CN115393333A - Industrial product surface defect detection method based on attention generation countermeasure network - Google Patents
Industrial product surface defect detection method based on attention generation countermeasure network Download PDFInfo
- Publication number
- CN115393333A CN115393333A CN202211056663.6A CN202211056663A CN115393333A CN 115393333 A CN115393333 A CN 115393333A CN 202211056663 A CN202211056663 A CN 202211056663A CN 115393333 A CN115393333 A CN 115393333A
- Authority
- CN
- China
- Prior art keywords
- attention
- network
- training
- image
- generator
- 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
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
- G06T7/0004—Industrial image inspection
- G06T7/0006—Industrial image inspection using a design-rule based approach
-
- 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
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Quality & Reliability (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to the technical field of defect detection methods, and provides an industrial product surface defect detection method based on an attention generation confrontation network. The method comprises the following steps: 1) Preprocessing an image; 2) Generating an antagonistic network model based on attention; 3) A model training stage; 4) And a defect detection stage. According to the method, the problems of low detection accuracy and insufficient feature extraction of the traditional defect detection method under the condition of lacking defect samples are solved, the method for detecting the surface defects of the industrial products based on the attention generation countermeasure network is provided, the deep convolution network and the self-attention mechanism are effectively utilized, the countermeasure network is improved, and the defect detection capability of the model is improved.
Description
Technical Field
The invention belongs to the technical field of defect detection methods, and particularly relates to an industrial product surface defect detection method based on an attention generation countermeasure network.
Background
In the technical field of defect detection methods, a deep neural network can detect whether the surface of an industrial product has defects through good training and shows excellent performance. However, in practical application of the surface defect detection of the industrial product, there is a problem that a sufficient defect sample is not available, manual labeling of the defect sample is high in labor cost and complex in operation, and how to extract the surface features of the industrial product efficiently and accurately is also a problem that needs to be solved at present.
To solve the above problem, an auto-encoder may be used to learn without supervision. However, the automatic encoder does not contribute much in the objective function, which may cause that the defects remained in the reconstructed residual may be too weak to be identified, and the detection may not achieve a perfect effect, thereby affecting the accuracy of model classification.
Therefore, there is a need in the art for a defect detection method in the absence of defect samples to address the problem of industrial product surface defect detection, and a method of adding a self-attention mechanism for feature extraction to improve the model feature extraction capability.
Disclosure of Invention
The invention aims to provide an industrial product surface defect detection method based on attention generation countermeasure network, which can effectively utilize a deep convolution network, the generation countermeasure network and a self-attention mechanism to complete the industrial product surface defect detection task in the scene lacking enough defect samples.
In order to achieve the purpose, the invention provides the following scheme:
an attention-based method for detecting surface defects of industrial products by generating a confrontation network, comprising the following steps of:
step 1: and (5) image preprocessing. Acquiring an industrial product data set, extracting a defect-free industrial product image, randomly extracting a plurality of small blocks of each type to be respectively used as a training set and a verification set, and normalizing each small block after Gaussian denoising.
Step 2: and generating a countermeasure network model building based on the attention.
And step 3: and (5) a model training stage. And (3) superimposing Gaussian noise on the industrial product pictures in the industrial product data set, inputting the industrial product pictures into the attention-based generation countermeasure network model training constructed in the step (2), and respectively training the discriminator D, the generator G and the inverter T. When the training times reach certain iteration times, obtaining a trained antagonistic network model generated based on attention;
and 4, step 4: and a defect detection stage. After training, the models D, T and G learned in step 3 are ready for industrial surface inspection. The detection process mainly comprises likelihood map generation, image reconstruction, residual map creation, image fusion and thresholding.
Further, in step 1, a specific method of image preprocessing includes:
step 1.1, an industrial product data set is obtained, and for each type of industrial product, a plurality of small blocks are randomly extracted from 2-5 defect-free images with the size of 256 × 256 to be used as training data. The size of each patch was set to 32 x 32.
And 1.2, randomly dividing the small blocks into two groups, wherein one group is used as a training set, and the other group is used as a verification set for evaluating the performance of the model.
