改进YOLOv3的无纺布表面缺陷检测研究

Study on Surface Defect Detection of Non-Woven Fabric by Improved YOLOv3

针对在无纺布表面缺陷检测中存在小缺陷目标漏检及识别率不高的问题,提出了一种改进YOLOv3的无纺布表面缺陷检测算法。首先在网络模型当中生成一个新的特征图层,来提取更多小缺陷目标的特征,其次引入K-Means++算法对聚类先验框中心点的提取进行改进,选取更为合适的Anchor Box,使定位更加精准,提高检测精度。最后,在自制的无纺布表面缺陷数据集上进行对比检测,研究结果表明:改进后的YOLOv3算法在测试集上的mAP值为85.83%,比原始的YOLOv3算法提高了6.99%,单张图片的平均检测时间为0.168s,与原始算法检测时间基本持平,检测性能也优于Faster R-CNN。

In allusion to the phenomenon of minor defects fail to detect and low recognition rate in non-woven fabric surface defect detection,an improved YOLOv3 non-woven fabric surface defect detection method is raised.Firstly,a new feature layer is generated in the network model to extract more features of the target with small defects Secondly,Use K-Means ++ method to optimize the selection of the center point of the clustering prior box.A more suitable Anchor Box makes the positioning more accurate and improves the detection accuracy.Finally,a comparative test was conducted on the self-made non-woven fabric surface defect data set.The results prove that the mAP value of the modified method on the test set is 85.83%,which is 6.99% higher than the initial YOLOv3 method.The average detection time of a single image is 0.168s,which is basically the same as that of the initial method,and the detection capability exceeds Faster R-CNN.