WO2024007602A1 - 一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法及装置 - Google Patents
一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法及装置 Download PDFInfo
- Publication number
- WO2024007602A1 WO2024007602A1 PCT/CN2023/079045 CN2023079045W WO2024007602A1 WO 2024007602 A1 WO2024007602 A1 WO 2024007602A1 CN 2023079045 W CN2023079045 W CN 2023079045W WO 2024007602 A1 WO2024007602 A1 WO 2024007602A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- image
- defect
- photovoltaic panel
- images
- normal
- Prior art date
Links
- 230000007547 defect Effects 0.000 title claims abstract description 270
- 238000012549 training Methods 0.000 title claims abstract description 112
- 238000000034 method Methods 0.000 title claims abstract description 63
- 230000006870 function Effects 0.000 claims description 82
- 230000002950 deficient Effects 0.000 claims description 66
- 238000005457 optimization Methods 0.000 claims description 12
- 238000004590 computer program Methods 0.000 claims description 9
- 238000001514 detection method Methods 0.000 abstract description 12
- 238000010586 diagram Methods 0.000 description 7
- 238000005516 engineering process Methods 0.000 description 4
- 238000013507 mapping Methods 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 239000000428 dust Substances 0.000 description 2
- 238000010248 power generation Methods 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 238000013256 Gubra-Amylin NASH model Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 210000003608 fece Anatomy 0.000 description 1
- 229910017053 inorganic salt Inorganic materials 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
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
- G06T7/001—Industrial image inspection using an image reference 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
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- 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
- G06T2207/20084—Artificial neural networks [ANN]
Definitions
- This application relates to the field of artificial intelligence technology, and in particular to a training method and device for a CycleGAN model used to generate defect images of photovoltaic panels.
- photovoltaic panel defects can be detected through artificial intelligence image recognition methods.
- the accuracy of photovoltaic panel defect detection is low. Therefore, how to solve the problem of few photovoltaic panel defect images and thereby improve the accuracy of photovoltaic panel defect detection has become an urgent problem to be solved.
- This application provides a training method for a CycleGAN model used to generate photovoltaic panel defect images.
- a CycleGAN model used to generate photovoltaic panel defect images Through the target CycleGAN model used to generate photovoltaic panel defect images, normal images of photovoltaic panels can be converted into different types of photovoltaic panel defect images. This solves the problem of fewer photovoltaic panel defect images and improves the accuracy of photovoltaic panel defect detection.
- a training method for a CycleGAN model for generating photovoltaic panel defect images including: obtaining training samples, wherein the training samples include normal sample images and defective sample images of photovoltaic panels; The training samples are input into the CycleGAN model to be trained for generating photovoltaic panel defect images, and candidate images of the photovoltaic panels output by the CycleGAN model are obtained, where the candidate images include normal candidate images of the photovoltaic panels and Defect candidate images of photovoltaic panels; train the CycleGAN model according to the training samples and the candidate images to obtain a target CycleGAN model for generating photovoltaic panel defect images.
- inputting the training sample into a CycleGAN model to be trained for generating photovoltaic panel defect images, and obtaining candidate images of photovoltaic panels output by the CycleGAN model further includes: according to The normal sample image generates a corresponding first defective candidate image and a third normal candidate image, and the corresponding first normal candidate image is generated based on the first defective candidate image; and the corresponding third normal candidate image is generated based on the defective sample image.
- two normal candidate images and a third defective candidate image and generate a corresponding second defective candidate image according to the second normal candidate image.
- the CycleGAN model includes a first generator and a second generator, and the method further includes:
- the normal sample image is input to the first generator, the first generator outputs the first defect candidate image according to the normal sample image, and the first defect candidate image is input to the second A generator, the second generator outputs the first normal candidate image according to the first defect candidate image, and inputs the normal sample image into the second generator, and the second generator outputs the first normal candidate image according to the first defect candidate image.
- the normal sample image outputs the third normal candidate image.
- the CycleGAN model includes a first discriminator and a second discriminator
- the method further includes: according to the first discriminator, comparing the normal sample image of the photovoltaic panel and the photovoltaic The normal candidate image of the photovoltaic panel is discriminated; according to the second discriminator, the defect sample image of the photovoltaic panel and the defect candidate image of the photovoltaic panel are discriminated.
- training the cycle generative adversarial network CycleGAN model according to the training sample and the candidate image to obtain a target CycleGAN model for generating photovoltaic panel defect images further comprising: based on the The discrimination results of the first discriminator and the second discriminator are used to obtain the total loss function of the CycleGAN model; adjust the model parameters of the CycleGAN model based on the total loss function, and continue with the adjusted CycleGAN model for the next time Train until the training end conditions are met to obtain the target CycleGAN model.
- a photovoltaic panel defect image generation method based on the CycleGAN model including: acquiring a photovoltaic panel image to be processed; inputting the photovoltaic panel image to be processed into a computer for generating photovoltaic panel defects Image target CycleGAN model to obtain defect images of photovoltaic panels.
- a training device for a CycleGAN model that generates photovoltaic panel defect images including: a first acquisition module for acquiring training samples, wherein the training samples include normal images of photovoltaic panels. Sample images and defective sample images; an output module for inputting the training samples into the CycleGAN model to be trained for generating photovoltaic panel defect images, and obtaining candidates for the photovoltaic panels output by the CycleGAN model. images, wherein the candidate images include normal candidate images of photovoltaic panels and defective candidate images of photovoltaic panels; a training module configured to train the CycleGAN model based on the training samples and the candidate images to obtain Targeted CycleGAN model for generating photovoltaic panel defect images.
- the output module is further configured to: generate a corresponding first defect candidate image and a third normal candidate image according to the normal sample image, and generate a corresponding first defect candidate image according to the first defect candidate image.
- the first normal candidate image ; generates a corresponding second normal candidate image and a third defective candidate image based on the defective sample image, and generates a corresponding second defective candidate image based on the second normal candidate image.
- the CycleGAN model includes a first generator and a second generator, and the output module is further configured to: input the normal sample image into the first generator, and generate it from the third generator.
- a generator outputs the first defect candidate image according to the normal sample image, and inputs the first defect candidate image into the second generator, and the second generator outputs the first defect candidate image according to the first defect candidate image.
- the first normal candidate image is output, the normal sample image is input to the second generator, and the second generator outputs the third normal candidate image according to the normal sample image.
- the CycleGAN model includes a first discriminator and a second discriminator, and the device is further configured to: according to the first discriminator, compare normal sample images of photovoltaic panels and normal candidates of photovoltaic panels. Images are judged. According to the second discriminator, the defect sample image of the photovoltaic panel and the defect candidate image of the photovoltaic panel are distinguished.
- the training module is further configured to: obtain the total loss function of the CycleGAN model based on the discrimination results of the first discriminator and the second discriminator; based on the total loss function Adjust the model parameters of the CycleGAN model, and continue the next training of the adjusted CycleGAN model until the training end conditions are met, and the target CycleGAN model is obtained.
- a photovoltaic panel defect image generation device based on the CycleGAN model, including: an acquisition module for acquiring an image of the photovoltaic panel to be processed; an input module to generate the photovoltaic panel defects to be processed The image is input into the target CycleGAN model used to generate photovoltaic panel defect images to obtain photovoltaic panel defect images.
- the seventh aspect of the present application proposes a computer program product, including a computer program that, when executed by a processor, implements the CycleGAN model for generating photovoltaic panel defect images according to the first aspect
- the training method or the photovoltaic panel defect image generation method based on the CycleGAN model described in the second aspect proposes a computer program product, including a computer program that, when executed by a processor, implements the CycleGAN model for generating photovoltaic panel defect images according to the first aspect
- This application provides a training method for a CycleGAN model used to generate photovoltaic panel defect images.
- a CycleGAN model used to generate photovoltaic panel defect images Through the target CycleGAN model used to generate photovoltaic panel defect images, normal images of photovoltaic panels can be converted into different types of photovoltaic panel defect images. This solves the problem of fewer photovoltaic panel defect images and improves the accuracy of photovoltaic panel defect detection.
- Figure 1 is a schematic flowchart of a training method for a CycleGAN model for generating photovoltaic panel defect images provided by an embodiment of the present application;
- Figure 2 is a schematic flow chart of another CycleGAN model training method for generating photovoltaic panel defect images provided by an embodiment of the present application;
- Figure 4 is a schematic structural diagram of a CycleGAN model provided by an embodiment of the present application.
- Figure 6 is a schematic flow chart of another CycleGAN model training method for generating photovoltaic panel defect images provided by an embodiment of the present application
- Figure 7 is a schematic flow chart of another CycleGAN model training method for generating photovoltaic panel defect images provided by an embodiment of the present application.
- Figure 8 is a schematic flow chart of a photovoltaic panel defect image generation method based on the CycleGAN model provided by an embodiment of the present application;
- Figure 9 is a schematic structural diagram of a model training device provided by an embodiment of the present application.
- Figure 10 is a schematic structural diagram of a photovoltaic panel defect image generation device based on the CycleGAN model provided by an embodiment of the present application;
- FIG. 11 is a schematic structural diagram of an electronic device provided by an embodiment of the present application.