And step 1.3, denoising each small block by using a 3 x 3 Gaussian filter, and normalizing the small blocks in a [0,1] interval.
Further, in step 2, a countermeasure network model is built based on attention generation: the attention-based generation confrontation network model is composed of a generator network, a discriminator network and an inverter network, wherein the generator network, the discriminator network and the inverter network all adopt deep convolution networks and are fused with a self-attention mechanism.
The input layer of generator G is a 64-dimensional noise vector followed by a 1024-dimensional Fully Connected (FC) layer. The FC layer is further reshape 2 x 256, where each dimension represents height (H), width (W), and depth, respectively. All layers connected are then Deconv layers of decreasing depth. Wherein the steps of all Deconv layers are 2 x 2, and each previous layer is upsampled. Inserting a self-attention module between layers, which module references SAGAN, adding a self-attention mechanism can make the model more aware of global features. The concrete formula is as follows:
y i ＝γO i +x i
wherein, beta j,i Representing the degree of influence of the model on the ith position when synthesizing the jth region, the output of the layer of interest is
Gamma is a proportional parameter, x i Is an element diagram.
The output layer size is specifically set at 32 x 32. Batch Normalization (BN) is performed on all layers except the output of generator G and activated using a rectifying linear unit (ReLU).
The inverter T and the discriminator D are both constructed reversely from G. The difference between them is the shape of the final output layer and the insertion position of the self attention module, all using the LeakyReLU activation except for the output layer.
Further, in step 3, the specific method in the model training phase includes:
step 3.1, gaussian noise is added to the image generated by the model G to "break" the similarity between the real image and the false image. The specific formula is as follows:
in the formula (I), the compound is shown in the specification,
in the image after the noise is superimposed, X is a defect-free image of the industrial product, and H represents gaussian noise following a normal distribution.
And 3.2, training a discriminator D. When the average likelihood of a batch of validation data sets reaches a threshold P at each optimization iteration, training is stopped and the parameters of discriminator D at that time are saved. The specific formula is as follows:
where n is the batch size. P is an empirically determined threshold.
And 3.3, training a generator G. Step 3.2 and training generator G to train continuously, and the network parameters are trained by using the following loss functions:
where E represents the mathematical expectation of real data x and potential spatial data z; the G network being a generator, P z (z) is the underlying spatial data distribution, typically Gaussian, resulting in a data-generating distribution Pg (x) that we expect to be very close to Pdata (x) to fit close to the true distribution. The D network is a discriminant function, needs to solve the traditional two-classification problem, and has the responsibility of effectively distinguishing real distribution and generated distribution, namely measuring the difference between Pg (x) and Pdata (x), and performing repeated iterative training.
And 3.4, training the inverter T. The inverter T defines an inverse mapping T (x): x → z with respect to the generator G, mapping the normal image block x back to the potential space. Specifically, an L2 norm loss function is used to optimize the inverter T, which is defined as:
in the optimization process, the parameters of G are fixed, and only the parameters of E are updated.
Further, in step 4, the specific method in the defect detection stage includes:
step 4.1, a set of image patches is extracted from a given query image g, wherein the size of the patches is the same as the size of the patches in the training set, which is 32 x 32. Given a patch x (i, j) located at (i, j) in g, discriminator D estimates the probability that the patch belongs to the normal class. Then, the likelihood map s of g is constructed as:
S(i，j)＝D(x(i，j))
the generator G and inverter T learned in step 4.2, step 3 are jointly used to reconstruct the input image patch x' (i, j), with the following specific formula:
x′(i，j)＝G(T(x(i，j)))
all reconstructed patches x '(i, j) are reorganized to form a full-size reconstructed image f'.
Step 4.3, carrying out normalization operation on the reconstructed image f', wherein a specific formula is as follows:
where u () is the mean of the image and σ () is the standard deviation of the image.
Step 4.4, constructing a residual error map, wherein the specific formula is as follows:
r＝|f-f r |⊙f err
wherein an is an element matrix multiplication, f err Is an error map composed of block-by-block computed reconstruction errors.
