基于双通道多尺度注意力机制的光伏板裂缝检测方法  

Photovoltaic Panel Crack Detection Method Based on Dual-Channel Multi-Scale Attention Mechanism

作　　者：强浩[1,2] 叶波 唐文祺 QIANG Hao;YE Bo;TANG Wenqi(Intelligent Manufacturing Branch,School of Mechanical Rail Transit,Changzhou University,Changzhou213164,China;Energy and Power Advanced Equipment Engineerng Research Center,Changzhou 213164,China)

机构地区：[1]常州大学机械与轨道交通学院,智能制造产业学院,江苏常州213164 [2]江苏省能源动力高端装备工程研究中心,江苏常州213164

出　　处：《计算机测量与控制》2023年第12期84-89,264,共7页Computer Measurement &Control

基　　金：江苏省研究生实践创新计划项目(SJCX21_1272)。

摘　　要：针对目前传统边缘检测方法提取出的图像边缘轮廓模糊、不连续等问题,提出一种基于双通道多尺度注意力机制的光伏板裂缝检测方法,实现对图像低级边缘、边界、目标轮廓的检测;首先构建了双通道主干网络,包含语义分支通道和空间细节分支通道;其次,基于多尺度原则构建了多尺度及注意力机制模块,对特征图像的高、宽、通道的维度变换,分配特征权重,在捕捉跨通道信息的同时,还能够捕捉方向感知和位置感知的信息;最后将空洞融合模块融合到语义分支通道中,提升网络提取特征信息的能力。实验结果表明,所提出的算法对光伏板图像边缘检测性能有提升,相较HED、RCF与FCN算法,F_(1)值提升了2.83%、0.37%与1.54%,获得了较为清晰的裂缝图像。Aiming at the problems of fuzzy and discontinuous image edge contours extracted by traditional edge detection methods,a photovoltaic panel crack detection method based on dual-channel multi-scale attention mechanism is proposed to detect the detections of low-level edges,boundaries,and target contours.Firstly,the dual-channel backbone network was constructed,including the se-mantic branch channel and spatial detail branch channel.Secondly,based on the multi-scale principle,the multi-scale and attention mechanism was built to transform the dimensions of the height,width,and channel for the feature image,allocate the feature weights,capture the cross-channel information,and capture the direction and position information.Finally,the hole fusion module was integrated into the semantic branch channel,enhancing the ability of network to extract the feature information.Experimental re-sults show that the proposed algorithm improves the edge detection performance of photovoltaic panel images.Compared to the holis-tically-nested edge detection(HED),richer convolutional features(RCF)and Fully convolutional network(FCN)algorithms,the F,value of the proposed algorithm was increased by 2.83%,0.37%and 1.54%,respectively,which obtained the clearer crack images.

关 键 词：裂缝检测 多尺度 注意力机制 双通道网络 空洞融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TM615[自动化与计算机技术—计算机科学与技术]