CN-121982005-A - Power distribution network inspection image defect detection method and system based on collaborative feature learning

CN121982005ACN 121982005 ACN121982005 ACN 121982005ACN-121982005-A

Abstract

The invention belongs to the technical field of image defect detection, and relates to a method and a system for detecting defects of a power distribution network inspection image based on collaborative feature learning, wherein the method comprises the steps of carrying out super-resolution reconstruction on an input power distribution network inspection image to generate a high-resolution image; splitting a high-resolution image into two parallel branches, carrying out parallel processing, connecting along a channel dimension to obtain a first-stage basic feature, carrying out similar processing to obtain a second-stage basic feature and a third-stage basic feature, generating self-adaptive features based on channel standardized preprocessing, dynamic convolution mixing blocks and self-adaptive convolution, carrying out internal coarse-fine granularity information exchange to generate interactive features, generating a multi-scale fusion feature based on a cross-scale attention mechanism, carrying out defect detection on the multi-scale fusion feature through detection heads with different resolutions, and outputting the type, position and confidence information of the defect. The method can remarkably improve the precision and the robustness of the inspection image defect detection of the power distribution network through a multi-level and multi-scale characteristic enhancement and fusion mechanism.

Inventors

LIU QIHAO
WANG BO
YI JIFENG
Pang chaoyang
Lv Haokun
ZHANG HAO
LI LEI
LIU YONG
HU BINGXU
Sun shiyang
Bi Shuaipeng
WANG NINGXIN
GUO SHUAI
LI XUEGANG
SUN XIANG
MA MING
CHANG JIAN
HU BO
LU FEI
WU HAOYANG
WANG JINGCHAO
LIU CHUANHUI

Assignees

国网山东省电力公司桓台县供电公司

Dates

Publication Date: 20260505
Application Date: 20260128

Claims (10)

