is carried out on the ROI image and the residual image between the reconstructed image generated by the GAN network and the input image according to the proportion of 0.5:0.5, wherein the calculation formula is as follows:
result＝residual_image*0.5+roi_image*0.5
and finally, performing corrosion-first and expansion-later processing on the segmentation result by using a 3 x 3-sized cross-shaped inner core, eliminating the existence of isolated points, and obtaining a primary defect detection image.
The further improvement of the invention is that, in the step 4), Laplace edge detection is performed on the reconstructed image of the GAN network, and the detection result is utilized to remove the unnecessary edge area from the preliminary defect detection result, so as to obtain the final defect detection result.
The invention has at least the following beneficial effects:
the invention provides a defect detection method for image restoration based on GAN. The method designs an image restoration network based on GAN, realizes a learning task based on a small number of non-defective samples by preprocessing non-defective sample image data, and solves the problem of unbalance of positive and negative samples in an actual industrial application scene; by combining the network with the edge detection Loss, the element recovery capability of the network is enhanced; meanwhile, a post-processing method of region of interest (ROI) extraction and edge removal is adopted, so that the defect detection precision of the network is greatly improved, and the defect detection task of the industrial element product is well realized.
Further, the preprocessing module in the invention mainly preprocesses the original 1000 training image data of the defect-free sample, namely, the image data is subjected to image data amplification, image data gray value adjustment and image data noise addition in sequence to generate 4000 training image data for GAN network training, thereby solving the problem of imbalance of positive and negative samples.
Furthermore, the edge detection Loss module in the invention adds the edge detection Loss, and performs weighted summation with the generator Loss and the discriminator Loss of the GAN network to be used as the total Loss of the generator of the GAN network for training the generator to ensure that the edge part of the reconstructed image is closer to the input image, thereby reducing the error of the edge area and improving the detection precision of the defect detection network.
Furthermore, the region of interest (ROI) extraction module extracts the ROI from the image to be detected, generates an ROI image, performs image fusion on the ROI image and the residual image, performs threshold processing on the ROI image, generates a preliminary defect detection image, and improves the detection precision of a defect detection network.
Further, the edge removal module in the invention performs edge detection on the reconstructed image of the GAN network, and removes the unnecessary edge region in the preliminary defect detection image by using the detection result to obtain the final defect detection result, thereby improving the detection accuracy of the defect detection network.
Further, in order to verify the effectiveness of the method in the defect detection task, performance evaluation is performed on the basis of 150 pieces of industrial element product test image data (139 pieces of defective and 11 pieces of non-defective), and in the detection results, the undetected rate escape is 0.0, the undetected rate overkill is 0.0071428, the detection precision mIOU at the pixel level is 0.5936287, the final detection score is 0.9165828, the detection speed is 7fps, and the calculation mode of the detection score is as shown in formula (1):
score＝0.5*(1-escape)+0.3*(1-overkill)+0.2*(mIOU) (1)
in summary, the present invention provides a defect detection method for GAN-based image repair. Aiming at the problem of unbalance of positive and negative samples in an actual industrial application scene, the method adopts a method of image preprocessing (image data amplification, image data gray value adjustment and image data noise addition) based on a GAN (generic area network) repair network, and realizes a learning task based on a small number of non-defective samples; by combining the GAN network with the edge detection Loss, the element recovery capability of the network is enhanced; meanwhile, a post-processing method of region of interest (ROI) extraction and edge removal is adopted, so that the defect detection precision of the network is greatly improved. In a defect detection task of 150 pieces of industrial element product test image data (139 pieces of defective and 11 pieces of non-defective), the method reduces the undetected rate which has the greatest influence on the factory quality of industrial products to 0.0, reduces the overdetected rate which has the influence only next to the undetected rate as much as possible on the basis, reduces the value of the undetected rate as little as possible, improves the value of the pixel-level detection precision as much as possible, finally obtains a higher detection score, and has good application value.
Drawings
Fig. 1 is a flowchart of a training phase of a GAN-based image repair defect detection method according to the present invention.
FIG. 2 is a flowchart of a testing phase of a GAN-based image repair defect detection method according to the present invention.
