基于改进YOLOv5算法的带钢表面缺陷检测

Strip surface defect detection based on improved YOLOv5 algorithm

针对热轧带钢表面缺陷检测精度低的问题,提出一种改进的YOLOv5算法。首先,设计了一种多尺度融合平行骨干网络,其中,辅助骨干中的各个特征层经过转换,作为输入逐级复合连接到主骨干的各个特征层,利用辅助骨干提高主骨干的表达能力,从而提升了网络的表达能力;其次,在骨干网络和颈部中引入基于内容感知的特征重组模块(content-aware reassembly of features,CARAFE),获取检测任务所需的丰富语义信息,改善上采样语义信息丢失的问题;最后,在多尺度融合平行骨干中加入坐标注意力模块(coordinate attention,CA),增强网络的特征提取能力,更加准确定位目标位置。实验结果表明,改进后的YOLOv5算法在武钢热轧带钢表面缺陷数据集上平均准确率(mean average precision,mAP)达到93.1%,较原始YOLOv5算法提升3.9%,检测速度FPS保持在78.4,具有较高的检测精度。

To address the issue of low detection accuracy of surface defects in hot-rolled strip,an improved YOLOv5 algorithm was proposed.Firstly,a multi-scale fusion parallel backbone network was designed,in which each feature layer in the auxiliary backbone was connected to each feature layer of the main backbone as input.The expression capacity was improved by utilizing the auxiliary backbone.Secondly,a content-aware reassembly of features(CARAFE)module was introduced into the backbone network and neck to obtain rich semantic information required for detection tasks and mitigate the loss of up-sampled semantic information.Finally,a coordinate attention module(CA)was applied to the multi-scale fusion parallel backbone to enhance the feature extraction capacity and accurately locate the target position.Experimental results demonstrate that the improved YOLOv5 algorithm achieves a mean average precision(mAP)of 93.1%on the WISCO hot-rolled strip surface defect dataset,which is 3.9%higher than the original YOLOv5 algorithm,while maintaining a detection speed of 78.4 frame per second.The above indicators prove the effectiveness and practicability of the proposed algorithm.