CN-122023220-A - Composite insulator infrared edge enhancement method and system based on generation countermeasure network

CN122023220ACN 122023220 ACN122023220 ACN 122023220ACN-122023220-A

Abstract

The invention discloses a composite insulator infrared edge enhancement method and a system based on a generated countermeasure network, which belong to the technical field of infrared image processing, wherein the method constructs the generated countermeasure network comprising an edge perception generator and a cross-branch interaction multi-scale discriminator; the edge perception generator is embedded with a self-adaptive noise suppression module and a weak edge enhancement attention module to realize noise suppression and edge enhancement, the cross-branch interaction multi-scale discriminator adopts an image and edge double-branch structure to synchronously evaluate image authenticity and edge accuracy, and a dynamic collaborative composite loss function fused with various losses is designed to be optimized by combining a self-adaptive training strategy. The invention can effectively inhibit the noise of the infrared image, obviously enhance the edge detail of the insulator and improve the accuracy and the robustness of defect detection.

Inventors

HU RUIZHE
LI TANGBING
HU AN
TU ZHAN
ZHOU YOUWU
XU JIAN
ZHOU LONGWU

Assignees

南昌科晨电力试验研究有限公司
国网江西省电力有限公司电力科学研究院

Dates

Publication Date: 20260512
Application Date: 20260403

Claims (10)

1. The composite insulator infrared edge enhancement method based on the generation countermeasure network is characterized by comprising the following steps of: Step S1, collecting infrared images of extra-high voltage composite insulators under different operation conditions as original images, marking each original image to obtain a standard edge image, adding Gaussian noise to obtain an infrared image of the extra-high voltage composite insulator containing noise, and dividing the infrared image into a training set and a testing set; Step S2, constructing and generating an countermeasure network model, wherein the countermeasure network model comprises an edge perception generator and a cross-branch interaction multi-scale discriminator, the edge perception generator is an improved U-Net network and sequentially comprises an encoding stage, a bottleneck stage, a decoding stage and an output stage, the encoding stage comprises an adaptive noise suppression module, the adaptive noise suppression is realized by means of weighted fusion of original features and smooth features, the bottleneck stage adopts dynamic cavity convolution, the cavity rate of the dynamic cavity convolution is dynamically adjusted according to the periodic probability of an insulator umbrella skirt in a feature map, and the decoding stage comprises an infrared weak edge enhancement attention module for generating an edge probability map and enhancing weak edge features according to the edge probability map; step S3, training the generated countermeasure network model by using a training set and combining a dynamic cooperative composite loss function and a self-adaptive training strategy, and performing model verification by using a test set, wherein the dynamic cooperative composite loss function at least comprises countermeasure loss, edge perception loss and edge-noise cooperative loss; and S4, inputting the ultra-high voltage composite insulator infrared image to be enhanced into a training-completed edge perception generator for generating an countermeasure network model, and outputting an edge enhanced image.
2. The composite insulator infrared edge enhancement method based on generation of countermeasure network according to claim 1, wherein in the modified U-Net network: The coding stage consists of four coding units, wherein each coding unit is formed by sequentially connecting a3 multiplied by 3 convolution layer, a BN layer, leakyReLU activation function and an adaptive noise suppression module, the coding units extract image multi-scale features through downsampling, and the adaptive noise suppression module realizes adaptive noise suppression through weighted fusion of original features and smooth features; The bottleneck stage adopts a combined structure of dynamic cavity convolution and 1 multiplied by 1 standard convolution, and the cavity rate is dynamically adjusted according to the periodic probability of the insulator umbrella skirt in the feature diagram; the decoding stage consists of four decoding units, and each decoding unit is formed by sequentially connecting a transposed convolution layer, a jump connection mechanism and an infrared weak edge enhancement attention module; And in the output stage, the characteristic image output by the infrared weak edge enhancement attention module is mapped into a single-channel image through a 1X 1 convolution layer, and a normalized edge enhancement image is output by adopting a Tanh activation function.
3. The composite insulator infrared edge enhancement method based on generation countermeasure network according to claim 2, wherein the adaptive noise suppression module performs the following adaptive noise suppression process: ; In the formula, For the feature map output by the adaptive noise suppression module, X is the pixel abscissa, y is the pixel ordinate, For the convolutional layer output features, i.e. the original features, For a3 x 3 mean filtering result, i.e. a smoothing feature, As an insulator edge protection factor, Is a noise suppression coefficient.
4. The method for enhancing the infrared edge of the composite insulator based on the generation countermeasure network according to claim 2, wherein the calculation method of the void ratio is as follows: ; In the formula, Is a pixel The void fraction at the location(s) is (are) the void fraction, Is a pixel The periodic probability of the insulator umbrella skirt, For the adjustment of the coefficients, e is a natural constant.
5. The method for enhancing the infrared edge of the composite insulator based on the generation countermeasure network according to claim 2, wherein the calculation process of the infrared weak edge enhancement attention module is as follows: ; In the formula, The output profile of the attention module is enhanced for infrared weak edges, For the weak edge enhancement factor to be used, For the input feature map of the current decoding unit at pixel (x, y), An edge probability map at pixel (x, y) is computed based on feature gradients for the infrared weak edge enhancement attention module interior.
6. The method for enhancing the infrared edge of a composite insulator based on a generated countermeasure network according to claim 1, wherein the cross-branch interaction multi-scale discriminator adopts a double-branch structure consisting of an image branch and an edge branch, wherein the image branch and the edge branch each comprise three scale branches, each scale branch comprises four 3×3 convolution layers, a cross-branch feature fusion module is arranged for realizing feature interaction of the image branch and the edge branch in an intermediate feature layer, and the cross-branch interaction multi-scale discriminator is input into an original image and an edge enhanced image Standard edge map And outputting a two-dimensional discrimination result of the authenticity score and the edge accuracy score.
7. The method of claim 1, wherein the dynamic cooperative composite loss function is formed by weighted combination of a countering loss, an edge perception loss, a structural similarity loss, a noise suppression loss and an edge-noise cooperative loss, and the weights are dynamically adjusted during training of the countering network.
8. The method for enhancing the infrared edge of a composite insulator based on a generation countermeasure network according to claim 7, wherein the countermeasure loss is improved by a wasperstein distance, as shown in the following formula: ; In the formula, In order to combat the loss of this, For the distribution of the training set data, Representing the expectation, a mean calculation representing the corresponding sample distribution, Representing the final output of the cross-branch interaction multi-scale arbiter, As the original image is to be taken, For the edge-enhanced image output by the edge-aware generator, Is a standard edge map.
9. The method for enhancing an infrared edge of a composite insulator based on a generation countermeasure network according to claim 7, wherein the edge-noise synergy loss Calculated by the following formula: ; in H, W, the standard edge maps are respectively Is defined by the height and width of the substrate, For a 3 x 3 local variance of the edge enhanced image at pixel (x, y), Is the mean local variance of the noiseless insulator infrared image, An edge probability map at pixel (x, y) is computed based on feature gradients for the infrared weak edge enhancement attention module interior.
10. A composite insulator infrared edge enhancement system based on a generation countermeasure network for executing the composite insulator infrared edge enhancement method based on a generation countermeasure network as claimed in any one of claims 1 to 9, comprising: The data acquisition module is used for acquiring infrared images of the extra-high voltage composite insulator under different operation conditions as original images, labeling each original image to obtain a standard edge image, adding Gaussian noise to the standard edge image to obtain an infrared image of the extra-high voltage composite insulator containing noise, and dividing the infrared image into a training set and a testing set; Generating an countermeasure network module, including generating a countermeasure network model, the generating the countermeasure network model including an edge awareness generator and a cross-branch interaction multi-scale discriminant; The model training module is connected with the data acquisition module and the countermeasure network generation module and is used for initializing the countermeasure network generation model, the training set is used for carrying out iterative training on the model by combining a dynamic cooperative composite loss function and a self-adaptive training strategy, in the training process, a dynamic cooperative composite loss function consisting of countermeasure loss, edge perception loss, structural similarity loss, noise suppression loss and edge-noise cooperative loss is calculated, model parameters are updated according to loss values until the model converges, and the model parameters after training is completed are saved; the image enhancement module is connected with the generation countermeasure network module and used for loading the model parameters after training, inputting the ultra-high voltage composite insulator infrared image to be enhanced into the edge perception generator for forward reasoning, and outputting an edge enhanced image through encoding, bottleneck, decoding and output stage processing.

Description

Composite insulator infrared edge enhancement method and system based on generation countermeasure network Technical Field The invention belongs to the technical field of infrared image processing, and particularly relates to a composite insulator infrared edge enhancement method and system based on a generated countermeasure network. Background The composite insulator is used as a core insulating component of the ultra-high voltage transmission line, and the running state of the composite insulator directly determines the safety of a power grid. The infrared thermal imaging technology is widely used for detecting defects of composite insulators due to the advantages of non-contact and strong real-time performance, and the edge definition of an infrared image of an insulator is a key premise of defect identification, and edge information bears core characteristics such as an umbrella skirt structure of the insulator, a defective area outline and the like. The extra-high voltage composite insulator can lead to small integral structure and fuzzy edge characteristics of the insulator in an infrared image even if the extra-high voltage composite insulator is subjected to sectional shooting due to the influence of the length of the extra-high voltage composite insulator. The existing infrared edge enhancement technology is mainly divided into a traditional method and a deep learning basic method, wherein the traditional method is used for extracting edges through manually designed gradient operators, such as Sobel operators, canny operators and Laplace enhancement. However, the infrared image of the extra-high voltage composite insulator has inherent defects of strong noise interference (complicated electromagnetic environment of a power transmission line), low contrast (small difference between the insulator and background infrared radiation), blurred edges (poor uniformity of infrared radiation of an organic material of the composite insulator), easy edge breakage, multiple false edges, detail loss and the like in the traditional method, and cannot meet the high-precision requirement of defect detection. The deep learning basic method, such as an edge enhancement model based on U-Net, can extract edges through semantic segmentation ideas, but is not optimized for the characteristics of low signal to noise ratio and strong noise coupling of infrared images, and the lack of contrast constraint leads to excessive or insufficient edge enhancement, so that the method is difficult to adapt to complex operation scenes of the extra-high voltage composite insulator. Therefore, an enhancement method which can adapt to the infrared image characteristics of the extra-high voltage composite insulator and has noise robustness and edge accuracy is needed. Disclosure of Invention Aiming at the problems of contradiction between noise suppression and edge reservation, poor adaptability to complex scenes and the like in the ultra-high voltage composite insulator infrared edge enhancement in the prior art, the invention provides a composite insulator infrared edge enhancement method and system based on an antagonism network, which realize the collaborative optimization of noise robustness and edge accuracy by generating an antagonism network model, innovative loss function design and a self-adaptive training strategy. The invention is realized by the following technical scheme: A composite insulator infrared edge enhancement method based on a generation countermeasure network comprises the following steps: Step S1, collecting infrared images of extra-high voltage composite insulators under different operation conditions as original images, marking each original image to obtain a standard edge image, adding Gaussian noise to obtain an infrared image of the extra-high voltage composite insulator containing noise, and dividing the infrared image into a training set and a testing set; Step S2, constructing and generating an countermeasure network model, wherein the countermeasure network model comprises an edge perception generator and a cross-branch interaction multi-scale discriminator, the edge perception generator is an improved U-Net network and sequentially comprises an encoding stage, a bottleneck stage, a decoding stage and an output stage, the encoding stage comprises an adaptive noise suppression module, the adaptive noise suppression is realized by means of weighted fusion of original features and smooth features, the bottleneck stage adopts dynamic cavity convolution, the cavity rate of the dynamic cavity convolution is dynamically adjusted according to the periodic probability of an insulator umbrella skirt in a feature map, and the decoding stage comprises an infrared weak edge enhancement attention module for generating an edge probability map and enhancing weak edge features according to the edge probability map; step S3, training the generated countermeasure network model by using a training set and combining a dynam