Fig. 3 is a general flowchart of a GAN-based image repair defect detection method according to the present invention.
Detailed Description
The invention is explained in detail below with reference to the drawings:
the flow chart of the training phase of the present invention is shown in fig. 1.
In the training stage, inputting a defect-free image, preprocessing the defect-free image, namely amplifying image data, adjusting the gray value of the image data and adding noise to the image data to generate a noise-added image, inputting the noise-added image into a generator of the GAN network, repairing the image, outputting a reconstructed image, and then inputting the output reconstructed image into a discriminator of the GAN network to judge whether the image is true or false.
The loss of the generator is obtained by carrying out weighted summation on the L2_ loss between the input defect-free image and the output reconstructed image, the L2_ loss between the edge detection results of the input defect-free image and the output reconstructed image and the error result of true and false judgment of a discriminator of the output reconstructed image, and the loss of the generator is used for training the generator; and calculating error results of true and false judgment of the discriminators of the input non-defective image and the output reconstructed image respectively, summing the error results and the output reconstructed image to obtain the loss of the discriminators, and using the loss to train the discriminators to finally realize the training of the whole network.
The test phase flow diagram of the present invention is shown in fig. 2.
In the testing stage, an image to be detected is input, the image to be detected is input into a generator of a trained GAN network, image restoration is carried out, a reconstructed image is output, the residual error between the input image to be detected and the output reconstructed image is calculated, meanwhile, a region of interest (ROI) is extracted from the input image to be detected, an ROI image is generated, image fusion is carried out on the ROI image and the residual error image, then threshold processing is carried out, a preliminary defect detection image is generated, finally, edge detection is carried out on the output reconstructed image, and the detection result is utilized to remove the edge region which is not needed in the preliminary defect detection image, so that the final defect detection result is obtained.
The general flow chart of the present invention is shown in fig. 3.
The invention provides a defect detection method for image restoration based on GAN, which mainly comprises the following modules:
1) a pre-processing module, which corresponds to the pre-processing module in fig. 1.
The module mainly analyzes the main defect types of the elements and mainly comprises the following components: the size and color of the defect, the specific form of the defect and the position of the defect in the element are combined with the color characteristics of the element, and the original 1000 pieces of defect-free sample training image data are preprocessed.
Firstly, image data is amplified, the image data is respectively subjected to horizontal overturning, vertical overturning and diagonal overturning, filtering is carried out on the basis by adopting a 50% probability, the image is smoother, and three types of mean filtering, median filtering and Gaussian filtering are randomly selected in a filtering mode to generate 4000 data amplification images.
And secondly, adjusting the gray value of the image data, changing the gray value range of 4000 image data after the data is amplified, and randomly and integrally adjusting the gray value of the 4000 image data by 0-80 (the black part with the gray value less than 50 still keeps the original gray value and is not changed) on the basis of the original gray value, so that the training image data is closer to the test image data.
And finally, adding noise to the image data, determining whether to add the noise to the image data according to the probability of 90%, wherein the noise type randomly selects four types of curve noise, linear noise, circular noise and elliptic noise, the gray value of the noise randomly selects any gray value within the range of 0-255, and the number and the size of the noise are randomly selected within a preset value range.
And sending each 4000 pieces of image data before and after noise addition as final required training data into the GAN network for training.
2) And an edge detection Loss module, which corresponds to the edge detection Loss module in fig. 1.
The module adds the Loss of edge detection in the total Loss of the generator of the GAN network, namely, Sobel edge detection is respectively carried out on the input image and the reconstructed image of the GAN network, and the L2_ Loss between the edge detection results is calculated to be used as the Loss of edge detection, and the Loss of the generator of the GAN network and the Loss of a discriminator are subjected to weighted summation to be used as the total Loss of the generator of the GAN network for training of the generator, wherein the calculation formula is shown as formula (2):
g_total_loss＝2*g_loss+0.05*d_loss+2*edge_loss (2)
3) a region of interest (ROI) extraction module, which corresponds to the region of interest (ROI) extraction module in fig. 2.