- Figure 1 is a schematic flowchart of a training method for a CycleGAN model for generating photovoltaic panel defect images provided by an embodiment of the present application.
- this embodiment proposes a training method for a CycleGAN model for generating photovoltaic panel defect images.
- the CycleGAN model is a cycle generative adversarial network model, which specifically includes the following steps:
- the training samples can be annotated and divided into normal sample images and defective sample images of the photovoltaic panel.
- S102 Input the training sample into the CycleGAN model to be trained to generate defect images of photovoltaic panels, and obtain candidate images of photovoltaic panels output by the CycleGAN model, where the candidate images include normal candidate images of photovoltaic panels and Defect candidate images of photovoltaic panels.
- the training samples can be input into the CycleGAN model, and then the normal candidate images of the photovoltaic panels and the defect candidate images of the photovoltaic panels output by the CycleGAN model are obtained.
- the CycleGAN model includes 2 generators, namely the first generator and the second generator.
- the normal sample image x of the photovoltaic panel in the training sample into the first generator in the CycleGAN model, and obtain the first defect candidate image x' corresponding to the normal sample image x.
- the defective sample image y of the board is input into the second generator in the CycleGAN model, and the second normal candidate image y' corresponding to the defective sample image y can be obtained.
- the first normal candidate image x" corresponding to the first defect candidate image x' can be obtained, and the second normal candidate image y' can be obtained.
- the second defective candidate image y" corresponding to the second normal candidate image y' can be obtained.
- the CycleGAN model can be trained based on the training samples and candidate images, and then the target CycleGAN model for generating photovoltaic panel defect images is obtained.
- the normal sample images of the photovoltaic panels can be converted into different types of photovoltaic panel defect images based on the target CycleGAN model.
- the training method of the CycleGAN model for generating photovoltaic panel defect images is to obtain training samples, where the training samples include normal sample images and defective sample images of photovoltaic panels, and input the training samples to the to-be-trained In the CycleGAN model that generates photovoltaic panel defect images, the candidate images of the photovoltaic panels output by the CycleGAN model are obtained.
- the candidate images include normal candidate images of the photovoltaic panels and defective candidate images of the photovoltaic panels, and are based on the training samples and The candidate images train the CycleGAN model to obtain the target CycleGAN model for generating photovoltaic panel defect images.
- this application can convert normal images of photovoltaic panels into different types of photovoltaic panel defect images through the target CycleGAN model used to generate photovoltaic panel defect images, solving the problem of fewer photovoltaic panel defect images and improving the quality of photovoltaic panels. Defect detection accuracy.
- the normal image of the photovoltaic panel output by the CycleGAN model can be obtained based on the first generator and the second generator in the CycleGAN model.
- Candidate images and defect candidate images can be obtained based on the first generator and the second generator in the CycleGAN model.
- the training sample is input into the CycleGAN model, and the specific process of obtaining the candidate image of the photovoltaic panel output by the CycleGAN model includes the following step:
- the normal sample image can be input into the first generator, the first generator outputs the first defect candidate image based on the normal sample image, and the first defect candidate image is input into the second generator, and the second generator outputs the first defect candidate image according to the normal sample image.
- the generator outputs a first normal candidate image based on the first defective candidate image.
- the normal sample image can also be input into the second generator, and the second generator outputs the third normal candidate image based on the normal sample image.
- the defective sample image can be input into the second generator, and the second generator can output the second normal candidate image according to the defective sample image, and the second normal candidate image can be input into the first generator, and the second generator can output the second normal candidate image according to the defective sample image.
- the generator outputs a second defective candidate image based on the second normal candidate image.
- the defect sample image can also be input into the first generator, and the first generator outputs the third defect candidate image according to the defect sample image.
- the CycleGAN model includes two discriminators, namely the first discriminator and the second discriminator. After obtaining the candidate image of the photovoltaic panel output by the CycleGAN model, it can be based on the first discriminator and the second discriminator.
- the second discriminator distinguishes the normal sample image of the photovoltaic panel and the normal candidate image of the photovoltaic panel, and discriminates the defect sample image of the photovoltaic panel and the defect candidate image of the photovoltaic panel.
- the specific process of distinguishing normal candidate images and defective candidate images of photovoltaic panels in the above steps includes the following steps:
- the second discriminator can be used to distinguish defect sample images and defect candidate images.
- the generator G is the first generator
- the generator F is the second generator
- the discriminator Dx is the first discriminator
- the discriminator Dy is the second discriminator.
- the normal sample image x is input into the generator G (the first generator), the first defect candidate image x' can be obtained, and then the first defect candidate image x'
- the candidate image x' is input into the generator F (the second generator), and the first normal candidate image x" can be obtained.
- the normal sample image x is input into the generator F (the second generator), and the third normal candidate image can be obtained.
- the discriminator Dx (the first discriminator) can be used to distinguish the normal sample image x and the normal candidate images (the first normal candidate image x", the second normal candidate image y' and the third normal candidate image x" ')
- the discriminator Dy (second discriminator) can be used to distinguish the defect sample image y and the defect candidate image (the first defect candidate image x', the second defect candidate image y" and the third defect candidate image y"') .
- the defective sample image y is input into the generator F (the second generator), the second normal candidate image y' can be obtained, and then the second normal candidate image y'
- the candidate image y' is input into the generator G (the first generator), and the second defect candidate image y" can be obtained.
- the defect sample y image is input into the generator G (the first generator), and the third defect candidate can be obtained.
- the discriminator Dx (the first discriminator) can be used to distinguish the normal sample image x and the normal candidate image (the first normal candidate image x", the second normal candidate image y' and the third normal candidate image x" ')
- the discriminator Dy (second discriminator) can be used to distinguish the defect sample image y and the defect candidate image (the first defect candidate image x', the second defect candidate image y" and the third defect candidate image y"') .
- the CycleGAN model is trained based on the training samples and candidate images to obtain the target CycleGAN model for generating photovoltaic panel defect images.
- the specific process includes the following steps:
- the first loss function of the CycleGAN model, Loss GAN means that the defect candidate image generated by the generator G (the first generator) is judged as a defective sample image by the discriminator Dy (the second discriminator) as much as possible; so that the generator F The normal candidate image generated by the (second generator) is judged as a normal sample image by the discriminator Dx (first discriminator) as much as possible.
- the second loss function of the CycleGAN model means that the normal sample image x passes through the generator G (the first generator) to generate the first defect candidate image x', and then passes through the generator F (the second generator) to generate the first normal candidate image. x", so that the generated normal candidate image is as similar as possible to the normal sample image.
- the generator G the first generator
- the generator F the second generator
- the generator G the first generator
- the second loss function Loss Cycle of the CycleGAN model can be obtained with the following formula:
- S703 Obtain the third loss function of the CycleGAN model based on the similarity between the normal sample image and the third normal candidate image, and based on the similarity between the defect sample image and the third defect candidate image.
- the third loss function of the CycleGAN model can be obtained with the following formula:
- the total loss function Loss CycleGAN of the CycleGAN model can be obtained.
- the total loss function Loss CycleGAN Loss GAN + ⁇ 1LossCycle+ ⁇ 2 Loss Identity .
- ⁇ 1 and ⁇ 2 can be set according to the degree of influence on the total loss function of the CycleGAN model.
- the target optimization function can be obtained, and the model parameters of the CycleGAN generative adversarial network model can be adjusted according to the target optimization function and the total loss function.
- Loss CycleGAN is the total loss function
- the adjusted CycleGAN model can be trained for the next time until the training end conditions are met, and the target CycleGAN model is obtained.
- the training end conditions can be set according to the actual situation, and are not limited in this application.
- the training end condition can be set as the number of training times reaches the preset number of training times.
- the mapping relationship between the photovoltaic panel normal image and different types of defect images can be obtained based on the target CycleGAN model for generating photovoltaic panel defect images.
- the normal image of the photovoltaic panel can be converted into the defect images of different types to obtain a large number of defect images of the photovoltaic panel.
- the training method of the CycleGAN model for generating photovoltaic panel defect images can obtain the mapping relationship between the normal image of the photovoltaic panel and different types of defect images through the target CycleGAN model that generates the photovoltaic panel defect image, and then Normal images of photovoltaic panels are converted into different types of defect images, and a large number of defect images of photovoltaic panels are obtained, which solves the problem of fewer defect images of photovoltaic panels, improves the accuracy of defect detection of photovoltaic panels, and thereby improves the power generation efficiency of photovoltaic modules.
- FIG. 8 is a schematic flowchart of the photovoltaic panel defect image generation method based on the CycleGAN model provided by the embodiment of the present application.
- the photovoltaic panel image to be processed may be a normal image of the photovoltaic panel.
- the photovoltaic panel defect image to be processed is input into the target CycleGAN model used to generate photovoltaic panel defect images, and a large number of photovoltaic panel defect images can be obtained.
- the photovoltaic panel defect image generation method based on the CycleGAN model provided by this application obtains the photovoltaic panel image to be processed and inputs the photovoltaic panel defect image to be processed into the target CycleGAN model used to generate the photovoltaic panel defect image to obtain Image of photovoltaic panel defects. Therefore, this application can convert normal images of photovoltaic panels into different types of photovoltaic panel defect images based on the target CycleGAN model that generates photovoltaic panel defect images, solving the problem of fewer photovoltaic panel defect images and improving the detection of photovoltaic panel defects. Detection accuracy.
- FIG. 9 is a schematic structural diagram of a model training device provided by an embodiment of the present application.
- the training device 1000 for the CycleGAN model used to generate photovoltaic panel defect images includes: a first acquisition module 110, an output module 120 and a second acquisition module 130.
- the first acquisition module 110 is used to acquire training samples, where the training samples include normal sample images and defective sample images of photovoltaic panels;
- the output module 120 is used to input the training sample into the CycleGAN model to be trained for generating photovoltaic panel defect images, and obtain the candidate image of the photovoltaic panel output by the CycleGAN model, wherein, the The candidate images include normal candidate images of photovoltaic panels and defective candidate images of photovoltaic panels;
- the training module 130 is configured to train the Cycle GAN model based on the training sample and the candidate image to obtain a target CycleGAN model for generating photovoltaic panel defect images.
- the output module 120 is further configured to: generate a corresponding first defect candidate image and a third normal candidate image according to the normal sample image, and generate a corresponding first defect candidate image according to the first defect candidate image. the first normal candidate image; generate corresponding second normal candidate images and third defective candidate images based on the defective sample image, and generate corresponding second defective candidate images based on the second normal candidate image.
- the CycleGAN model includes a first generator and a second generator
- the output module 120 is also used to: input the normal sample image into the first generator, and use the The first generator outputs the first defect candidate image according to the normal sample image, and inputs the first defect candidate image into the second generator, and the second generator outputs the first defect candidate image according to the first defect candidate image.
- Image output is the first normal candidate image, and a normal sample image is input to the second generator, and the second generator outputs the third normal candidate image according to the normal sample image.
- the defective sample image is input to the second generator, the second generator outputs the second normal candidate image according to the defective sample image, and the second normal candidate image is input to the first generator
- the first generator outputs the second defect candidate image according to the second normal candidate image, and inputs the defect sample image into the first generator, and the first generator outputs the second defect candidate image according to the defect
- the sample image outputs the third defect candidate image.
- the CycleGAN model includes a first discriminator and a second discriminator
- the device 1000 is also used to: according to the first discriminator, compare the normal sample image of the photovoltaic panel and the normal image of the photovoltaic panel. Candidate images are judged.
- the second discriminator the defect sample image of the photovoltaic panel and the defect candidate image of the photovoltaic panel are distinguished.
- the training module 130 is further configured to: obtain the total loss function of the CycleGAN model based on the discrimination results of the first discriminator and the second discriminator; based on the total loss The function adjusts the model parameters of the CycleGAN model, and continues the next training of the adjusted CycleGAN model until the training end conditions are met, and the target CycleGAN model is obtained.
- the training module 130 is further configured to: obtain the determination results based on the normal sample image and the second normal candidate image and the defective sample image and the first defect candidate image.
- the first loss function of the CycleGAN model obtains the second loss function of the CycleGAN model based on the similarity between the normal sample image and the first normal candidate image and based on the similarity between the defective sample image and the second defective candidate image ;
- According to the similarity between the normal sample image and the third normal candidate image and according to the similarity between the defective sample image and the third defective candidate image obtain the third loss function of the CycleGAN model; according to the first The loss function, the second loss function and the third loss function are used to obtain the total loss function of the CycleGAN model.
- the device 1000 is further configured to: obtain a target optimization function; and adjust model parameters of the CycleGAN model according to the target optimization function and the total loss function.
- the training device provided by this application for generating a CycleGAN model of photovoltaic panel defect images obtains training samples, where the training samples include normal sample images and defective sample images of photovoltaic panels, and inputs the training samples to the system to be trained.
- the candidate images of the photovoltaic panels output by the CycleGAN model are obtained.
- the candidate images include normal candidate images of the photovoltaic panels and defective candidate images of the photovoltaic panels, and are based on the training samples and
- the candidate images train the CycleGAN model to obtain the target CycleGAN model for generating photovoltaic panel defect images.
- this application can convert normal images of photovoltaic panels into different types of photovoltaic panel defect images through the target CycleGAN model used to generate photovoltaic panel defect images, solving the problem of fewer photovoltaic panel defect images and improving the quality of photovoltaic panels. Defect detection accuracy.
- this embodiment provides a photovoltaic panel defect image generation device based on the CycleGAN model.
- Figure 10 is a schematic structural diagram of a photovoltaic panel defect image generation device based on the CycleGAN model provided in an embodiment of the present application.
- the training device 2000 for the CycleGAN model used to generate photovoltaic panel defect images includes: a second acquisition module 210 and an input module 220.
- the acquisition module 210 is used to acquire the photovoltaic panel image to be processed
- the input module 220 inputs the photovoltaic panel defect image to be processed into the target CycleGAN model used to generate the photovoltaic panel defect image to obtain the photovoltaic panel defect image.
- the photovoltaic panel defect image generation device based on the CycleGAN model provided by this application acquires the photovoltaic panel image to be processed and inputs the photovoltaic panel defect image to be processed into the target CycleGAN model used to generate the photovoltaic panel defect image to obtain the photovoltaic panel defect image. Board defect image. Therefore, this application can convert normal images of photovoltaic panels into different types of photovoltaic panel defect images based on the target CycleGAN model that generates photovoltaic panel defect images, solving the problem of fewer photovoltaic panel defect images and improving the detection of photovoltaic panel defects. Detection accuracy.
- this application also proposes an electronic device 3000, as shown in Figure 11, including: a memory 310, a processor 320, and a computer program stored on the memory 310 and executable on the processor 320, so When the processor executes the program, the training method of the CycleGAN model for generating photovoltaic panel defect images as described in the first aspect or the photovoltaic panel defect image generation method based on the CycleGAN model as described in the second aspect is implemented.
- the present application proposes a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause the computer to execute the method for generating photovoltaic panel defects described in the first aspect.
- this application also proposes a computer program product, including a computer program that, when executed by a processor, implements the training of the CycleGAN model for generating photovoltaic panel defect images described in the first aspect method or the photovoltaic panel defect image generation method based on the CycleGAN model described in the second aspect.
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)
- Photovoltaic Devices (AREA)
Abstract
本申请提供了一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法及装置,该方法包括:获取训练样本,将所述训练样本输入至待训练的用于生成光伏板缺陷图像的循环生成对抗网络CycleGAN模型中,获取由所述CycleGAN模型输出的光伏板的候选图像,并根据所述训练样本和所述候选图像对所述循环生成对抗网络CycleGAN模型进行训练,以获取用于生成光伏板缺陷图像的目标CycleGAN模型。由此,本申请通过用于生成光伏板缺陷图像的目标CycleGAN模型,可以将光伏板的正常图像转换成不同类型的光伏板的缺陷图像,解决了光伏板缺陷图像较少的问题,提高了光伏板缺陷检测的准确率。
Description
相关申请的交叉引用
本申请基于申请号为202210790763.5、申请日为2022年07月06日的中国专利申请提出,并要求该中国专利申请的优先权,该中国专利申请的全部内容在此引入本申请作为参考。
本申请涉及人工智能技术领域,特别涉及一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法及装置。
随着近年来光伏产业的飞速发展,光伏板的规模已经十分庞大,由于集中式光伏场站大多建设在偏远地区,相关人员也不易到达,光伏板的运行维护成为影响光伏产业发展的主要因素之一,在光伏板的运行过程中,可能存在表面破损、灰尘、遮挡、脏污、热斑等缺陷,导致光伏组件发电效率不高,严重地会损坏光伏板,影响光伏场站的安全运行并会造成一定的经济损失。
相关技术中,可以通过人工智能图像识别方法进行光伏板缺陷的检测,然而,由于光伏板缺陷图像较少,导致光伏板缺陷检测的准确率较低。因此,如何解决光伏板缺陷图像较少的问题,进而提高光伏板缺陷检测的准确率,已成为了亟待解决的问题。
发明内容
本申请提供了一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法,通过用于生成光伏板缺陷图像的目标CycleGAN模型,可以将光伏板的正常图像转换成不同类型的光伏板缺陷图像,解决了光伏板缺陷图像较少的问题,提高了光伏板缺陷检测的准确率。
根据本申请的第一方面,提供了一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法,包括:获取训练样本,其中,所述训练样本包括光伏板的正常样本图像和缺陷样本图像;将所述训练样本输入至待训练的用于生成光伏板缺陷图像的CycleGAN模型中,获取由所述CycleGAN模型输出的光伏板的候选图像,其中,所述候选图像包括光伏板的正常候选图像以及光伏板的缺陷候选图像;根据所述训练样本和所述候选图像对所述CycleGAN模型进行训练,以获取用于生成光伏板缺陷图像的目标CycleGAN模型。
另外,根据本申请上述实施例的一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法,
还可以具有如下附加的技术特征:
根据本申请的一个实施例,所述将所述训练样本输入至待训练的用于生成光伏板缺陷图像的CycleGAN模型中,获取由所述CycleGAN模型输出的光伏板的候选图像,还包括:根据所述正常样本图像生成对应的第一缺陷候选图像和第三正常候选图像,并根据所述第一缺陷候选图像生成对应的所述第一正常候选图像;根据所述缺陷样本图像生成对应的第二正常候选图像和第三缺陷候选图像,并根据所述第二正常候选图像生成对应的第二缺陷候选图像。
根据本申请的一个实施例,所述CycleGAN模型包括第一生成器和第二生成器,所述方法还包括:
将所述正常样本图像输入所述第一生成器,由所述第一生成器根据所述正常样本图像输出所述第一缺陷候选图像,并将所述第一缺陷候选图像输入所述第二生成器,由所述第二生成器根据所述第一缺陷候选图像输出所述第一正常候选图像,并将所述正常样本图像输入所述第二生成器,由所述第二生成器根据所述正常样本图像输出所述第三正常候选图像。
将所述缺陷样本图像输入所述第二生成器,由所述第二生成器根据所述缺陷样本图像输出所述第二正常候选图像,将所述第二正常候选图像输入所述第一生成器,由所述第一生成器根据所述第二正常候选图像输出所述第二缺陷候选图像,并将所述缺陷样本图像输入所述第一生成器,由所述第一生成器根据所述缺陷样本图像输出所述第三缺陷候选图像。
根据本申请的一个实施例,所述CycleGAN模型包括第一判别器和第二判别器,所述方法还包括:根据所述第一判别器,对所述光伏板的正常样本图像和所述光伏板的正常候选图像进行判别;根据所述第二判别器,对所述光伏板的缺陷样本图像和光伏板的缺陷候选图像进行判别。
根据本申请的一个实施例,根据所述训练样本和所述候选图像对所述循环生成对抗网络CycleGAN模型进行训练,以获取用于生成光伏板缺陷图像的目标CycleGAN模型,还包括:基于所述第一判别器和所述第二判别器的判别结果,获取所述CycleGAN模型的总损失函数;基于所述总损失函数调整所述CycleGAN模型的模型参数,并对调整后的CycleGAN模型继续下一次训练,直至满足训练结束条件,得到所述目标CycleGAN模型。
根据本申请的一个实施例,所述基于所述第一判别器和所述第二判别器的判别结果,获取所述CycleGAN模型的总损失函数,包括:根据所述正常样本图像和所述第二正常候选图像以及根据所述缺陷样本图像和第一缺陷候选图像的判别结果,获取所述CycleGAN模型的第一损失函数;根据所述正常样本图像和所述第一正常候选图像以及根据所述缺陷样本图像和第二缺陷候选图像的相似度,获取所述CycleGAN模型的第二损失函数;根据所述正常样本图像和所述第三正常候选图像以及根据所述缺陷样本图像和所述第三缺陷候选图像的相似度,获取所述CycleGAN模型的第三损失函数;根据所述第一损失函数、第二损失函数和第三损失函数,获取所述CycleGAN模型的所述总损失函数。
根据本申请的一个实施例,所述方法还包括:获取目标优化函数;根据所述目标优化函数和所述
总损失函数,调整所述CycleGAN模型的模型参数。
根据本申请的第二方面,提供了一种基于CycleGAN模型的光伏板缺陷图像生成方法,包括:获取待处理的光伏板图像;将所述待处理的光伏板图像输入至用于生成光伏板缺陷图像的目标CycleGAN模型中,以得到光伏板的缺陷图像。
根据本申请的第三方面,提供了一种用于生成光伏板缺陷图像的CycleGAN模型的训练装置,包括:第一获取模块,用于获取训练样本,其中,所述训练样本包括光伏板的正常样本图像和缺陷样本图像;输出模块,用于将所述训练样本输入至待训练的用于生成光伏板缺陷图像的循环生成对抗网络CycleGAN模型中,获取由所述CycleGAN模型输出的光伏板的候选图像,其中,所述候选图像包括光伏板的正常候选图像以及光伏板的缺陷候选图像;训练模块,用于根据所述训练样本和所述候选图像对所述CycleGAN模型进行训练,以获取用于生成光伏板缺陷图像的目标CycleGAN模型。
根据本申请上述实施例的一种用于生成光伏板缺陷图像的CycleGAN模型的训练装置,还可以具有如下附加的技术特征:
根据本申请的一个实施例,所述输出模块,还用于:根据所述正常样本图像生成对应的第一缺陷候选图像和第三正常候选图像,并根据所述第一缺陷候选图像生成对应的所述第一正常候选图像;根据所述缺陷样本图像生成对应的第二正常候选图像和第三缺陷候选图像,并根据所述第二正常候选图像生成对应的第二缺陷候选图像。
根据本申请的一个实施例,所述CycleGAN模型包括第一生成器和第二生成器,所述输出模块,还用于:将所述正常样本图像输入所述第一生成器,由所述第一生成器根据所述正常样本图像输出所述第一缺陷候选图像,并将所述第一缺陷候选图像输入所述第二生成器,由所述第二生成器根据所述第一缺陷候选图像输出所述第一正常候选图像,并将正常样本图像输入所述第二生成器,由所述第二生成器根据所述正常样本图像输出所述第三正常候选图像。将所述缺陷样本图像输入所述第二生成器,由所述第二生成器根据所述缺陷样本图像输出所述第二正常候选图像,将所述第二正常候选图像输入所述第一生成器,由所述第一生成器根据所述第二正常候选图像输出所述第二缺陷候选图像,并将缺陷样本图像输入所述第一生成器,由所述第一生成器根据所述缺陷样本图像输出所述第三缺陷候选图像。
根据本申请的一个实施例,所述CycleGAN模型包括第一判别器和第二判别器,所述装置,还用于:根据第一判别器,对光伏板的正常样本图像和光伏板的正常候选图像进行判别。根据第二判别器,对光伏板的缺陷样本图像和光伏板的缺陷候选图像进行判别。
本申请的一个实施例,所述训练模块,还用于:基于所述第一判别器和所述第二判别器的判别结果,获取所述CycleGAN模型的总损失函数;基于所述总损失函数调整所述CycleGAN模型的模型参数,并对调整后的CycleGAN模型继续下一次训练,直至满足训练结束条件,得到所述目标CycleGAN模型。
本申请的一个实施例,所述训练模块,还用于:根据所述正常样本图像和所述第二正常候选图像以及根据所述缺陷样本图像和第一缺陷候选图像的判别结果,获取所述CycleGAN模型的第一损失函数;根据所述正常样本图像和所述第一正常候选图像以及根据所述缺陷样本图像和第二缺陷候选图像的相似度,获取所述CycleGAN模型的第二损失函数;根据所述正常样本图像和所述第三正常候选图像以及根据所述缺陷样本图像和所述第三缺陷候选图像的相似度,获取所述CycleGAN模型的第三损失函数;根据所述第一损失函数、第二损失函数和第三损失函数,获取所述CycleGAN模型的所述总损失函数。
本申请的一个实施例,所述装置还用于:获取目标优化函数;根据所述目标优化函数和所述总损失函数,调整所述CycleGAN模型的模型参数。
根据本申请的第四方面,提供了一种基于CycleGAN模型的光伏板缺陷图像生成装置,包括:获取模块,用于获取待处理的光伏板图像;输入模块,将所述待处理的光伏板缺陷图像输入至用于生成光伏板缺陷图像的目标CycleGAN模型中,以得到光伏板缺陷图像。
为了实现上述目的,本申请第五方面提出了一种电子设备,其特征在于,包括:存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序,所述处理器执行所述程序时,实现如第一方面所述的用于生成光伏板缺陷图像的CycleGAN模型的训练方法或者第二方面所述的基于CycleGAN模型的光伏板缺陷图像生成方法。
为了实现上述目的,本申请第六方面提出了一种存储有计算机指令的非瞬时计算机可读存储介质,其中,所述计算机指令用于使所述计算机执行第一方面所述的用于生成光伏板缺陷图像的CycleGAN模型的训练方法或者第二方面所述的基于CycleGAN模型的光伏板缺陷图像生成方法。
为了实现上述目的,本申请第七方面提出了一种计算机程序产品,包括计算机程序,所述计算机程序在被处理器执行时实现根据第一方面所述的用于生成光伏板缺陷图像的CycleGAN模型的训练方法或者第二方面所述的基于CycleGAN模型的光伏板缺陷图像生成方法。
本申请实施例提供的技术方案至少包括如下有益效果:
本申请提供了一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法,通过用于生成光伏板缺陷图像的目标CycleGAN模型,可以将光伏板的正常图像转换成不同类型的光伏板缺陷图像,解决了光伏板缺陷图像较少的问题,提高了光伏板缺陷检测的准确率。
应当理解,本部分所描述的内容并非旨在标识本申请的实施例的关键或重要特征,也不用于限制本申请的范围。本申请的其它特征将通过以下的说明书而变得容易理解。
附图用于更好地理解本方案,不构成对本申请的限定。其中:
图1为本申请实施例提供的一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法的流程示意图;
图2为本申请实施例提供的另一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法的流程示意图;
图3为本申请实施例提供的另一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法的流程示意图;
图4为本申请实施例提供的一种CycleGAN模型的结构示意图;
图5为本申请实施例提供的一种CycleGAN模型的结构示意图;
图6为本申请实施例提供的另一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法的流程示意图;
图7为本申请实施例提供的另一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法的流程示意图;
图8为本申请实施例提供的一种基于CycleGAN模型的光伏板缺陷图像生成方法的流程示意图;
图9为本申请实施例提供的一种模型训练装置的结构示意图;
图10为本申请实施例提供的一种基于CycleGAN模型的光伏板缺陷图像生成装置的结构示意图;
图11为本申请实施例提供的一种电子设备的结构示意图。
以下结合附图对本申请的示范性实施例做出说明,其中包括本申请实施例的各种细节以助于理解,应当将它们认为仅仅是示范性的。因此,本领域普通技术人员应当认识到,可以对这里描述的实施例做出各种改变和修改,而不会背离本申请的范围和精神。同样,为了清楚和简明,以下的描述中省略了对公知功能和结构的描述。
下面采用实施例对本申请的用于生成光伏板缺陷图像的CycleGAN模型的训练方法进行详细说明。
图1为本申请实施例提供的用于生成光伏板缺陷图像的CycleGAN模型的训练方法的流程示意图。
如图1所示,本实施例提出的用于生成光伏板缺陷图像的CycleGAN模型的训练方法,所述CycleGAN模型为循环生成对抗网络模型,具体包括以下步骤:
S101、获取训练样本,其中,训练样本包括光伏板的正常样本图像和缺陷样本图像。
需要说明的是,在获取训练样本后,可以对训练样本进行标注,将训练样本划分为光伏板的正常样本图像和缺陷样本图像。
进一步地,光伏板的缺陷样本图像可以为光伏板表面破损、光伏板表面有灰尘、光伏板表面被遮挡、光伏板表面有脏污等。进一步地,光伏板表面被遮挡的类型可以分为光伏板表面被阴影遮挡、光伏板表面被植物遮挡、光伏板表面被建筑物遮挡、光伏板表面被其他组件遮挡等,进一步地光伏
板表面有脏污的类型可以分光伏板表面有鸟粪、光伏板表面有泥垢、光伏板表面有无机盐结垢等。
S102、将训练样本输入至待训练的用于生成光伏板缺陷图像的循环生成对抗网络CycleGAN模型中,获取由CycleGAN模型输出的光伏板的候选图像,其中,候选图像包括光伏板的正常候选图像以及光伏板的缺陷候选图像。
在本申请实施例中,在获取到训练样本后,可以将训练样本输入CycleGAN模型中,进而获取由CycleGAN模型输出的光伏板的正常候选图像以及光伏板的缺陷候选图像。
其中,CycleGAN模型包括2个生成器,即第一生成器和第二生成器。
举例而言,将训练样本中的光伏板的正常样本图像x,输入CycleGAN模型中的第一生成器,可以获取与正常样本图像x对应的第一缺陷候选图像x’,将训练样本中的光伏板的缺陷样本图像y,输入CycleGAN模型中的第二生成器,可以获取与缺陷样本图像y对应的第二正常候选图像y’。
进一步地,在获取到第一缺陷候选图像和第二正常候选图像后,可以输入CycleGAN模型中,以获取与第一缺陷候选图像对应的第一正常候选图像以及与第二正常候选图像对应的第二缺陷候选图像。
举例而言,将第一缺陷候选图像x’输入CycleGAN模型中的第二生成器,可以获取与第一缺陷候选图像x’对应的第一正常候选图像x”,将第二正常候选图像y’输入CycleGAN模型中的第一生成器,可以获取与第二正常候选图像y’对应的第二缺陷候选图像y”。