And 4.5, fusing the likelihood map s constructed in the step 4.1 and the residual map r constructed in the step 4.4 to obtain a fused map u. The concrete formula is as follows:
u(i，j)＝[1(i，j)-s(i，j)]⊙r(i，j)
where 1 (i, j) -s (i, j) represents inverting the likelihood map s so that the background intensity approaches zero. So that the fused map shows a cleaner background than the corresponding residual map.
And 4.6, carrying out threshold processing on the fusion graph. The specific formula is as follows:
wherein u is 0 Is a reference fusion map obtained from a defect-free sample, and t is an empirically determined control constant.
The invention has the advantages that:
according to the industrial product surface defect detection method based on the attention generation countermeasure network, the deep convolution is applied to the generation countermeasure network, and the self-attention mechanism is fused with the generator, the discriminator and the inverter of the model, so that the performance of the model is improved, and the detection accuracy of the industrial product surface defects is improved.
Drawings
FIG. 1 is a flow chart of a method for detecting defects on a surface of an industrial product based on an attention-generated countermeasure network according to the present invention;
FIG. 2 is a block diagram of a model training process proposed by the present invention;
fig. 3 is a network structure diagram of the self-attention mechanism of the present invention.
Detailed Description
To facilitate understanding and practice of the invention by those of ordinary skill in the art, the invention is described in further detail below by way of examples and with reference to the accompanying drawings.
As shown in fig. 1, a method for detecting surface defects of industrial products based on attention-generated countermeasure network includes the following steps:
step 1, image preprocessing. An industrial product data set was acquired and 5000 small pieces were randomly extracted from 2-5 defect free images with a size of 256 x 256 for each type of industrial product for use as training data. The size of each small block is set to be 32 x 32.
Then, 5000 small blocks were randomly divided into two groups, one group of 4500 blocks as training set and the other group of 500 blocks as validation set for model performance evaluation. And each small block is denoised by a 3 x 3 Gaussian filter and normalized in the [0,1] interval.
And 2, generating a countermeasure network model building based on attention. As shown in FIG. 2, the attention-based generation confrontation network model is composed of a generator network, a discriminator network and an inverter network, wherein the generator network, the discriminator network and the inverter network all adopt deep convolution networks and merge a self-attention mechanism.
The input layer of generator G is a 64-dimensional noise vector followed by a 1024-dimensional Fully Connected (FC) layer. The FC layer is further reshape of 2 x 256, where each dimension represents height (H), width (W), and depth, respectively. All layers connected are then Deconv layers of decreasing depth. Wherein the steps of all Deconv layers are 2 x 2, and each previous layer is up-sampled. Inserting a self-attention module between the penultimate layer and the output layer, which module references SAGAN, as shown in FIG. 3, adding a self-attention mechanism may make the model more aware of global features. The specific formula is as follows:
y i ＝γo i +x i
wherein, beta j,i Representing the degree of influence of the model on the ith position when synthesizing the jth region, the output of the layer of interest is
Gamma is a proportional parameter, x i Is an element diagram.
The output layer is specifically sized 32 x 32. Batch Normalization (BN) is performed on all layers except the output of generator G and activated using a rectifying linear unit (ReLU).
The inverter T and the discriminator D are both constructed reversely from G. The difference is the shape of the final output layer. And the self-attention module of the inverter T and the discriminator D is interposed between the penultimate layer and the output layer. Except for the output layer, they are all activated using LeakyReLU.
And 3, a model training stage. Superimposing Gaussian noise on an industrial product picture in an industrial product data set, wherein a specific formula is as follows:
in the formula (I), the compound is shown in the specification,
for the image after superimposing noise, X is a defect-free image of the industrial productAnd H denotes gaussian noise following a normal distribution.
Then will be
And (3) inputting the attention-based generation pairing network model training constructed in the step (2), and respectively training a discriminator D, a generator G and an inverter T. When the training times reach certain iteration times, obtaining a trained antagonistic network model generated based on attention;
when training the discriminator D, stopping training and saving the parameters of the discriminator D when the average likelihood of a batch of verification data sets reaches the threshold value P during each optimization iteration. The specific formula is as follows:
where n is the batch size. P is an empirically determined threshold.