The module extracts a region of interest (ROI, an element internal region) of an image to be detected, makes the region of interest white and a region of no interest black, and performs image fusion on a residual image between a reconstructed image generated by an ROI image and a GAN network and an input image according to the proportion of 0.5:0.5, wherein the calculation formula is shown as a formula (3):
result＝residual_image*0.5+roi_image*0.5 (3)
and finally, performing corrosion-first and expansion-later processing on the segmentation result by using a 3 x 3-sized cross-shaped inner core, eliminating the existence of isolated points, and obtaining a primary defect detection image.
4) An edge removal module, which corresponds to the edge removal module in fig. 2.
The module carries out Laplace edge detection on the reconstructed image of the GAN network, and removes the unneeded edge area from the preliminary defect detection result by using the detection result to obtain a final defect detection result.
To test the effectiveness of the present invention for defect detection, performance evaluations were developed on the basis of 150 production test image data of industrial components (139 defective and 11 non-defective), and the specific detection performance of the present invention is shown in table 1. According to the test result, the defect detection method based on GAN image repair provided by the invention has the advantages that the undetected rate escape is 0.0, the undetected rate overkill is 0.0071428, the pixel-level detection precision mIOU is 0.5936287, the final detection score is 0.9165828, and the detection speed is 7 fps.
Compared with the traditional defect detection method, the method is based on the expansion training of the image data of the defect-free sample, so that the model has a higher robust type; the defect detection task in the method is mainly completed by a generator which comprises 15 convolution layers, 3 maximum value down-sampling layers and 3 anti-convolution layers, parameters can be directly updated through learning data, a complex algorithm process of manual design is avoided, the whole method is simple and easy to implement, the training and testing tasks of the network can be completed in a short time, and the calculated amount is small; on the basis, a post-processing method of region of interest (ROI) extraction and edge removal is adopted, so that the defect detection precision is greatly improved, and the defect detection task of the industrial element product is well completed.
Because the method is based on the image data expansion training of the defect-free sample, and any image data with a label is not used in the training process, the method has good generalization capability, can be suitable for various industrial scenes, and has good application value.
TABLE 1 Performance evaluation results of the present invention on 150 test image data
Claims (4)
1. A defect detection method for image restoration based on GAN is characterized by comprising the following steps:
1) preprocessing the original 1000 defect-free sample training image data, namely sequentially performing image data amplification, image data gray value adjustment and image data noise addition on the original 1000 defect-free sample training image data to generate 4000 training image data for GAN network training;
2) adding edge detection Loss, and performing weighted summation on the edge detection Loss, the generator Loss of the GAN network and the discriminator Loss to serve as the total Loss of the generator of the GAN network for training of the generator; the method specifically comprises the following steps: increasing the Loss of edge detection in the total Loss of the generator of the GAN network, namely performing Sobel edge detection on the input image and the reconstructed image of the GAN network respectively, calculating L2_ Loss between the edge detection results as the Loss of edge detection, performing weighted summation with the generator Loss of the GAN network and the discriminator Loss as the total Loss of the generator of the GAN network for training the generator, wherein the calculation formula is as follows:
g_total_loss＝2*g_loss+0.05*d_loss+2*edge_loss
3) extracting an ROI (region of interest) aiming at an image to be detected, generating an ROI image, carrying out image fusion on the ROI image and a residual image, and then carrying out threshold processing to generate a primary defect detection image; the method specifically comprises the following steps: extracting an interested region ROI of an image to be detected, enabling the interested region to be white and enabling an uninteresting region to be black, and carrying out image fusion on the ROI image and a residual image between a reconstructed image generated by a GAN network and an input image according to the proportion of 0.5:0.5, wherein the calculation formula is as follows:
result＝residual_image*0.5+roi_image*0.5
performing gray level histogram calculation on the fused image result to determine the optimal threshold value of the image, performing image segmentation by using the threshold value, and finally performing corrosion-first and expansion-later treatment on the segmentation result by using a 3 x 3-sized cross-shaped inner core to eliminate the existence of isolated points to obtain a primary defect detection image;
4) and performing edge detection on the reconstructed image of the GAN network, and removing unnecessary edge regions in the preliminary defect detection image by using the detection result to obtain a final defect detection result.
