基于YOLOv5s改进的机油瓶表面缺陷检测方法

Improving the Defect Detection Method of YOLOv5s for Oil Bottle Surfaces

摘要:

针对机油瓶表面瑕疵点目标过小、ROI候选框鲁棒性较差导致液位线定位不准等问题，提出一种改进YOLOv5s的缺陷检测算法。利用K-Means++代替K-Means进行聚类中心的初始化，使得生成的先验框更加接近检测目标的真实形状和大小；同时在主干网络中引入可变形卷积，提高特征提取的灵活性，并引入SE注意力机制，对特征图不同通道进行权重调整；此外，在颈部网络中使用BiFPN代替原有的PANet，实现对不同尺度信息的自适应特征融合。实验结果表明，改进的YOLOv5s算法mAP达到了96.9%,较YOLOv5s算法提升了6%，准确率提升了4%。经实验验证了改进后的YOLOv5s在检测准确率方面优于原始的YOLOv5s算法，解决了小目标漏检及液位线定位存在偏差的问题。

Abstract:

To address the issues of small target size of surface defects in engine oil bottles and the poor robustness of ROI candidate boxes leading to inaccurate liquid level positioning, an improved defect detection algorithm for YOLOv5s is proposed. Firstly, K-Means++ is used to initialize the clustering centers instead of K-Means, making the generated prior boxes closer to the real shape and size of the detection targets. At the same time, deformable convolution is introduced into the backbone network to improve the flexibility of feature extraction, and the SE attention mechanism is introduced to adjust the weights of different channels in the feature maps. Additionally, BiFPN is used in the neck network instead of the original PANet to achieve adaptive feature fusion of different scale information. Experimental results show that the improved YOLOv5s algorithm achieves an mAP of 96.9%, which is 6% higher than that of the YOLOv5s algorithm, and the accuracy is increased by 4%. The experimental results verify that the improved YOLOv5s algorithm outperforms the original YOLOv5s algorithm in terms of detection accuracy, solving the problem of missing small targets and inaccurate liquid level positioning.