1. The detection method for the defects of the inspection images of the power distribution network based on collaborative feature learning is characterized by comprising the following steps: S1, extracting shallow layer characteristics F 0 from an input power distribution network inspection image through convolution, further processing F 0 by utilizing a residual characteristic block, fusing characteristics F RB processed by the residual characteristic block with F 0 , and performing channel expansion and pixel rearrangement operation on the fused characteristics F fusion to generate a high-resolution image I HR ; S2, splitting the I HR into two parallel branches X 1 and X 2 , dividing the X 1 into a plurality of non-overlapping sub-tensors, connecting the non-overlapping sub-tensors in series along a channel dimension to obtain a total tensor, convolving the total tensor to obtain an output tensor Y 1 of the X 1 , obtaining an output tensor Y 2 of the X 2 through maximum pooling and convolution operation, and connecting the Y 1 and the Y 2 along the channel dimension to obtain a first-stage basic feature F 1 ; S3, processing the F 1 again according to S2 to obtain a second-stage basic feature F 2 , and processing the F 2 again according to S2 to obtain a third-stage basic feature F 3 ; s4, for F 1 、F 2 、F 3 , performing multi-receptive field fusion based on channel standardization pretreatment, a dynamic convolution mixing block and self-adaptive convolution respectively to generate self-adaptive characteristics A 1 、A 2 、A 3 corresponding to F 1 、F 2 、F 3 , wherein the dynamic convolution mixing block comprises a Inception mixer and a convolution gating linear unit; S5, respectively carrying out internal coarse-fine granularity information exchange on the A 1 、A 2 、A 3 to generate interaction characteristics I 1 、I 2 、I 3 corresponding to the A 1 、A 2 、A 3 ; S6, generating a multi-scale fusion feature C 1 、C 2 、C 3 corresponding to the I 1 、I 2 、I 3 based on a cross-scale attention mechanism; and S7, setting three detection heads with different resolutions, respectively detecting the defects of the C 1 、C 2 、C 3 , and outputting the type, the position and the confidence information of the defects.
2. The method for detecting defects of inspection images of a power distribution network based on collaborative feature learning according to claim 1, wherein in S1, for the input inspection images I LR , the process of generating I HR is specifically as follows: f 0 was extracted by convolution, expressed as: ; In the formula, Representing a3 x 3 convolution; The F 0 is further processed by a single layer residual feature block, denoted as: ; In the formula, Representing a1 x 1 convolution; representing a ReLU activation function; Feature fusion of F RB with F 0 gave F fusion , expressed as: ; Processing F fusion through the channel expansion and pixel rearrangement operations yields I HR , denoted as: ; In the formula, Representing a pixel rearrangement operation.
3. The method for detecting the defects of the inspection image of the power distribution network based on collaborative feature learning according to claim 1, wherein in the step S2, the process of obtaining F 1 is specifically as follows: I HR is split into two parallel branches X 1 and X 2 by an average pooling and splitting operation, denoted as: ; In the formula, Representing a split operation; representing an average pooling operation; For X 1 , the input tensor is divided into four non-overlapping sub-tensors by splitting the spatial dimension, then the four sub-tensors are connected in series along the channel dimension to obtain a total tensor, and finally Y 1 is obtained by convolution, which is expressed as: ; In the formula, Representing a1 x 1 convolution; representing connections along the channel dimension; for X 2 , Y 2 is obtained by a max-pooling and convolution operation, expressed as: ; In the formula, Representing a maximum pooling operation; Y 1 and Y 2 were connected along the channel dimension to give F 1 .
4. The method for detecting the defects of the inspection images of the power distribution network based on collaborative feature learning according to claim 3, wherein the shape of the I HR is that Wherein b represents the batch size, c represents the number of input channels, w and h represent the spatial dimensions, and X 1 and X 2 are shaped as The shape of each sub-tensor when splitting X 1 is The shape of the total tensor is 。
5. The method for detecting the defects of the inspection images of the power distribution network based on collaborative feature learning according to claim 1, wherein in the step S4, the process of generating a 1 is specifically as follows: Pretreatment F 1 by channel normalization, denoted: ; In the formula, Representation SiLU activates a function; representing a batch normalization operation; F CBS represents the feature after the normalized pretreatment of the feature channel; F CBS is divided into a main path F main and a jump connection path F shortcut , expressed as: ; In the formula, Representing a split operation; F main is processed through three sequential dynamic convolution mixing blocks, denoted as: ; In the formula, Representing a dynamic convolution mixing block; 、、 respectively representing output results of the three dynamic convolution mixing blocks; For a pair of 、、 Performing adaptive convolution processing to obtain A 1 , which is expressed as: ; In the formula, Representing connections along the channel dimension; A 2 、A 3 is generated based on F 2 、F 3 in the same manner.
6. The method for detecting the defects of the inspection images of the power distribution network based on collaborative feature learning according to claim 5, wherein in the dynamic convolution mixing block, a dynamic Inception mixer is used Expressed as: ; ; In the formula, 、 Representing input and output characteristics of the dynamic Inception mixer, respectively; 、 Respectively represent splitting Two intermediate features obtained; And Depth separable convolutions with convolution kernels 3 and 5, respectively; convolution gating linear unit Expressed as: ; wherein W 1 、B 1 、W 2 、B 2 represents a learnable parameter; Representation GRLU activates a function; representing element-by-element multiplication; Representing a depth convolution; 、 respectively representing input and output characteristics of the convolution gating linear unit; The output characteristics of the dynamic convolution hybrid block are obtained and expressed as: ; In the formula, Representing a normalization operation; 、、 Representing the input, intermediate, output characteristics of the dynamic convolution mixing block, respectively.
