Re-YOLOv5:一种基于结构重参数化的钢材缺陷检测方法

钢材在生产的过程中很容易产生裂纹、斑点等缺陷，而目前对于所产生缺陷的检测技术还不是很成熟。为了实现对工业钢材生产过程中所产生的钢材缺陷进行实时鲁棒检测，以YOLOv5为基础，引入了结构重参数化方法，建立了Re-YOLOv5工业钢材缺陷检测模型。在该模型中，将YOLOv5的Neck层与Head层合并为Head层，用作预测，并且加入RepVGG模块和卷积层，输出预测结果。Backbone用作特征提取，可以在改善模型推理速度的同时提高检测准确率。同时，采用改进后的空间金字塔池化模块SPP^*对候选框进行分类和修正，以获取多尺度特征信息，并引入了有助于模型加深的CCBL模块。在公开的NEU-DET钢材缺陷图片数据集上进行测试，提出的模型的检测精度可达77.8%，与基线模型YOLOv5s相比，实现了6%的精度提升，且单幅图片的推理时间仅为8.9 ms，满足工业生产实时性需求。此外，该模型所占内存较小，便于部署到工业设备中。

Re-YOLOv5: a Method of Steel Defect Detection Based on Structural Re-Parameterization

Defects are easy to occur in the process of steel production,but the detection technology for the steel defects is not mature at present.To realize real-time robust detection of steel defects produced in the process of industrial production,a novel structural re-parameterization method based on YOLOv5 and construct the Re-YOLOv5 industrial steel defect detection model is proposed.The Neck layer and the Head layer of YOLOv5 are merged into the Head layer for prediction,while the RepVGG module and convolution are added to output the prediction results.Backbone is used for feature extraction,which can improve the reasoning speed and the detection accuracy at the same time.The candidate boxes are classified and modified by the improved spatial pyramid pooling scheme SPP^* to obtain multi-scale features,and the CCBL components are introduced to deepen the model structure.Extensive experiments prove that the proposed model can achieve superior results as compared to other methods tested on the public steel defect dataset.The detection accuracy can reach 77.8%.Especially,compared to the baseline model YOLOv5s,it achieves a 6% improvement.The reasoning time of a single image is only 8.9 ms,which meets the real-time requirements of industrial production.In addition,the model occupies less memory and is easy to be transplanted to industrial equipment.