基于改进的YOLOv4-tiny钢卷端面缺陷检测

Defect detection of steel coil based on improved Yolov4-Tiny

针对带钢生产过程中钢卷端面出现的缺损和拉丝缺陷,本文提出了一种改进的YOLOv4-tiny检测方法。首先,在主干网络中加入了注意模块,增强检测模型对缺陷特征的聚焦能力,增加一个上采样层,优化了深度特征和浅层特征的特征融合;其次,使用Focal损失函数替换置信度和分类的二分交叉熵损失函数,解决分类过程中存在正负类样本分布不均衡问题;最后,利用加权K-means聚类算法重新聚类得到新的先验框。实验结果表明,本文改进后的模型参数量和检测速度与原模型相当,但检测精确度上取得了更好的效果,更适用带钢生产的实时检测任务。

For strip steel edge of production process of edge defect and drawing defects,this paper puts forward an improved YOLOv4-tiny detection method.Firstly,an attention module is added to the backbone network to enhance the ability of the detection model to focus on defect features.An upper sampling layer is added to optimize the feature fusion of depth features and shallow features.Then,the Focal loss function was used to replace the confidence and dichotomous cross entropy loss function of classification to solve the problem of unbalanced distribution of positive and negative samples in the classification process.Finally,the weighted K-means clustering algorithm is used to get a new prior box.The experimental results show that the number of parameters and detection speed of the improved model are similar to the original model,but the detection accuracy is better,and it is more suitable for the real-time detection task of strip production.