2. The method as claimed in claim 1, wherein the step 1) of analyzing the main defect types of the components comprises: the size and color of the defect, the specific form of the defect and the position of the defect in the element are combined with the color characteristics of the element, and the original 1000 pieces of defect-free sample training image data are preprocessed.
3. The method for detecting defects of GAN-based image restoration according to claim 2, wherein the step 1) is implemented as follows:
firstly, image data is augmented, the image data is respectively subjected to horizontal overturning, vertical overturning and diagonal overturning, filtering is carried out on the basis by adopting 50% probability, so that the image is smoother, and three types of mean filtering, median filtering and Gaussian filtering are randomly selected in a filtering mode to generate 4000 data augmented images;
secondly, adjusting the gray value of the image data, changing the gray value range of 4000 image data after the data is amplified, randomly and integrally adjusting the gray value of the 4000 image data to be 0-80 on the basis of the original gray value, keeping the original gray value of the black part with the gray value being less than 50, and keeping the training image data to be closer to the test image data without changing the black part;
finally, image data is subjected to noise addition, whether noise addition is carried out on the image data is determined according to the probability of 90%, four types of curve noise, linear noise, circular noise and elliptic noise are randomly selected according to the noise type, the gray value of the noise is randomly selected from any gray value within the range of 0-255, and the number and the size of the noise are randomly selected within a preset value range;
and sending each 4000 pieces of image data before and after noise addition as final required training data into the GAN network for training.
4. The method as claimed in claim 1, wherein in step 4), Laplace edge detection is performed on the reconstructed image of the GAN network, and the detection result is used to remove the unwanted edge region from the preliminary defect detection result, so as to obtain a final defect detection result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011024273.1A CN112085734B (en) | 2020-09-25 | 2020-09-25 | GAN-based image restoration defect detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011024273.1A CN112085734B (en) | 2020-09-25 | 2020-09-25 | GAN-based image restoration defect detection method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112085734A CN112085734A (en) | 2020-12-15 |
| CN112085734B true CN112085734B (en) | 2022-02-01 |
Family
ID=73739035
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011024273.1A Active CN112085734B (en) | 2020-09-25 | 2020-09-25 | GAN-based image restoration defect detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112085734B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113763320A (en) * | 2021-07-29 | 2021-12-07 | 国网江苏省电力有限公司电力科学研究院 | Cable accessory lead sealing construction defect detection method and system |
| CN113516656B (en) * | 2021-09-14 | 2021-12-14 | 浙江双元科技股份有限公司 | Defect image data processing simulation method based on ACGAN and Cameralink cameras |
| WO2023229590A1 (en) * | 2022-05-25 | 2023-11-30 | Innopeak Technology, Inc. | Deep learning based video super-resolution |
| CN116452906B (en) * | 2023-03-03 | 2024-01-30 | 哈尔滨市科佳通用机电股份有限公司 | Railway wagon fault picture generation method based on text description |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109345469A (en) * | 2018-09-07 | 2019-02-15 | 苏州大学 | It is a kind of that speckle denoising method in the OCT image of confrontation network is generated based on condition |
| CN110348445A (en) * | 2019-06-06 | 2019-10-18 | 华中科技大学 | A kind of example dividing method merging empty convolution sum marginal information |
| CN110516201A (en) * | 2019-08-20 | 2019-11-29 | Oppo广东移动通信有限公司 | Image processing method, device, electronic equipment and storage medium |
| CN110706308A (en) * | 2019-09-07 | 2020-01-17 | 创新奇智(成都)科技有限公司 | GAN-based steel coil end face edge loss artificial sample generation method |
| CN111241958A (en) * | 2020-01-06 | 2020-06-05 | 电子科技大学 | Video image identification method based on residual error-capsule network |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10572979B2 (en) * | 2017-04-06 | 2020-02-25 | Pixar | Denoising Monte Carlo renderings using machine learning with importance sampling |
