CN-122023410-A - Insulator defect detection method based on super-resolution reconstruction and deep learning fusion

CN122023410ACN 122023410 ACN122023410 ACN 122023410ACN-122023410-A

Abstract

The invention discloses an insulator defect detection method based on super-resolution reconstruction and deep learning fusion, and belongs to the technical field of computer vision and electric power inspection. The method comprises the steps of obtaining an original aerial image, inputting a pre-trained super-resolution reconstruction model for quality enhancement, outputting the super-resolution reconstruction image, and inputting the reconstruction image into an improved FS-YOLOv8 detection model for reasoning to obtain a defect detection result. The dynamic snake convolution module based on the micro-arc priori is integrated in the main network of the improved model, and initial sampling points of the dynamic snake convolution module are distributed along a preset arc and are used for enhancing the geometric modeling capability of the slender arc morphological characteristics of the insulator. Optionally, the method further comprises a result verification step, wherein the real defects and the shadow artifacts are distinguished through multi-factor weighted judgment of direction consistency, texture contrast and chromaticity consistency. The invention solves the contradiction between the quality degradation of the aerial image and the difficulty in detecting the small defects, and remarkably improves the detection precision and the robustness.

Inventors

ZHANG JINGYI
Huang Lianshu
Dong Pengrui
YANG XIAOXIA
ZHANG JIANYONG
HUANG YU
WU TONG
YANG YUQI
CHEN LIFAN

Assignees

成都理工大学

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (10)

