基于改进YOLOv5算法的钢管焊缝缺陷检测

A Method for Detecting Defects in Steel Pipe Welds with Improved YOLOv5

针对X光图像下钢管焊缝缺陷对比度弱﹑缺陷尺寸大小不一﹑同类缺陷形状变化大等因素导致的缺陷检测率不高的问题,提出一种改进YOLOv5的焊缝缺陷检测算法。首先,对X光图像进行去模糊处理,获得较清晰的焊缝图像;其次,在YOLOv5的主干网络中引入动态区域感知卷积代替标准卷积,保证参数不增加的情况下,增强特征提取能力;进一步针对YOLOv5中CSP特征金字塔融合准则过于简单的问题,采用了一种高效的特征融合机制以增强特征表达能力;最后,在检测头部分引入可学习权重参数,实现检测头中的特征自适应融合。实验结果表明,与传统YOLOv5算法相比,虽然检测速度从32.2 fps降到27.5 fps,但是检测的mAP提高了3.3%,达到94.6%,初步满足实际生产中钢管焊缝缺陷自动检测需求。

Aiming at the problem of low defect detection rate caused by factors such as weak contrast of steel pipe weld defects,different defect sizes,and large changes in the shape of similar defects under X-ray images,an improved YOLOv5 weld defect detection algorithm is proposed.First,deblur the X image to obtain a clearer weld image;then introduce dynamic region-aware convolution to replace the standard convolution in the backbone network of YOLOv5 to enhance feature extraction capabilities without increasing parameters;Second,since the original CSP-FPN in YOLOv5 adopts a simple criterion for fusing multiscale features which may lead to the degeneration of semantic information,this paper introduces an efficient feature aggregation mechanism to promote the multiscale feature presentation ability;Finally,the learnable weight parameters are introduced in the detection head part to realize the adaptive fusion of features in the detection head. The experimental results show that compared with the traditional YOLOv5 algorithm,the detection speed is reduced from 32.2 fps to 27.5 fps,and the detected mAP is increased by 3.3%to 94.6%,which preliminarily meets the needs of automatic detection of steel pipe weld defects in actual production.