| CN109242841A (en) * | 2018-08-30 | 2019-01-18 | 广东工业大学 | A kind of transmission tower defect inspection method based on generation confrontation network |
| US11568645B2 (en) * | 2019-03-21 | 2023-01-31 | Samsung Electronics Co., Ltd. | Electronic device and controlling method thereof |
| CN110097543B (en) * | 2019-04-25 | 2023-01-13 | 东北大学 | Hot-rolled strip steel surface defect detection method based on generation type countermeasure network |
| CN110120038B (en) * | 2019-05-07 | 2021-02-26 | 重庆同枥信息技术有限公司 | Pavement crack defect detection method based on countermeasure generation network |
-
2020
- 2020-09-25 CN CN202011024273.1A patent/CN112085734B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109345469A (en) * | 2018-09-07 | 2019-02-15 | 苏州大学 | It is a kind of that speckle denoising method in the OCT image of confrontation network is generated based on condition |
| CN110348445A (en) * | 2019-06-06 | 2019-10-18 | 华中科技大学 | A kind of example dividing method merging empty convolution sum marginal information |
| CN110516201A (en) * | 2019-08-20 | 2019-11-29 | Oppo广东移动通信有限公司 | Image processing method, device, electronic equipment and storage medium |
| CN110706308A (en) * | 2019-09-07 | 2020-01-17 | 创新奇智(成都)科技有限公司 | GAN-based steel coil end face edge loss artificial sample generation method |
| CN111241958A (en) * | 2020-01-06 | 2020-06-05 | 电子科技大学 | Video image identification method based on residual error-capsule network |
Non-Patent Citations (4)
| Title |
|---|
| Deep learning fault diagnosis method based on global optimization GAN for unbalanced data;Funa Zhou;《Knowledge-Based Systems》;20200131;1-19 * |
| Defect Enhancement Generative Adversarial Network for Enlarging Data Set of Microcrack Defect;Song Lin;《IEEE Access》;20191007;148413-148423 * |
| Self-Supervised Ultrasound to MRI Fetal Brain Image Synthesis;Jianbo Jiao;《IEEE Transactions on Medical Imaging》;20200824;4413-4424 * |
| 产品质量检测的生成对抗网络方法;杜卉然;《中国优秀博硕士学位论文全文数据库(硕士) 信息科技辑》;20200215;I140-307 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112085734A (en) | 2020-12-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112085734B (en) | GAN-based image restoration defect detection method | |
| CN108961217B (en) | Surface defect detection method based on regular training | |
| CN111223088B (en) | Casting surface defect identification method based on deep convolutional neural network | |
| CN105913415B (en) | A kind of image sub-pixel edge extracting method with extensive adaptability | |
| CN113362326A (en) | Method and device for detecting welding spot defects of battery | |
| CN111223093A (en) | AOI defect detection method | |
| CN111383209A (en) | Unsupervised flaw detection method based on full convolution self-encoder network | |
| CN109191421B (en) | Visual detection method for pits on circumferential surface of cylindrical lithium battery | |
| CN110992363B (en) | Three-dimensional magnetic flux leakage detection defect reconstruction method based on generation countermeasure network | |
| CN107490582A (en) | A kind of streamline Work Piece Verification System Based | |
| CN111798409A (en) | Deep learning-based PCB defect data generation method | |
| CN114419014A (en) | Surface defect detection method based on feature reconstruction | |
| CN110782385A (en) | Image watermark removing method based on deep learning | |
| CN107316287A (en) | A kind of defect identification method in rectangle ferrite magnetic disk sheet face | |
| CN115830004A (en) | Surface defect detection method, device, computer equipment and storage medium | |
| CN114972397A (en) | Infrared image defect contour detection method for wind power blade composite material | |
| CN114782410A (en) | Insulator defect detection method and system based on lightweight model | |
| CN113971681A (en) | Edge detection method for belt conveyor in complex environment | |
| CN117252815A (en) | Industrial part defect detection method, system, equipment and storage medium based on 2D-3D multi-mode image | |
| CN112614113A (en) | Strip steel defect detection method based on deep learning | |
| CN117197146A (en) | Automatic identification method for internal defects of castings | |
| CN116433978A (en) | Automatic generation and automatic labeling method and device for high-quality flaw image | |
| CN114626445B (en) | Dam termite video identification method based on optical flow network and Gaussian background modeling | |
| CN116363097A (en) | Defect detection method and system for photovoltaic panel | |
| CN115953387A (en) | Radiographic image weld defect detection method based on deep learning |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |