基于图像增强与改进YOLOv5s的斜纹针织物疵点检测算法  

Detection algorithm for twill knitted fabric defect based on image enhancement and improved YOLOv5s

作　　者：蒋宇亮 孙卫红 梁曼 邵铁锋[1] JIANG Yuliang;SUN Weihong;LIANG Man;SHAO Tiefeng(China Jiliang University,Hangzhou,310018,China)

机构地区：[1]中国计量大学,浙江杭州310018

出　　处：《棉纺织技术》2024年第7期40-46,共7页Cotton Textile Technology

基　　金：浙江省基础公益研究计划项目(LGG20E050014)。

摘　　要：针对疵点检测中图像光照不均、疵点尺度变化大等因素导致检测精度低的问题,提出一种基于图像增强与改进YOLOv5s的斜纹针织物疵点检测算法。首先,对采集图像进行预处理,利用改进的MSR算法获得去除环境光照射分量后的织物图像,然后通过Gabor滤波获取疵点图像ROI区域,并结合Mask掩膜及伽马变换以突出疵点特征。其次,对YOLOv5s网络进行改进,在主干网络中加入带SE注意力机制的RepVGG模块以提高主干网络的特征提取能力;在颈部网络中增加并行检测层以提升对多尺度目标的检测性能。最后,引入EIoU边框回归损失函数进一步提高模型性能。试验结果表明:相较于原YOLOv5s算法,改进后算法的mAP值达到89.7%,提升了7.4个百分点,模型推理速度达119.9帧/s,满足实际检测需求。Aimed at the problems of lower detection accuracy caused by uneven image illumination,different scale of defects and so on during defect detection,a twill fabric defect detection algorithm based on image enhancement and improved YOLOv5s was proposed.Firstly,preprocess was performed for acquired images.The improved MSR algorithm was used to obtain the fabric image after removing the reflection component.Then the ROI region of the defects image was obtained through Gabor filtering.Gamma transform and Mask were combined to highlight the defects features.Secondly,the YOLOv5s network was improved.RepVGG module with SE attention mechanism was added to the backbone network to improve the feature extraction capability of the backbone network.A parallel detection layer was added to the neck network to improve the detection performance of multi-scale targets.Finally,the EIoU border regression loss function was introduced to further improve the model performance.The experimental result showed that the mAP value of the improved algorithm was reached to 89.7% compared with original YOLOv5s algorithm,which was 7.4 percentages higher than that of the original.The model inference speed was 119.9 frames/s,which could meet the actual detection requirements.

关 键 词：疵点检测 MSR算法 YOLOv5 RepVGG 多尺度目标检测 

分 类 号：TS107[轻工技术与工程—纺织工程]