基于轻量化YOLOv7的织物疵点检测算法研究  

Fabric defect detection algorithm research based on lightweight YOLOv7

作　　者：赵英宝[1] 刘姝含 黄丽敏[1] 武晓晶[1] ZHAO Yingbao;LIU Shuhan;HUANG Limin;WU Xiaojing(Hebei University of Science and Technology,Shijiazhuang,050018,China)

机构地区：[1]河北科技大学,河北石家庄050018

出　　处：《棉纺织技术》2024年第11期53-61,共9页Cotton Textile Technology

基　　金：国家自然科学基金项目(62003129,61903122)。

摘　　要：由于检测工艺的不完善和外界因素的影响,织物疵点检测过程中会存在目标漏检和误检的情况,并且为了在移动设备和嵌入式设备中部署,提出了一种基于轻量化YOLOv7的织物疵点检测算法(LFD-YOLOv7)。首先,针对YOLOv7算法网络结构复杂和参数量较大的问题,结合GhostNet网络构建EGM模块来取代主干网络中的ELAN模块,降低了网络的复杂度和计算瓶颈,增强网络的学习能力;其次,基于ShuffleNetv2的思想,将其与残差网络相融合构造了S-SPPCSPC模块,使网络更加轻量化;然后,引入CA注意力机制来抑制背景噪声对目标检测的影响,提高小目标的准确率;最后采用SIoU损失函数来优化输出预测框边界,提高算法收敛速度。试验结果表明:与YOLOv7算法相比,LFD-YOLOv7算法平均检测精度提升了5.59个百分点,参数量减少了30.3%,检测速度达到41帧/s,满足纺织工业生产对织物疵点的准确性和实时性要求。Due to the imperfect detection process and the influence of external factors,there may be cases of missed and false detection of targets in the fabric defect detection process.In order to deploy it in mobile and embedded devices,a lightweight YOLOv7 based on fabric defect detection algorithm was proposed,i.e LFD-YOLOv7.Firstly,in response to the problem of complex network structure and large parameter quantity in YOLOv7 algorithm,EGM module was constructed by combining GhostNet network structure to replace ELAN module in the backbone network,which could reduce the complexity and computational bottlenecks of the network,and enhance the learning ability of the network.Secondly,based on the idea of ShuffleNetv2,S-SPPCSPC module was constructed by integrating it with residual networks,making the network more lightweight.Then,CA attention mechanism was introduced to suppress the influence of background noise on target detection and improve the accuracy to small targets.Finally,SIoU loss function was used to optimize the output prediction box boundary and improve the convergence speed of the algorithm.The experiment results showed that compared with YOLOv7 algorithm,the average detection accuracy of LFD-YOLOv7 algorithm was improved by 5.59 percentage points,the parameters were reduced by 30.3%,and the detection speed was reached 41 frame/s,the requirements of accuracy and real-time of fabric defects in textile industry was reached.

关 键 词：织物疵点 YOLOv7 注意力机制 残差网络 轻量化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]