S103、根据训练样本和候选图像对CycleGAN模型进行训练,以获取用于生成光伏板缺陷图像的目标CycleGAN模型。
需要说明的是,相关技术中,光伏板的缺陷图像较少,本申请可以基于目标CycleGAN模型,将大量的光伏板的正常图像转换成不同类型的缺陷候选图像,解决了现有技术中光伏板的缺陷图像较少的问题。
在本申请实施例中,在获取到训练样本和候选图像后,可以根据训练样本和候选图像对CycleGAN模型进行训练,进而获取用于生成光伏板缺陷图像的目标CycleGAN模型。
进一步地,在获取到用于生成光伏板缺陷图像的目标CycleGAN模型后,可以基于目标CycleGAN模型,将光伏板的正常样本图像转换成不同类型的光伏板缺陷图像。
本申请提供的用于生成光伏板缺陷图像的CycleGAN模型的训练方法,通过获取训练样本,其中,训练样本包括光伏板的正常样本图像和缺陷样本图像,并将训练样本输入至待训练的用于生成光伏板缺陷图像的循环生成对抗网络CycleGAN模型中,获取由CycleGAN模型输出的光伏板的候选图像,其中,候选图像包括光伏板的正常候选图像以及光伏板的缺陷候选图像,并根据训练样本和候选图像对CycleGAN模型进行训练,以获取用于生成光伏板缺陷图像的目标CycleGAN模型。由此,本申请通过用于生成光伏板缺陷图像的目标CycleGAN模型,可以将光伏板的正常图像转换成不同类型的光伏板缺陷图像,解决了光伏板缺陷图像较少的问题,提高了光伏板缺陷检测的准确率。
需要说明的是,本申请中,在试图获取由CycleGAN模型输出的光伏板的候选图像时,可以基于CycleGAN模型中的第一生成器和第二生成器,获取由CycleGAN模型输出的光伏板的正常候选图像以及缺陷候选图像。
作为一种可能的实现方式,如图2所示,在上述步骤的基础上,上述步骤S101中将训练样本输入CycleGAN模型中,获取由CycleGAN模型输出的光伏板的候选图像的具体过程,包括以下步骤:
S201、根据正常样本图像生成对应的第一缺陷候选图像和第三正常候选图像,并根据第一缺陷候选图像生成对应的第一正常候选图像。
在本申请实施例中,可以将正常样本图像输入第一生成器,由第一生成器根据正常样本图像输出第一缺陷候选图像,并将第一缺陷候选图像输入第二生成器,由第二生成器根据第一缺陷候选图像输出第一正常候选图像。
进一步地,在本申请实施例中,还可以将正常样本图像输入第二生成器,由第二生成器根据正常样本图像输出第三正常候选图像。
S202、根据缺陷样本图像生成对应的第二正常候选图像和第三缺陷候选图像,并根据第二正常候选图像生成对应的第二缺陷候选图像。
在本申请实施例中,可以将缺陷样本图像输入第二生成器,由第二生成器根据缺陷样本图像输出第二正常候选图像,并将第二正常候选图像输入第一生成器,由第一生成器根据第二正常候选图像输出第二缺陷候选图像。
进一步地,在本申请实施例中,还可以将缺陷样本图像输入第一生成器,由第一生成器根据缺陷样本图像输出第三缺陷候选图像。
需要说明的是,本申请中,CycleGAN模型包括两个判别器,即第一判别器和第二判别器,在获取到由CycleGAN模型输出的光伏板的候选图像后,可以基于第一判别器和第二判别器,对光伏板的正常样本图像和光伏板的正常候选图像进行判别,对光伏板的缺陷样本图像和光伏板的缺陷候选图像进行判别。
作为一种可能的实现方式,如图3所示,在上述步骤的基础上,上述步骤中对光伏板的正常候选图像以及缺陷候选图像进行判别的具体过程,包括以下步骤:
S301、根据第一判别器,对光伏板的正常样本图像和光伏板的正常候选图像进行判别。
需要说明的是,第一判别器,可以用于区分正常样本图像和正常候选图像。
S302、根据第二判别器,对光伏板的缺陷样本图像和光伏板的缺陷候选图像进行判别。
需要说明的是,第二判别器,可以用于区分缺陷样本图像和缺陷候选图像。
下面对基于CycleGAN模型获取光伏板的候选图像、对光伏板的正常样本图像和光伏板的正常候选图像进行判别以及对光伏板的缺陷样本图像和光伏板的缺陷候选图像进行判别的过程进行解
释说明。
举例而言,其中,生成器G即为第一生成器,生成器F即为第二生成器,判别器Dx即为第一判别器,判别器Dy即为第二判别器。
如图4所示,针对正常样本图像x,缺陷样本图像y,将正常样本图像x输入至生成器G(第一生成器)中,可以获取第一缺陷候选图像x’,再将第一缺陷候选图像x’输入至生成器F(第二生成器)中,可以获取第一正常候选图像x”,将正常样本图像x输入生成器F(第二生成器),可以获取第三正常候选图像像x”’,判别器Dx(第一判别器)可以用于区分正常样本图像x和正常候选图像(第一正常候选图像x”、第二正常候选图像y’和第三正常候选图像x”’),判别器Dy(第二判别器)可以用于区分缺陷样本图像y和缺陷候选图像(第一缺陷候选图像x’、第二缺陷候选图像y”和第三缺陷候选图像y”’)。
如图5所示,针对正常样本图像x,缺陷样本图像y,将缺陷样本图像y输入至生成器F(第二生成器)中,可以获取第二正常候选图像y’,再将第二正常候选图像y’输入至生成器G(第一生成器)中,可以获取第二缺陷候选图像y”,将缺陷样本y图像输入生成器G(第一生成器)中,可以获取第三缺陷候选图像y”’,判别器Dx(第一判别器)可以用于区分正常样本图像x和正常候选图像(第一正常候选图像x”、第二正常候选图像y’和第三正常候选图像x”’),判别器Dy(第二判别器)可以用于区分缺陷样本图像y和缺陷候选图像(第一缺陷候选图像x’、第二缺陷候选图像y”和第三缺陷候选图像y”’)。
进一步地,可以根据判别结果,获取CycleGAN模型的总损失函数,进而调整CycleGAN模型的模型参数,以得到目标CycleGAN模型。
作为一种可能的实现方式,如图6所示,在上述步骤的基础上,上述步骤S103中根据训练样本和候选图像对CycleGAN模型进行训练,以获取用于生成光伏板缺陷图像的目标CycleGAN模型的具体过程,包括以下步骤:
S601、基于第一判别器和第二判别器的判别结果,获取CycleGAN模型的总损失函数。
作为一种可能的实现方式,如图7所示,在上述步骤的基础上,上述步骤S601中基于第一判别器和第二判别器的判别结果,获取CycleGAN模型的总损失函数的具体过程,包括以下步骤:
S701、根据正常样本图像和正常候选图像的判别结果以及根据缺陷样本图像和缺陷候选图像的判别结果,获取CycleGAN模型的第一损失函数。
其中,CycleGAN模型的第一损失函数,LossGAN表示使生成器G(第一生成器)生成的缺陷候选图像尽可能被判别器Dy(第二判别器)判定为缺陷样本图像;使生成器F(第二生成器)生成的正常候选图像尽可能被判别器Dx(第一判别器)判定为正常样本图像。
其中,CycleGAN模型的第一损失函数LossGAN可以用以下公式进行获取:
S702、根据正常样本图像和第一正常候选图像的相似度以及缺陷样本图像和第二缺陷候选图像的相似度,获取CycleGAN模型的第二损失函数。
其中,CycleGAN模型的第二损失函数表示正常样本图像x经过生成器G(第一生成器)生成第一缺陷候选图像x’,再经过生成器F(第二生成器)生成第一正常候选图像x”,使生成的正常候选图像与正常样本图像尽可能相似,同理地,缺陷样本图像y经过生成器F(第二生成器)生成第二正常候选图像y’后,再经过生成器G(第一生成器)生成第二缺陷图像y”,使生成的缺陷候选图像与样本缺陷图像尽可能相似。
其中,CycleGAN模型的第二损失函数LossCycle可以用以下公式进行获取:
S703、根据正常样本图像和第三正常候选图像的相似度,并根据缺陷样本图像和第三缺陷候选图像的相似度,获取CycleGAN模型的第三损失函数。
其中,CycleGAN模型的第三损失函数LossIdentity表示保证生成器G、F生成的图像与输入的图像在内容上基本保持一致。即生成器F(第二生成器)的目的是生成正常图像,输入正常样本图像x,生成的图像F(x)应与正常样本图像x尽可能相似,同理地,生成器G(第二生成器)的目的是生成缺陷图像,输入缺陷样本图像y,生成的图像G(y)应与缺陷样本图像y尽可能相似。
其中,CycleGAN模型的第三损失函数可以用以下公式进行获取:
S704、根据第一损失函数、第二损失函数和第三损失函数,获取CycleGAN模型的总损失函数。
需要说明的是,根据第一损失函数LossGAN、第二损失函数LossCycleGAN和第三损失函数LossIdentity,可以获取CycleGAN模型的总损失函数LossCycleGAN。
举例而言,总损失函数LossCycleGAN=LossGAN+λ1LossCycle+λ2LossIdentity。
其中,λ1和λ2可以根据对CycleGAN模型的总损失函数的影响程度进行设定。
S602、基于总损失函数调整CycleGAN模型的模型参数,并对调整后的CycleGAN模型继续下一次训练,直至满足训练结束条件,得到目标CycleGAN模型。
在本申请实施例中,在获取到总损失函数后,可以获取目标优化函数,并根据目标优化函数和总损失函数,调整CycleGAN生成对抗网络模型的模型参数。
举例而言,在试图根据目标优化函数和总损失函数,调整CycleGAN生成对抗网络模型的模型参数时,可以根据以下公式进行调整:
其中,为目标优化函数、LossCycleGAN为总损失函数。
进一步地,可以对调整后的CycleGAN模型继续下一次训练,直至满足训练结束条件,得到目标CycleGAN模型。
其中,训练结束条件,可以根据实际情况进行设定,本申请不作限定。
可选地,可以设定训练结束条件为训练次数达到预设训练次数。
举例而言,可以设定训练结束条件为训练次数达到10000次。
进一步,在获取到生成光伏板缺陷图像的目标CycleGAN模型后,可以根据生成光伏板缺陷图像的目标CycleGAN模型获取光伏板正常图像和不同类型的缺陷图像之间的映射关系。
进一步地,在获取到光伏板正常图像和不同类型的缺陷图像之间的映射关系后,可以将光伏板正常图像转换成不同类型的缺陷图像,得到光伏板大量的缺陷图像。
本申请提供的用于生成光伏板缺陷图像的CycleGAN模型的训练方法,可以通过生成光伏板缺陷图像的目标CycleGAN模型,获取到光伏板正常图像和不同类型的缺陷图像之间的映射关系,进而将光伏板正常图像转换成不同类型的缺陷图像,得到光伏板大量的缺陷图像,解决了光伏板缺陷图像较少的问题,提高了光伏板缺陷检测的准确率,进而提高了光伏组件的发电效率。
下面采用实施例对本申请的基于CycleGAN模型的光伏板缺陷图像生成方法进行详细说明。
图8为本申请实施例提供的基于CycleGAN模型的光伏板缺陷图像生成方法的流程示意图。
S801、获取待处理的光伏板图像。
可选地,待处理的光伏板图像可以为光伏板的正常图像。
S802、将待处理的光伏板缺陷图像输入至用于生成光伏板缺陷图像的目标CycleGAN模型中,以得到光伏板缺陷图像。
在本申请实施例中,在获取到待处理图像后,将待处理的光伏板缺陷图像输入至用于生成光伏板缺陷图像的目标CycleGAN模型中,进而可以得到大量的光伏板缺陷图像。
本申请提供的基于CycleGAN模型的光伏板缺陷图像生成方法,通过获获取待处理的光伏板图像,将待处理的光伏板缺陷图像输入至用于生成光伏板缺陷图像的目标CycleGAN模型中,以得到光伏板缺陷图像。由此,本申请通过基于生成光伏板缺陷图像的目标CycleGAN模型,可以将光伏板的正常图像转换成不同类型的光伏板缺陷图像,解决了光伏板缺陷图像较少的问题,提高了光伏板缺陷检测的准确率。
为了实现上述实施例,本实施例提供了一种用于生成光伏板缺陷图像的CycleGAN模型的训练装置,图9为本申请实施例提供的一种模型训练装置的结构示意图。
如图9所示,该用于生成光伏板缺陷图像的CycleGAN模型的训练装置1000,包括:第一获取模块110,输出模块120和第二获取模块130。
第一获取模块110,用于获取训练样本,其中,所述训练样本包括光伏板的正常样本图像和缺陷样本图像;