7. The method for detecting the defects of the inspection images of the power distribution network based on collaborative feature learning according to claim 1, wherein in the step S5, the process of generating I 1 is specifically as follows: Processing a 1 to obtain coarse grain memory m c and fine grain memory m f , expressed as: ; In the formula, Representing a3 x 3 convolution; Representing a linear layer; generating a degree of interest weight for m c 、m f using a Softmax activation function, expressed as: ; ; In the formula, 、 Respectively representing the attention weights of m c 、m f ; representing a Softmax activation function; representing element-by-element multiplication; Will be 、 Addition gives I 1 , expressed as: ; The same applies to the generation of I 2 、I 3 based on A 2 、A 3 .
8. The method for detecting the defects of the inspection image of the power distribution network based on collaborative feature learning according to claim 1, wherein in the step S6, the process of generating C 1 is specifically as follows: query, key and value features of I 1 are extracted, expressed as: ; In the formula, Q s 、k s 、v s represents the query feature, key feature and value feature extracted when the expansion rate is s; Enhanced value feature with a computation scale s Expressed as: ; ; In the formula, W a represents an adaptive weight matrix; representing element-by-element multiplication; Representing a pooling operation; an attention weighting coefficient representing a scale s; A cross-scale attention feature is calculated, expressed as: ; in the formula, S represents the total number of scales involved in cross-scale attention fusion, namely the maximum value of S; the superscript T denotes the transpose, d k denotes the feature dimension of k s ; representing a cross-scale attention feature; In combination with the linear transformation, C 1 is obtained, expressed as: ; In the formula, Representing a linear layer; C 2 、C 3 is generated based on I 2 、I 3 .
9. The method for detecting the defects of the inspection image of the power distribution network based on collaborative feature learning according to claim 1, wherein in S7, the specific process of detecting the defects of C 1 、C 2 、C 3 is as follows: setting a high-resolution detection head, a medium-resolution detection head and a low-resolution detection head which are respectively used for carrying out parallel defect detection on C 1 、C 2 、C 3 , wherein the detection heads are expressed as follows: ; Wherein K represents the index of the detection head and is 1,2 or 3; C K represents C 1 、C 2 or C 3 ; representing the classification probability of the kth detection head; Frame regression representing the kth head; Indicating the target presence confidence of the kth detection head; Expressed as: ; In the formula, Representing the input as C K ; Representing a Sigmoid activation function; representing element-by-element multiplication; representing a batch normalization operation; Representing a3 x 3 convolution; a convolution weight matrix representing a classification branch of the kth detection head; Expressed as: ; In the formula, Representing the input as C K ; Representing a1 x 1 convolution; Expressed as: ; In the formula, Representing the input as C K ; A convolution weight matrix representing the target confidence branch of the kth detection head.
10. A power distribution network inspection image defect detection system based on collaborative feature learning, configured to implement the method of any one of claims 1-9, comprising: The super-resolution reconstruction module is used for extracting shallow layer features F 0 from an input power distribution network inspection image through convolution, further processing F 0 by utilizing a residual feature block, fusing the features F RB processed by the residual feature block with F 0 , and performing channel expansion and pixel rearrangement operation on the fused features F fusion to generate a high-resolution image I HR ; The downsampling compression module is used for splitting the I HR into two parallel branches X 1 and X 2 , dividing the X 1 into a plurality of non-overlapping sub-tensors, connecting the non-overlapping sub-tensors in series along a channel dimension to obtain a total tensor, convolving the total tensor to obtain an output tensor Y 1 of the X 1 , obtaining an output tensor Y 2 of the X 2 through maximum pooling and convolution operation, and connecting the Y 1 and Y 2 along the channel dimension to obtain a first-stage basic feature F 1 ; Processing F 1 again according to the downsampling compression module to obtain a second-stage basic feature F 2 , and processing F 2 again according to the downsampling compression module to obtain a third-stage basic feature F 3 ; The dynamic kernel self-adaptive convolution module performs multi-receptive field fusion on the basis of channel standardization pretreatment, a dynamic convolution mixing block and self-adaptive convolution for F 1 、F 2 、F 3 respectively to generate self-adaptive characteristics A 1 、A 2 、A 3 corresponding to F 1 、F 2 、F 3 , wherein the dynamic convolution mixing block comprises a Inception mixer and a convolution gating linear unit; The hierarchical granularity interaction module is used for respectively carrying out internal coarse-fine granularity information exchange on the A 1 、A 2 、A 3 to generate interaction characteristics I 1 、I 2 、I 3 corresponding to the A 1 、A 2 、A 3 ; The cross-resolution attention module is used for generating a multi-scale fusion feature C 1 、C 2 、C 3 corresponding to the I 1 、I 2 、I 3 based on a cross-scale attention mechanism; and the multi-scale parallel detection module is provided with three detection heads with different resolutions, and respectively detects the defects of the C 1 、C 2 、C 3 and outputs the type, the position and the confidence information of the defects.