The generator G is then trained. The discriminator D and generator G are trained continuously, the network parameters are trained with the following loss functions:
where E represents the mathematical expectation of real data x and potential spatial data z; the G network is a generator, P z (z) is the underlying spatial data distribution, typically a Gaussian distribution, resulting in a data-generating distribution Pg (x) that we expect to be very close to Pdata (x) to fit an approximation to the true distribution. The D network is a discriminant function, needs to solve the traditional two-classification problem, and has the responsibility of effectively distinguishing real distribution and generated distribution, namely measuring the difference between Pg (x) and Pdata (x) and carrying out repeated iterative training.
When training the inverter T, the inverter T defines an inverse mapping T (x) with respect to the generator G: x → z, mapping the normal image block x back to the potential space. Specifically, an L2 norm loss function is used to optimize the inverter T, which is defined as:
in the optimization process, the parameters of G are fixed, and only the parameters of E are updated.
And 4, a defect detection stage. After training, the models D, T and G learned in step 3 are ready for industrial surface inspection.
First, a set of image patches is extracted from a given query image g, where the patch size is the same as the size of the patches in the training set, which is 32 x 32. Given a patch x (i, j) located at (i, j) in g, discriminator D estimates the probability that the patch belongs to the normal class. Then, the likelihood map s of g is constructed as:
s(i，j)＝D(x(i，j))
the learned generator G and inverter T are then jointly used to reconstruct the input image patch x' (i, j), with the following specific formula:
x′(i，j)＝G(T(x(i，j)))
all reconstructed patches x '(i, j) are reorganized to form a full-size reconstructed image f'. Then, the reconstructed image f' is normalized, and the specific formula is as follows:
where u () is the mean of the image and σ () is the standard deviation of the image.
Then, a residual error map is constructed according to the difference value between the original image and the reconstructed image, and the specific formula is as follows:
r＝|f-f r |⊙f err
where, for example, means element matrix multiplication, ferr is an error map composed of reconstruction errors calculated block by block.
And then fusing the constructed likelihood map s and the constructed residual map r to obtain a fusion map u. The specific formula is as follows:
u(i，j)＝[1(i，j)-s(i，j)]⊙r(i，j)
where 1 (i, j) -s (i, j) represents inverting the likelihood map s so that the background intensity approaches zero. So that the fused map shows a cleaner background than the corresponding residual map.
And finally, carrying out threshold processing on the fusion graph. The concrete formula is as follows:
where uO is a reference fusion map obtained from a defect-free sample and t is an empirically determined control constant.
Claims (5)
1. An industrial product surface defect detection method based on attention generation countermeasure network is characterized by comprising the following steps:
1) Image preprocessing, namely acquiring an industrial product data set, extracting a defect-free industrial product image, randomly extracting a plurality of small blocks of each type to be respectively used as a training set and a verification set, and normalizing each small block after Gaussian denoising;
2) Generating a countermeasure network model based on attention;
3) In the model training stage, after Gaussian noise is superposed on the industrial product pictures in the industrial product data set, inputting the images into the attention-based generation countermeasure network model training constructed in the step 2, respectively training the discriminator D, the generator G and the inverter T, and when the training times reach certain iteration times, obtaining the trained attention-based generation countermeasure network model;
4) And in the defect detection stage, after training, the models D, T and G learned in the step 3 are prepared for industrial product surface detection, and the detection process mainly comprises likelihood map generation, image reconstruction, residual map creation, image fusion and thresholding.
2. The method for detecting the surface defects of the industrial products based on the attention-generating antagonistic network as claimed in claim 1, characterized in that in the step 1): the image preprocessing comprises the following specific steps:
1) Acquiring an industrial product data set, and randomly extracting a plurality of small blocks from 2-5 defect-free images with the size of 256 × 256 for each type of industrial product to be used as training data, wherein the size of each small block is set to be 32 × 32;
2) Randomly dividing the small blocks into two groups, wherein one group is used as a training set, and the other group is used as a verification set for evaluating the performance of the model;
3) Each patch was denoised with a 3 x 3 gaussian filter and normalized to the [0,1] interval.
3. The method for detecting the surface defects of the industrial products based on the attention-generating antagonistic network as claimed in claim 1, characterized in that in the step 2): the construction of the countermeasure network model based on attention generation specifically comprises the following steps:
the attention-based generation confrontation network model consists of a generator network, a discriminator network and an inverter network, wherein the generator network, the discriminator network and the inverter network all adopt deep convolution networks and are fused with a self-attention mechanism;
the input layer of the generator G is a 64-dimensional noise vector followed by a 1024-dimensional Fully Connected (FC) layer, which is further reshape 2 × 256, where each dimension represents height (H), width (W), and depth, respectively; then, all connected layers are Deconv layers with gradually reduced depth, wherein the step of each Deconv layer is 2 x 2, and each previous layer is subjected to upsampling; inserting a self-attention module between layers, wherein the self-attention module refers to SAGAN, and adding a self-attention mechanism can enable the model to pay more attention to global characteristics, and a specific formula is as follows:
y i ＝γo i +x i
wherein, the first and the second end of the pipe are connected with each other,β j,i the degree of influence of the model on the ith position when synthesizing the jth region is shown, and the output of the attention layer is
Gamma is a proportional parameter, x i Is a key map;
the output layers are specifically sized 32 × 32, performing Batch Normalization (BN) on all layers except the output of the generator G and activation using a rectifying linear unit (ReLU);
inverter T and discriminator D are both constructed in reverse from G; the difference between them is the shape of the final output layer and the insertion position of the self attention module, all of which use LeakyReLU activation, except for the output layer.
4. The method for detecting the surface defects of the industrial products based on the attention-generating antagonistic network as claimed in claim 1, characterized in that in step 3): the specific steps of the model training phase are as follows:
1) Gaussian noise is added to the image generated by the model G, and the specific formula is as follows:
in the formula (I), the compound is shown in the specification,
the image after the noise is superposed, X is a defect-free image of the industrial product, and H represents Gaussian noise obeying normal distribution;
2) Training a discriminator D, stopping training and storing parameters of the discriminator D when the average likelihood of a batch of verification data sets reaches a threshold value P during each optimization iteration, wherein the specific formula is as follows:
where n is the batch size and P is an empirically determined threshold;
3) Training generator G, step 3.2 and training generator G continuously, the network parameters are trained with the following loss function:
where E represents the mathematical expectation of real data x and potential spatial data z; the G network is a generator, P z (z) is potential spatial data distribution, generally Gaussian distribution, to obtain a distribution Pg (x) for generating data, and we hope that the Pg (x) is very close to Pdata (x) to fit and approximate to real distribution, and a D network is a discriminant function, so that the traditional two-class problem needs to be solved, and the responsibility is to effectively distinguish the real distribution from the generated distribution, namely, measure the difference between the Pg (x) and the Pdata (x), and carry out repeated iterative training;
4) Training an inverter T, the inverter T defining an inverse mapping T (x) with respect to the generator G, x → z, mapping the normal image block x back to the latent space, specifically optimizing the inverter T using an L2 norm loss function, defined as:
in the optimization process, the parameters of G are fixed, and only the parameters of E are updated.
