基于YOLOv5的轻量级布匹疵点检测模型

A Light weight fabric defect detection model based on YOLOv5

针对人工检测布匹疵点效率低、漏检误检严重的现象,文中提出一种基于YOLOv5算法的布匹疵点检测模型G-YOLOv5。该模型首先利用Ghost卷积机制代替传统卷积,减少冗余的参数量和计算量;其次在骨干网络的最后加入协同注意力机制,加强对小目标物体的分类和定位性能;同时,使用轻量化的上采样算子CARAFE减少在特征处理过程中的特征损失。实验结果表明,改进后的算法在布匹疵点检测数据集上的平均准确率为88.4%,相比于YOLOv5算法提高了2.2个百分点,参数量缩减了一半,能够在较小的模型下达到较高的检测精度,满足实际工业的检测需求。

On the basis of the low efficiency of manual detection of cloth defects and false detection,a fabric defect detection model G-YOLOv5 based on YOLOv5 algorithm is proposed.Firstly,Ghost module is adopted to replace the traditional convolution to reduce the amount of redundant parameters and calculation.Secondly,Coordinate Attention is added at the end of the backbone,strengthening the classification and positioning performance of small target objects.Meanwhile,the lightweight up-sampling operator CARAFE is used to reduce the feature loss in the process of feature processing.The results show the mean average accuracy of the improved algorithm on the fabric defect detection data set is 88.4%,which is 2.2 percentage points higher than that of YOLOv5 algorithm.And the amount of parameters is reduced to half of YOLOv5,which can achieve high detection accuracy in a small model and meet the detection needs of contemporary industry.