基于改进YOLOv5s的轻量级光伏板缺陷检测网络  

An improved YOLOv5s-based lightweight PV panel defect detection network

作　　者：陈子璇 陈辉[1] CHEN Zixuan;CHEN Hui(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《山东理工大学学报(自然科学版)》2025年第3期70-78,共9页Journal of Shandong University of Technology:Natural Science Edition

基　　金：国家自然科学基金项目(61170060)。

摘　　要：针对已有光伏板缺陷检测算法模型参数多、识别速度与精度难以取得较好平衡的问题,以YOLOv5s模型为基础,提出用于光伏板表面缺陷检测的轻量级网络LPV-YOLO。为降低模型的参数量,利用Ghost模块和Mish激活函数提出GhostMConv和C3MGhost模块,将其用于YOLOv5s模型中,实现骨干网络轻量化;提出一种融合SimAM注意力机制和空间金字塔池化层的注意力金字塔池化,在不增加参数和计算量的前提下,弥补骨干网络轻量化带来的精度丢失;在颈部网络中嵌入SE通道注意力模块,以提升模型对缺陷关键信息的捕捉能力。实验表明:与原始的YOLOv5s相比,改进后LPV-YOLO网络的参数量降低49%,体积减少46%,计算量减少50%,而mAP只下降0.6%,在可接受范围之内;在光伏板缺陷数据集上的mAP仍然能够达到93.8%,帧率可以达到70.42 FPS,满足实时性的要求;与YOLOv7、SSD300、RetinaNet等端到端网络相比,LPV-YOLO检测精度最优,模型参数量也最小。该模型在保持较高检测精度的同时,大幅降低了模型复杂度,可为采用资源有限的移动端设备进行光伏板缺陷检测提供一种有效的方法。To address the challenge of achieving an optimal balance between the speed and accuracy in the detection of photovoltaic(PV)panel defect,this study proposes a lightweight network,LPV-YOLO,based on the YOLOv5s model for surface defect detection.Specifically,we first employ the Ghost module and the Mish activation function to design GhostMConv and C3MGhost modules,which are used in the YOLOv5s model to realize the lightweight of the backbone network,thereby reducing the number of parameters of the model.Second,an attention pyramid pooling module combining the SimAM attention and the spatial pyramid pooling layer is proposed.This module compensates for the accuracy loss caused by the lightweight of the backbone network without increasing the number of parameters and calculations.Finally,the channel attention squeeze and excitation(SE)module is embedded in the neck section to improve the ability of the model to extract critical defect information.Extensive experiments demonstrate that the LPV-YOLO model reduces the number of parameters by 49%,the model size by 46%and the computation cost by 50%,compared with the original YOLOv5s.Despite these reductions,the mean average precision(mAP)of LPV-YOLO remains high at 93.8%,with only a marginal decrease of 0.6%,which is within an acceptable range.Moreover,the recognition speed of the model can reach 70.42 FPS on the PV panel defect dataset,which satisfies real-time requirements.Compared with other end-to-end networks such as YOLOv7,SSD300,RetinaNet,LPV-YOLO maintains high detection accuracy while greatly reducing the complexity of the model,making it suitable for mobile devices with limited resources.

关 键 词：光伏板缺陷检测 实时检测 YOLOv5s 空间金字塔池化 轻量级注意力 

分 类 号：TB532.1[理学—物理] TB553[理学—声学]