Description

Power distribution network inspection image defect detection method and system based on collaborative feature learning Technical Field The invention belongs to the technical field of image defect detection, and particularly relates to a method and a system for detecting image defects of power distribution network inspection based on collaborative feature learning. Background The power distribution network is used as a key link of a power system, the safe and stable operation of the power distribution network directly determines the power supply reliability and the power utilization safety, and the inspection work is a core means for guaranteeing the normal operation of the power distribution network. The traditional manual inspection mode is limited by factors such as manpower and environment, the problems of low efficiency, high cost, strong subjectivity of detection results and the like generally exist, and especially, small defects such as fine damage of wires, slight aging of insulators and the like are difficult to accurately identify, so that the large-scale and fine inspection requirements of a modern power distribution network can not be met. With the wide application of unmanned aerial vehicle aerial photography, fixed camera monitoring and other technologies in power distribution network inspection, massive inspection images are rapidly acquired, an automatic defect detection technology based on computer vision becomes an industrial research hotspot by virtue of the high-efficiency and objective advantages, and various deep learning target detection algorithms such as YOLO, faster R-CNN and the like are gradually tried to be applied to the field. However, the particularity of the inspection scene of the power distribution network brings multiple serious challenges to the existing detection technology, so that the actual application effect is poor, on one hand, the inspection image is often influenced by the performance of shooting equipment, flying height, environmental shielding and the like, the resolution is limited, the small defect characteristic is extremely low in proportion, meanwhile, complex background interference such as trees, buildings and cable interweaving exists around the power distribution network equipment, the illumination condition is changed severely along with weather and time periods, the confusion of the defect characteristic and the background is further aggravated, the stability of characteristic extraction is seriously influenced, on the other hand, the existing general target detection algorithm is not fully adapted to the specific requirement of the inspection of the power distribution network, and lacks systematic design of low-resolution image enhancement, complex background suppression and multi-scale fusion characteristic adaptation, even if a part of improvement method specifically optimizes a single link (such as only improving small target detection capability or optimizing background segmentation), the difficulty of the multiple scene is still difficult to overcome, finally, the detection accuracy is insufficient, the error rate is higher, the model complexity is often higher, and the real-time processing performance on the edge computing equipment is difficult to match the field application requirement of the actual inspection. Disclosure of Invention According to the defects in the prior art, the invention aims to provide the method and the system for detecting the defects of the inspection image of the power distribution network based on collaborative feature learning, and the precision and the robustness of the detection of the defects of the inspection image of the power distribution network can be remarkably improved through a multi-level and multi-scale feature enhancement and fusion mechanism. In order to achieve the above purpose, the invention provides a power distribution network inspection image defect detection method based on collaborative feature learning, which comprises the following steps: S1, extracting shallow layer characteristics F 0 from an input power distribution network inspection image through convolution, further processing F 0 by utilizing a residual characteristic block, fusing characteristics F RB processed by the residual characteristic block with F 0, and performing channel expansion and pixel rearrangement operation on the fused characteristics F fusion to generate a high-resolution image I HR; S2, splitting the I HR into two parallel branches X 1 and X 2, dividing the X 1 into a plurality of non-overlapping sub-tensors, connecting the non-overlapping sub-tensors in series along a channel dimension to obtain a total tensor, convolving the total tensor to obtain an output tensor Y 1 of the X 1, obtaining an output tensor Y 2 of the X 2 through maximum pooling and convolution operation, and connecting the Y 1 and the Y 2 along the channel dimension to obtain a first-stage basic feature F 1; S3, processing the F 1 again accordin