1. The insulator defect detection method based on super-resolution reconstruction and deep learning fusion is characterized by comprising the following steps of: S1, acquiring an original aerial image containing an insulator; s2, inputting an original aerial image into a pre-trained super-resolution reconstruction model, carrying out image quality enhancement, and outputting a super-resolution reconstruction image, wherein the image quality enhancement at least comprises the restoration of motion blur, compression artifacts and random noise; And S3, inputting the super-resolution reconstructed image into an improved FS-YOLOv detection model for reasoning to obtain a detection result of the insulator defect, wherein a dynamic snake-shaped convolution module based on a micro-arc priori is integrated in a feature extraction main network of the improved FS-YOLOv detection model, and initial sampling points of the dynamic snake-shaped convolution module based on the micro-arc priori are distributed along a preset arc for enhancing the geometric modeling capability of the long and thin arc morphological feature of the insulator.
2. The insulator defect detection method based on the fusion of super-resolution reconstruction and deep learning according to claim 1, wherein the super-resolution reconstruction model generates an impedance network and comprises a generator and a discriminator; The generator comprises a shallow feature extraction network, a dense feature extraction network and an up-sampling network which are sequentially arranged; the system comprises a shallow feature extraction network, an up-sampling network, a high-resolution reconstruction image processing module and a low-resolution reconstruction image processing module, wherein the shallow feature extraction network is realized by adopting a convolution module and is used for inputting a low-resolution image; The discriminator comprises a downsampling module and two parallel U-NET networks, wherein the downsampling module is connected with one U-NET network and used for inputting a reconstructed image output by a generator, the other U-NET network is used for inputting a real high-resolution image, the two U-NET networks respectively output pixel-by-pixel true and false discrimination results of the reconstructed image and the real image, a loss function is calculated based on the difference of the two discrimination results, and the loss function is fed back to a dense feature extraction network of the generator to update network parameters.
3. The method for detecting the insulator defect based on the super-resolution reconstruction and the deep learning fusion according to claim 2, wherein the generator maps the pixel values of the input image to [ -0.5,0.5] intervals by adopting a normalization strategy during training, and the loss function at least comprises anti-loss and perception loss, wherein: The countermeasures are calculated based on the difference of the discrimination results of the two U-NET networks, and are used for guiding the generator to learn and generate a reconstructed image which cannot be distinguished from a real image on the pixel level; the perception loss is calculated based on depth feature differences extracted by a pre-trained classification network and is used for guiding a generator to learn and keep consistency of the reconstructed image and the real image on high-level semantic features.
4. The method for detecting the insulator defect based on the super-resolution reconstruction and the deep learning fusion according to claim 1, wherein the dynamic serpentine convolution module based on the micro-arc prior performs the following steps: S31, setting an arc priori, wherein an arc angle theta and the number N of sampling points are preset, the N sampling points are distributed on the arc at equal intervals, the arc is symmetrical about the convolution center, and the angle of the ith sampling point is equal to the angle of the ith sampling point Determined by the following formula: wherein i=1, 2, N; s32, generating arc sampling point coordinates, namely taking the anchor point position of the current convolution operation as a convolution center Taking a preset sampling radius r as the distance from a sampling point to a convolution center to generate initial sampling point coordinates ; S33, predicting the original offset delta P from the input feature map F by 1X 1 convolution, and multiplying the original offset delta P by a constraint coefficient alpha after constraint by a tanh function to obtain the learning offset of each sampling point : ; Step S34, applying a dynamic offset mechanism, namely adding the corresponding learning offset to the initial sampling point coordinate to obtain a final sampling point coordinate ; Step S35, bilinear interpolation sampling, namely obtaining the characteristic value of each final sampling point position through bilinear interpolation ; Step S36, weighting and summing the output, namely performing equal weight average on the characteristic values of the N sampling points to obtain output characteristics 。
5. The method for detecting the defects of the insulators based on the super-resolution reconstruction and deep learning fusion as claimed in claim 1, wherein the improved FS-YOLOv8 detection model specifically comprises: The device comprises a main network, a main C2f module and a dynamic snake-shaped convolution module, wherein the main network is constructed based on CSPDARKNET architecture and comprises a plurality of cascaded convolution layers, a main C2f module and the dynamic snake-shaped convolution module based on a micro-arc priori, wherein the dynamic snake-shaped convolution module based on the micro-arc priori is respectively arranged in a feature extraction layer with the downsampling multiplying power of 8 times, 16 times and 32 times and is used for extracting arc morphological features of an insulator on multiple scales; A feature fusion neck adopts a bidirectional feature pyramid network to receive and weight bidirectional fusion of P3, P4 and P5 multi-scale feature images output by the main network, wherein P3 is a feature image which is downsampled by 8 times and has the size of 160 multiplied by 160, P4 is a feature image which is downsampled by 16 times and has the size of 80 multiplied by 80, and P5 is a feature image which is downsampled by 32 times and has the size of 40 multiplied by 40; The decoupling detection head comprises an insulator detection branch and a defect detection branch, wherein the insulator detection branch predicts based on an 80×80 scale and 40×40 scale feature map output by a neck network and is trained by adopting WIoU loss functions for detecting insulator body and medium scale defects, and the defect detection branch predicts based on a 160×160 scale feature map output by the neck network and does not comprise a 20×20 scale detection head for detecting micro defects with pixel sizes below 8×8.
6. The insulator defect detection method based on super-resolution reconstruction and deep learning fusion according to claim 5, wherein the feature fusion neck is of a weighted bidirectional feature pyramid network structure, and the weighted bidirectional fusion of the multi-scale features is realized through a top-down up-sampling path and a bottom-up down-sampling path; The neck network receives the P3, P4 and P5 multi-scale feature graphs output by the backbone network, performs information interaction through the weighted feature fusion nodes, performs feature enhancement through the cascaded neck C2f modules respectively, and outputs the feature enhancement to the decoupling detection head.
7. The method for detecting the defects of the insulator based on the fusion of super-resolution reconstruction and deep learning according to claim 1, further comprising the following result checking step after the step S3: s4, acquiring illumination information, unmanned aerial vehicle attitude information and geometric information of an onboard light supplementing lamp when an original aerial image is shot, and calculating a theoretical shadow direction vector of a comprehensive light source on the surface of an insulator, wherein the comprehensive light source comprises a natural light source and the onboard light supplementing lamp; S5, extracting geometric extension direction vectors of each suspected defect feature in the detection result obtained in the step S3, and calculating cosine values of included angles between the geometric extension direction vectors and the theoretical shadow direction vectors Cosine value The calculation formula of (2) is as follows: wherein Is a vector of the geometric extension direction, Is a theoretical shadow direction vector; step S6, extracting a local texture contrast factor TCF and a local chromaticity consistency factor CCF of the suspected defect area, and constructing a weighted scoring model: Wherein sigma is a Sigmoid function, w 1 、w 2 、w 3 is a preset weight, and different weight values are respectively given to the direction consistency, the texture contrast and the chromaticity consistency; Step S7, if And if the suspected defect feature is not less than the preset judgment threshold, judging the suspected defect feature as a shadow artifact and inhibiting the shadow artifact, otherwise, keeping the suspected defect feature as a real defect.
8. The method for detecting an insulator defect based on super-resolution reconstruction and deep learning fusion as claimed in claim 7, wherein the local texture contrast factor TCF is determined by extracting gray level co-occurrence matrix contrast of the candidate region in a plurality of directions as local contrast Extracting gray level co-occurrence matrix contrast of a set annular region at the periphery of a candidate frame as background contrast The local texture contrast factor TCF is calculated as follows: The local texture contrast factor TCF is used for representing whether the candidate region has microscopic texture characteristics of real defects; the local chromaticity consistency factor CCF is determined by converting a candidate region image into an HSV color space, calculating two-dimensional joint histogram distributions P and Q of pixels in the candidate region and pixels in the background region on hue components and saturation components, setting a histogram to have M bins, and adopting the distribution difference of KL divergence measurement and the like: the local chrominance consistency factor CCF is used to characterize whether the candidate region has a chrominance variation characteristic of a real defect.
9. The insulator defect detection method based on super-resolution reconstruction and deep learning fusion according to claim 1, wherein the improved FS-YOLOv8 detection model adopts INT8 quantization in reasoning, and model parameters participate in reasoning operation in the form of 8-bit integers.
10. Insulator defect detection system based on super-resolution rebuilding and deep learning fusion, which is characterized by comprising: The image acquisition module is used for acquiring an original aerial image through the unmanned aerial vehicle and transmitting the original aerial image to the ground processing terminal in real time through a wireless network; An image enhancement module integrated with the super-resolution reconstruction model according to claim 2 or 3, for performing quality enhancement and super-resolution reconstruction on the original aerial image, and outputting a super-resolution reconstructed image; a defect detection module integrated with the improved FS-YOLOv detection model according to any one of claims 1 to 9, for performing insulator defect identification on the super-resolution reconstructed image, and outputting a defect detection result; a result checking module, configured to implement the result checking step according to claim 7 or 8, and perform light and shadow artifact checking on the defect detection result; And the result output module is used for generating and outputting a detection report containing defect types, positions and confidence degrees and synchronizing the detection report to the power operation and maintenance system.

Description

Insulator defect detection method based on super-resolution reconstruction and deep learning fusion Technical Field The invention relates to the technical field of computer vision and intelligent inspection of power equipment, in particular to an insulator defect detection method based on super-resolution reconstruction and deep learning fusion. Background The insulator is a key component for ensuring the safe operation of the power grid in the power transmission line, and the timely detection of defects (such as cracks and breakage) is very important. At present, an automatic inspection method based on unmanned aerial vehicle aerial photography and deep learning target detection models (such as YOLO series) has become a mainstream research direction. However, the prior art has a remarkable bottleneck in coping with a real inspection scene, namely, firstly, motion blur is generated by shaking of an unmanned aerial vehicle in flight, block artifacts (JPEG blocking effect) are introduced by image transmission compression, random noise is brought to a complex field environment, and the degradation factors seriously mask the visual characteristics of tiny defects (such as cracks which only occupy 8×8 pixels) on the surface of an insulator. Second, the mainstream inspection model typically employs 640×640 low resolution input to pursue speed, and feature information that causes micro defects is severely diluted in the input stage. In addition, the insulator is an elongated, arc-shaped and shape-changeable target, the aerial background (sky, mountain forest and the like) is complex, and the positioning accuracy and the anti-interference capability of the traditional detection model on the target are insufficient. Existing improvements focus mostly on optimizing inside a single detection model, such as introducing attention mechanisms or improving feature pyramids. However, such schemes do not reach the root of the problem-the low input image quality. The image with serious degradation is directly detected, and the upper performance limit of the detection model is severely restricted like finding tiny dust through a blurred lens. Therefore, an integrated scheme is needed that can improve the image quality from the source and pertinently optimize the detection model to cooperatively solve the problems of image degradation and tiny target detection. Disclosure of Invention The invention aims to solve the problems that in the prior art, aerial photography of an unmanned aerial vehicle insulator image is difficult to detect micro defects due to quality degradation, and a detection model is poor in adaptability to an elongated target and is easy to be interfered by a complex background, and provides an insulator defect detection method based on super-resolution reconstruction and deep learning fusion, which can overcome the quality degradation problem of aerial photography of the unmanned aerial vehicle image and the characteristics that the insulator defect target is micro and the detection background is complex, so that the insulator defect automatic detection with high precision, high robustness and high efficiency is realized. The invention also discloses a system for realizing the insulator defect detection method based on the super-resolution reconstruction and deep learning fusion. The aim of the invention is mainly realized by the following technical scheme: the insulator defect detection method based on super-resolution reconstruction and deep learning fusion comprises the following steps: S1, acquiring an original aerial image containing an insulator; s2, inputting an original aerial image into a pre-trained super-resolution reconstruction model, carrying out image quality enhancement, and outputting a super-resolution reconstruction image, wherein the image quality enhancement at least comprises the restoration of motion blur, compression artifacts and random noise; And S3, inputting the super-resolution reconstructed image into an improved FS-YOLOv detection model for reasoning to obtain a detection result of the insulator defect, wherein a dynamic snake-shaped convolution module based on a micro-arc priori is integrated in a feature extraction main network of the improved FS-YOLOv detection model, and initial sampling points of the dynamic snake-shaped convolution module based on the micro-arc priori are distributed along a preset arc for enhancing the geometric modeling capability of the long and thin arc morphological feature of the insulator. The invention initiates a two-stage paradigm of 'first repair and then detection', explicitly repairs image degradation through a super-resolution reconstruction model, provides clear and high signal-to-noise ratio input for detection, radically improves a detection starting point, introduces dynamic snake-shaped convolution based on micro-arc priori in the detection model, leads convolution kernel initial sampling points to be distributed along an insulator standa