基于FS-YOLOv3及多尺度特征融合的棉布瑕疵检测

Cotton defect detection based on FS-YOLOv3 and multiscale feature fusion

为解决传统棉布生产工艺中瑕疵检测成本高、精度低、速度慢等问题,提出一种FS-YOLOv3(Four Scales YOLOv3)网络来自动检测棉布瑕疵.该网络结合K-Means++聚类算法,以交并比为距离度量获取较好尺寸的锚框,提高检测速度.设计了4个不同尺度的卷积特征图与深度残差网络中相应尺度的特征图进行融合,有效地学习样本特征.将Softer NMS算法作为预测框过滤机制,使得高分类置信度的边框位置更为准确.实验结果表明:使用FS-YOLOv3网络能有效提高低对比度、小尺度目标的棉布瑕疵检测精度,整体性能优于传统的检测方法.

To solve the problems of high cost,low accuracy and slow speed in defect detection in traditional cotton production processes,FS-YOLOv3(Four Scales YOLOv3)is proposed to detect cotton fabric defects automatically.It combines with K-Means++clustering algorithm and takes IOU as the distance measurement to get the anchor frame with better size to improve the detection speed.Four convolutional feature maps of different sacles are designed,which are fused with the feature maps of corresponding scales in the deep residual network to learn the features of the sample effectively.In addition,the Softer NMS algorithm is used as the prediction frame filtering mechanism,which makes the position of the border frame with the high classification confidence more accurate.Experimental results show that using the FS-YOLOv3 network can improve the detection accuracy of cotton fabric defects with low-contrast and small scale targets effectively,and the overall performance is better than traditional detection methods.