输出模块120,用于将所述训练样本输入至待训练的用于生成光伏板缺陷图像的循环生成对抗网络CycleGAN模型中,获取由所述CycleGAN模型输出的光伏板的候选图像,其中,所述候选图像包括光伏板的正常候选图像以及光伏板的缺陷候选图像;
训练模块130,用于根据所述训练样本和所述候选图像对所述循环生成对抗网络CycleGAN模型进行训练,以获取用于生成光伏板缺陷图像的目标CycleGAN模型。
根据本申请的一个实施例,所述输出模块120,还用于:根据所述正常样本图像生成对应的第一缺陷候选图像和第三正常候选图像,并根据所述第一缺陷候选图像生成对应的所述第一正常候选图像;根据所述缺陷样本图像生成对应的第二正常候选图像和第三缺陷候选图像,并根据所述第二正常候选图像生成对应的第二缺陷候选图像。
根据本申请的一个实施例,所述CycleGAN模型包括第一生成器和第二生成器,所述输出模块120,还用于:将所述正常样本图像输入所述第一生成器,由所述第一生成器根据所述正常样本图像输出所述第一缺陷候选图像,并将所述第一缺陷候选图像输入所述第二生成器,由所述第二生成器根据所述第一缺陷候选图像输出所述第一正常候选图像,并将正常样本图像输入所述第二生成器,由所述第二生成器根据所述正常样本图像输出所述第三正常候选图像。将所述缺陷样本图像输入所述第二生成器,由所述第二生成器根据所述缺陷样本图像输出所述第二正常候选图像,将所述第二正常候选图像输入所述第一生成器,由所述第一生成器根据所述第二正常候选图像输出所述第二缺陷候选图像,并将缺陷样本图像输入所述第一生成器,由所述第一生成器根据所述缺陷样本图像输出所述第三缺陷候选图像。
根据本申请的一个实施例,所述CycleGAN模型包括第一判别器和第二判别器,所述装置1000,还用于:根据第一判别器,对光伏板的正常样本图像和光伏板的正常候选图像进行判别。根据第二判别器,对光伏板的缺陷样本图像和光伏板的缺陷候选图像进行判别。
本申请的一个实施例,所述训练模块130,还用于:基于所述第一判别器和所述第二判别器的判别结果,获取所述CycleGAN模型的总损失函数;基于所述总损失函数调整所述CycleGAN模型的模型参数,并对调整后的CycleGAN模型继续下一次训练,直至满足训练结束条件,得到所述目标CycleGAN模型。
本申请的一个实施例,所述训练模块130,还用于:根据所述正常样本图像和所述第二正常候选图像以及根据所述缺陷样本图像和第一缺陷候选图像的判别结果,获取所述CycleGAN模型的第一损失函数;根据所述正常样本图像和所述第一正常候选图像以及根据所述缺陷样本图像和第二缺陷候选图像的相似度,获取所述CycleGAN模型的第二损失函数;根据所述正常样本图像和所述第三正常候选图像以及根据所述缺陷样本图像和所述第三缺陷候选图像的相似度,获取所述CycleGAN模型的第三损失函数;根据所述第一损失函数、第二损失函数和第三损失函数,获取所述CycleGAN模型的所述总损失函数。
本申请的一个实施例,所述装置1000,还用于:获取目标优化函数;根据所述目标优化函数和所述总损失函数,调整所述CycleGAN模型的模型参数。
本申请提供的用于生成光伏板缺陷图像的CycleGAN模型的训练装置,通过获取训练样本,其中,训练样本包括光伏板的正常样本图像和缺陷样本图像,并将训练样本输入至待训练的用于生成光伏板缺陷图像的循环生成对抗网络CycleGAN模型中,获取由CycleGAN模型输出的光伏板的候选图像,其中,候选图像包括光伏板的正常候选图像以及光伏板的缺陷候选图像,并根据训练样本和候选图像对CycleGAN模型进行训练,以获取用于生成光伏板缺陷图像的目标CycleGAN模型。由此,本申请通过用于生成光伏板缺陷图像的目标CycleGAN模型,可以将光伏板的正常图像转换成不同类型的光伏板缺陷图像,解决了光伏板缺陷图像较少的问题,提高了光伏板缺陷检测的准确率。
为了实现上述实施例,本实施例提供了一种基于CycleGAN模型的光伏板缺陷图像生成装置,图10为本申请实施例提供的一种基于CycleGAN模型的光伏板缺陷图像生成装置的结构示意图。
如图10所示,该用于生成光伏板缺陷图像的CycleGAN模型的训练装置2000,包括:第二获取模块210和输入模块220。
获取模块210,用于获取待处理的光伏板图像;
输入模块220,将所述待处理的光伏板缺陷图像输入至用于生成光伏板缺陷图像的目标CycleGAN模型中,以得到光伏板缺陷图像。
本申请提供的基于CycleGAN模型的光伏板缺陷图像生成装置,通过获取待处理的光伏板图像,将待处理的光伏板缺陷图像输入至用于生成光伏板缺陷图像的目标CycleGAN模型中,以得到光伏板缺陷图像。由此,本申请通过基于生成光伏板缺陷图像的目标CycleGAN模型,可以将光伏板的正常图像转换成不同类型的光伏板缺陷图像,解决了光伏板缺陷图像较少的问题,提高了光伏板缺陷检测的准确率。
为了实现上述实施例,本申请还提出了一种电子设备3000,如图11所示,包括:存储器310、处理器320及存储在存储器310上并可在处理器上320运行的计算机程序,所述处理器执行所述程序时,实现如第一方面所述的用于生成光伏板缺陷图像的CycleGAN模型的训练方法或者第二方面所述的基于CycleGAN模型的光伏板缺陷图像生成方法。
为了实现上述实施例,本申请提出了一种存储有计算机指令的非瞬时计算机可读存储介质,其中,所述计算机指令用于使所述计算机执行第一方面所述的用于生成光伏板缺陷图像的CycleGAN模型的训练方法或者第二方面所述的基于CycleGAN模型的光伏板缺陷图像生成方法。
为了实现上述实施例,本申请还提出了一种计算机程序产品,包括计算机程序,所述计算机程序在被处理器执行时实现第一方面所述的用于生成光伏板缺陷图像的CycleGAN模型的训练方法或者第二方面所述的基于CycleGAN模型的光伏板缺陷图像生成方法。
应该理解,可以使用上面所示的各种形式的流程,重新排序、增加或删除步骤。例如,本发申
请中记载的各步骤可以并行地执行也可以顺序地执行也可以不同的次序执行,只要能够实现本申请公开的技术方案所期望的结果,本文在此不进行限制。
上述具体实施方式,并不构成对本申请保护范围的限制。本领域技术人员应该明白的是,根据设计要求和其他因素,可以进行各种修改、组合、子组合和替代。任何在本申请开的精神和原则之内所作的修改、等同替换和改进等,均应包含在本申请保护范围之内。
Claims (13)
- 一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法,其特征在于,该方法包括:获取训练样本,其中,所述训练样本包括光伏板的正常样本图像和缺陷样本图像;将所述训练样本输入至待训练的用于生成光伏板缺陷图像的循环生成对抗网络CycleGAN模型中,获取由所述CycleGAN模型输出的光伏板的候选图像,其中,所述候选图像包括光伏板的正常候选图像以及光伏板的缺陷候选图像;根据所述训练样本和所述候选图像对所述CycleGAN模型进行训练,以获取用于生成光伏板缺陷图像的目标CycleGAN模型。
- 根据权利要求1所述的训练方法,其特征在于,所述将所述训练样本输入至待训练的用于生成光伏板缺陷图像的所述CycleGAN模型中,获取由所述CycleGAN模型输出的光伏板的候选图像,还包括:根据所述正常样本图像生成对应的第一缺陷候选图像和第三正常候选图像,并根据所述第一缺陷候选图像生成对应的第一正常候选图像;根据所述缺陷样本图像生成对应的第二正常候选图像和第三缺陷候选图像,并根据所述第二正常候选图像生成对应的第二缺陷候选图像。
- 根据权利要求2所述的训练方法,其特征在于,所述CycleGAN模型包括第一生成器和第二生成器,所述方法还包括:将所述正常样本图像输入所述第一生成器,由所述第一生成器根据所述正常样本图像输出所述第一缺陷候选图像,并将所述第一缺陷候选图像输入所述第二生成器,由所述第二生成器根据所述第一缺陷候选图像输出所述第一正常候选图像,并将所述正常样本图像输入所述第二生成器,由所述第二生成器根据所述正常样本图像输出所述第三正常候选图像;将所述缺陷样本图像输入所述第二生成器,由所述第二生成器根据所述缺陷样本图像输出所述第二正常候选图像,将所述第二正常候选图像输入所述第一生成器,由所述第一生成器根据所述第二正常候选图像输出所述第二缺陷候选图像,并将所述缺陷样本图像输入所述第一生成器,由所述第一生成器根据所述缺陷样本图像输出所述第三缺陷候选图像。
- 根据权利要求3所述的训练方法,其特征在于,所述CycleGAN模型包括第一判别器和第二判别器,所述方法还包括:根据所述第一判别器,对所述光伏板的正常样本图像和所述光伏板的正常候选图像进行判别;根据所述第二判别器,对所述光伏板的缺陷样本图像和所述光伏板的缺陷候选图像进行判别。
- 根据权利要求4所述的训练方法,其特征在于,所述根据所述训练样本和所述候选图像对所述CycleGAN模型进行训练,以获取用于生成光伏板缺陷图像的目标CycleGAN模型,还包括:基于所述第一判别器和所述第二判别器的判别结果,获取所述CycleGAN模型的总损失函数;基于所述总损失函数调整所述CycleGAN模型的模型参数,并对调整后的CycleGAN模型继续下一次训练,直至满足训练结束条件,得到所述目标CycleGAN模型。
- 根据权利要求5所述的训练方法,其特征在于,所述基于所述第一判别器和所述第二判别器的判别结果,获取所述CycleGAN模型的总损失函数,包括:根据所述正常样本图像和所述第二正常候选图像以及根据所述缺陷样本图像和所述第一缺陷候选图像的判别结果,获取所述CycleGAN模型的第一损失函数;根据所述正常样本图像和所述第一正常候选图像的相似度以及根据所述缺陷样本图像和所述第二缺陷候选图像的相似度,获取所述CycleGAN模型的第二损失函数;根据所述正常样本图像和所述第三正常候选图像的相似度以及根据所述缺陷样本图像和所述第三缺陷候选图像的相似度,获取所述CycleGAN模型的第三损失函数;根据所述第一损失函数、所述第二损失函数和所述第三损失函数,获取所述CycleGAN模型的所述总损失函数。
- 根据权利要求6所述的训练方法,其特征在于,所述方法还包括:获取目标优化函数;根据所述目标优化函数和所述总损失函数,调整所述CycleGAN模型的模型参数。
- 一种基于CycleGAN模型的光伏板缺陷图像生成方法,包括:获取待处理的光伏板图像;将所述待处理的光伏板图像输入至用于生成光伏板缺陷图像的目标CycleGAN模型中,以得到光伏板缺陷图像。
- 一种用于生成光伏板缺陷图像的CycleGAN模型的训练装置,其特征在于,所述装置包括:第一获取模块,用于获取训练样本,其中,所述训练样本包括光伏板的正常样本图像和缺陷样本图像;输出模块,用于将所述训练样本输入至待训练的用于生成光伏板缺陷图像的循环生成对抗网络CycleGAN模型中,获取由所述CycleGAN模型输出的光伏板的候选图像,其中,所述候选图像包括光伏板的正常候选图像以及光伏板的缺陷候选图像;训练模块,用于根据所述训练样本和所述候选图像对CycleGAN模型进行训练,以获取用于生成光伏板缺陷图像的目标CycleGAN模型。
- 一种基于CycleGAN模型的光伏板缺陷图像生成装置,包括:获取模块,获取待处理的光伏板图像;输入模块,将所述待处理的光伏板缺陷图像输入至用于生成光伏板缺陷图像的目标CycleGAN模型中,以得到光伏板缺陷图像。
- 一种电子设备,其特征在于,包括处理器和存储器;其中,所述处理器通过读取所述存储器中存储的可执行程序代码来运行与所述可执行程序代码对应的程序,以用于实现如权利要求1-7或权利要求8中任一项所述的方法。
- 一种计算机可读存储介质,其上存储有计算机程序,其特征在于,该程序被处理器执行时实现如权利要求1-7或权利要求8中任一项所述的方法。
- 一种计算机程序产品,包括计算机程序,所述计算机程序在被处理器执行时实现根据权利要求1-7或权利要求8中任一项所述的方法。
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210790763.5 | 2022-07-06 | ||