5. The method for detecting the surface defects of the industrial products based on the attention-generating antagonistic network as claimed in claim 1, characterized in that in the step 4): the defect detection stage comprises the following specific steps:
1) Extracting a set of image patches from a given query image g, where the patch size is the same as the patch size in the training set, being 32 x 32, given a patch x (i, j) located at (i, j) in g, discriminator D estimates the probability that this patch belongs to the normal category, and then constructs a likelihood map s of g as:
s(i,j)＝D(x(i,j))
2) The generator G and the inverter T learned in step 3 are jointly used to reconstruct the input image patch x' (i, j), with the following specific formula:
x'(i,j)＝G(T(x(i,j)))
reorganizing all the reconstructed small blocks x '(i, j) to form a full-size reconstructed image f';
3) Carrying out normalization operation on the reconstructed image f', wherein a specific formula is as follows:
wherein u (.) is the average of the image and σ (.) is the standard deviation of the image;
4) Constructing a residual error graph, wherein the specific formula is as follows:
r＝|f-f r |⊙f err
wherein an is an element matrix multiplication, f err Is an error map composed of block-by-block computed reconstruction errors;
5) And (3) fusing the likelihood map s constructed in the step (4.1) and the residual map r constructed in the step (4.4) to obtain a fusion map u, wherein the specific formula is as follows:
u(i,j)＝[1(i,j)-s(i,j)]⊙r(i,j)
wherein 1 (i, j) -s (i, j) represents that the likelihood map s is inverted to make the background intensity close to zero, so that the fusion map shows a cleaner background than the corresponding residual map;
6) Carrying out threshold processing on the fusion graph, wherein a specific formula is as follows:
wherein u is 0 Is a reference fusion map obtained from a defect-free sample, and t is an empirically determined control constant.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211056663.6A CN115393333A (en) | 2022-08-31 | 2022-08-31 | Industrial product surface defect detection method based on attention generation countermeasure network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211056663.6A CN115393333A (en) | 2022-08-31 | 2022-08-31 | Industrial product surface defect detection method based on attention generation countermeasure network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115393333A true CN115393333A (en) | 2022-11-25 |
Family
ID=84124339
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211056663.6A Pending CN115393333A (en) | 2022-08-31 | 2022-08-31 | Industrial product surface defect detection method based on attention generation countermeasure network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115393333A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117952983A (en) * | 2024-03-27 | 2024-04-30 | 中电科大数据研究院有限公司 | Intelligent manufacturing production process monitoring method and system based on artificial intelligence |
-
2022
- 2022-08-31 CN CN202211056663.6A patent/CN115393333A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117952983A (en) * | 2024-03-27 | 2024-04-30 | 中电科大数据研究院有限公司 | Intelligent manufacturing production process monitoring method and system based on artificial intelligence |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107169956B (en) | Color woven fabric defect detection method based on convolutional neural network | |
| CN108492281B (en) | Bridge crack image obstacle detection and removal method based on generation type countermeasure network | |
| CN112329588B (en) | Pipeline fault detection method based on Faster R-CNN | |
| CN112132012B (en) | High-resolution SAR ship image generation method based on generation countermeasure network | |
| CN111161224A (en) | Casting internal defect grading evaluation system and method based on deep learning | |
| CN111914902B (en) | Traditional Chinese medicine identification and surface defect detection method based on deep neural network | |
| CN110378254B (en) | Method and system for identifying vehicle damage image modification trace, electronic device and storage medium | |
| CN111598098A (en) | Water gauge water line detection and effectiveness identification method based on full convolution neural network | |
| CN111382785A (en) | GAN network model and method for realizing automatic cleaning and auxiliary marking of sample | |
| CN111815526B (en) | Rain image rainstrip removing method and system based on image filtering and CNN | |
| CN116071709A (en) | Crowd counting method, system and storage medium based on improved VGG16 network | |
| CN113420619A (en) | Remote sensing image building extraction method | |
| CN115512222A (en) | Method for evaluating damage of ground objects in disaster scene of offline training and online learning | |
| CN115272777A (en) | Semi-supervised image analysis method for power transmission scene | |
| CN115393333A (en) | Industrial product surface defect detection method based on attention generation countermeasure network | |
| CN114357372A (en) | Aircraft fault diagnosis model generation method based on multi-sensor data driving | |
| Shit et al. | An encoder‐decoder based CNN architecture using end to end dehaze and detection network for proper image visualization and detection | |
| CN115456957B (en) | Method for detecting change of remote sensing image by full-scale feature aggregation | |
| CN115719363B (en) | Environment sensing method and system capable of performing two-dimensional dynamic detection and three-dimensional reconstruction | |
| CN115690001A (en) | Method for detecting defects in steel pipe welding digital radiographic image | |
| CN115966006A (en) | Cross-age face recognition system based on deep learning model | |
| CN113077458B (en) | Cloud and shadow detection method and system in remote sensing image | |
| CN117372720B (en) | Unsupervised anomaly detection method based on multi-feature cross mask repair | |
| CN117523205B (en) | Segmentation and identification method for few-sample ki67 multi-category cell nuclei | |
| CN115830461B (en) | Hyperspectral image recognition method based on depth sequence convolution network |
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 |