基于改进YOLOX的钢材表面缺陷检测研究

Steel surface defect detection algorithm based on improved YOLOX

针对目前单阶段目标检测网络YOLOX的特征提取能力不足、特征融合不充分以及钢材表面缺陷检测精度不高等问题,提出一种改进YOLOX的钢材表面缺陷检测算法。首先,在Backbone部分引入改进的SE注意力机制,增添一条最大池化层分支,进行权重融合,强化重要的特征通道;其次,在Neck部分引入ASFF模块,充分利用不同尺度的特征,更好地进行特征融合;最后,针对数据集所呈现的特点,将IOU损失函数替换为EIOU损失函数,改善模型定位不准确的问题,提高缺陷检测精度。实验结果表明,改进的YOLOX算法具有良好的检测效果,在NEU⁃DET数据集上的mAP达到了75.66%,相比原始YOLOX算法提高了3.74%,较YOLOv6提升了2.76%,检测精度优于其他主流算法。

In view of the unsatisfied feature extraction capability,inadequate feature fusion,and low accuracy in steel surface defect detection in the current single⁃stage object detection network YOLOX,a steel surface defect detection algorithm based on improved YOLOX is proposed.An improved SE attention mechanism is introduced into the Backbone,adding a pooling layer branch to fuse the weight and strengthen important feature channels.An ASFF(adaptively spatial feature fusion)module is incorporated in the Neck to fully utilize features of different scales and achieve better feature fusion.On the basis of the characteristics of this dataset,the IOU loss function is replaced by EIOU loss function,so as to eliminate inaccurate model positioning and improve the accuracy of defect detection.Experimental results demonstrate that the improved algorithm has good detection performance,which achieves mAP(mean average precision)of 75.66%on the NEU⁃DET dataset,increasing 3.74%in comparison with that of the original YOLOX algorithm,and 2.76%over that of the YOLOv6 algorithm.Therefore,the detection accuracy of the proposed algorithm outperforms that of the other mainstream algorithms.