CN202210790763.5A CN115147377A (zh) | 2022-07-06 | 2022-07-06 | 一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法及装置 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2024007602A1 true WO2024007602A1 (zh) | 2024-01-11 |
Family
ID=83412835
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2023/079045 WO2024007602A1 (zh) | 2022-07-06 | 2023-03-01 | 一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法及装置 |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN115147377A (zh) |
WO (1) | WO2024007602A1 (zh) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117934470A (zh) * | 2024-03-22 | 2024-04-26 | 宁德时代新能源科技股份有限公司 | 模型训练方法、缺陷检测方法、装置、设备和存储介质 |
CN118505687A (zh) * | 2024-07-17 | 2024-08-16 | 合肥中科类脑智能技术有限公司 | 光伏板缺陷检测方法、存储介质和电子设备 |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115147377A (zh) * | 2022-07-06 | 2022-10-04 | 中国华能集团清洁能源技术研究院有限公司 | 一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法及装置 |
CN115311526B (zh) * | 2022-10-11 | 2023-04-07 | 江苏智云天工科技有限公司 | 基于改进的Cycle GAN网络的缺陷样本生成方法和系统 |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190333219A1 (en) * | 2018-04-26 | 2019-10-31 | Elekta, Inc. | Cone-beam ct image enhancement using generative adversarial networks |
CN112598611A (zh) * | 2020-12-16 | 2021-04-02 | 阳光保险集团股份有限公司 | 一种凸字体银行卡号图像的合成方法、识别方法及装置 |
CN112668623A (zh) * | 2020-12-22 | 2021-04-16 | 中国铁道科学研究院集团有限公司 | 基于生成对抗网络的双耳销钉缺陷样本的生成方法及装置 |
CN112906769A (zh) * | 2021-02-04 | 2021-06-04 | 国网河南省电力公司电力科学研究院 | 一种基于CycleGAN的输变电设备图像缺陷样本增广方法 |
CN112950561A (zh) * | 2021-02-22 | 2021-06-11 | 中国地质大学(武汉) | 光纤端面缺陷检测方法、设备及存储介质 |
JP2021168078A (ja) * | 2020-04-13 | 2021-10-21 | 株式会社小糸製作所 | 画像生成システム、画像生成システムの機械学習方法 |
CN113554630A (zh) * | 2021-07-30 | 2021-10-26 | 西安电子科技大学 | 芯片表面缺陷检测方法、系统、计算机设备和存储介质 |
CN113570603A (zh) * | 2021-09-26 | 2021-10-29 | 国网浙江省电力有限公司电力科学研究院 | 一种基于CycleGAN的接触网零部件缺陷数据生成方法及系统 |
CN115147377A (zh) * | 2022-07-06 | 2022-10-04 | 中国华能集团清洁能源技术研究院有限公司 | 一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法及装置 |
-
2022
- 2022-07-06 CN CN202210790763.5A patent/CN115147377A/zh active Pending
-
2023
- 2023-03-01 WO PCT/CN2023/079045 patent/WO2024007602A1/zh unknown
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190333219A1 (en) * | 2018-04-26 | 2019-10-31 | Elekta, Inc. | Cone-beam ct image enhancement using generative adversarial networks |
JP2021168078A (ja) * | 2020-04-13 | 2021-10-21 | 株式会社小糸製作所 | 画像生成システム、画像生成システムの機械学習方法 |
CN112598611A (zh) * | 2020-12-16 | 2021-04-02 | 阳光保险集团股份有限公司 | 一种凸字体银行卡号图像的合成方法、识别方法及装置 |
CN112668623A (zh) * | 2020-12-22 | 2021-04-16 | 中国铁道科学研究院集团有限公司 | 基于生成对抗网络的双耳销钉缺陷样本的生成方法及装置 |
CN112906769A (zh) * | 2021-02-04 | 2021-06-04 | 国网河南省电力公司电力科学研究院 | 一种基于CycleGAN的输变电设备图像缺陷样本增广方法 |
CN112950561A (zh) * | 2021-02-22 | 2021-06-11 | 中国地质大学(武汉) | 光纤端面缺陷检测方法、设备及存储介质 |
CN113554630A (zh) * | 2021-07-30 | 2021-10-26 | 西安电子科技大学 | 芯片表面缺陷检测方法、系统、计算机设备和存储介质 |
CN113570603A (zh) * | 2021-09-26 | 2021-10-29 | 国网浙江省电力有限公司电力科学研究院 | 一种基于CycleGAN的接触网零部件缺陷数据生成方法及系统 |
CN115147377A (zh) * | 2022-07-06 | 2022-10-04 | 中国华能集团清洁能源技术研究院有限公司 | 一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法及装置 |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117934470A (zh) * | 2024-03-22 | 2024-04-26 | 宁德时代新能源科技股份有限公司 | 模型训练方法、缺陷检测方法、装置、设备和存储介质 |
CN118505687A (zh) * | 2024-07-17 | 2024-08-16 | 合肥中科类脑智能技术有限公司 | 光伏板缺陷检测方法、存储介质和电子设备 |
Also Published As
Publication number | Publication date |
---|---|
CN115147377A (zh) | 2022-10-04 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2024007602A1 (zh) | 一种用于生成光伏板缺陷图像的CycleGAN模型的训练方法及装置 | |
JP6596164B2 (ja) | 単一ビューオブジェクト再構成のためのファイングレインデータセットにおける教師なしマッチング | |
WO2019179295A1 (zh) | 面部识别的方法和装置 | |
CN103984951A (zh) | 一种磁粉检测缺陷自动识别方法及系统 | |
KR20240050461A (ko) | 배터리 전극 시트 절연 코팅층 결함의 검출 방법, 장치 및 컴퓨터 장치 | |
CN103400139A (zh) | 一种混凝土裂缝特征信息的识别方法 | |
RU2013134132A (ru) | Способ и устройство для определения эталонного набора данных класса для классификации ценных документов | |
CN116453029B (zh) | 基于图像数据的楼宇火灾环境检测方法 | |
CN102567529B (zh) | 一种基于双视图主动学习技术的跨语言文本分类方法 | |
CN112115873A (zh) | 基于深度学习的硅藻自动检测方法及系统 | |
CN114037666B (zh) | 一种数据集扩充和阴影图像分类辅助的阴影检测方法 | |
CN117849193A (zh) | 钕铁硼烧结的裂纹损伤在线监测方法 | |
WO2024061141A1 (zh) | 共性知识约束的遥感样本迁移方法 | |
JP2019087050A (ja) | 構造物維持管理業務支援システム | |
CN111539395A (zh) | 一种基于光流法和支持向量机的挖掘机实时检测方法 | |
CN117152094A (zh) | 基于计算机视觉的钢板表面缺陷分析方法、装置及系统 | |
CN116894005A (zh) | 文件处理方法、装置、电子设备和存储介质 | |
CN116794553A (zh) | 用于高频电源的智能故障诊断方法及其系统 | |
Myrans et al. | Combining classifiers to detect faults in wastewater networks | |
Gao et al. | Early stage damage detection of wind turbine blades based on UAV images and deep learning | |
Li et al. | Pid-yolox: An x-ray prohibited items detector based on yolox | |
CN114813730B (zh) | 基于边缘计算的工业零器件的检测方法和装置 | |
Wenying et al. | A novel template weighted match degree algorithm for optical character recognition | |
Fuqiang et al. | Virus-evolutionary genetic algorithm based selective ensemble for steganalysis | |
Xia et al. | Detection of insulator defects based on improved YOLOv7 model |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 23834386 Country of ref document: EP Kind code of ref document: A1 |