基于YOLOv3的布匹瑕疵检测方法

Defect Detection Method for Textile Fabrics Based on YOLOv3

针对布匹瑕疵差异较大、分布不均匀等问题,在YOLOv3中引入SwinTransformerBlock模块,用自注意力机制专注于有效特征排除无效特征的干扰,解决瑕疵差异大、分布不均等问题.同时用可变形卷积v2替换普通卷积,增大网络的感受野和多尺度建模能力,更好地适应瑕疵的形状和位置变化,从而提高目标检测的准确性和鲁棒性.实验结果表明,改进后算法在mAP上比原算法提高了3.80%,在检测速度上下降了2.86帧每秒.

Regarding the problem of large differences in fabric defects and uneven distribution,this paper solves the problem of large differences in defects and uneven distribution by introducing the SwinTrans-formerBlock module in YOLOv3.The self-attention mechanism focuses on effective feature exclusion of inval-id feature interference.At the same time,deformable convolution is used to replace ordinary convolution to in-crease the network’s receptive field and multi-scale modeling ability so as to better adapt to the shape and po-sition changes of defects,thereby improving the accuracy and robustness of object detection.Experimental re-sults show that the improved algorithm has increased by 3.80%in mAP compared to the original algorithm,and decreased by 2.86 frames in detection speed.