基于YOLOv8-VCRA的一种轻量级的带钢表面缺陷检测研究

Research on a Lightweight Strip Surface Defect Detection Based on YOLOv8-VCRA

摘要: 钢表面缺陷检测的重点是快速识别和精确定位。目前在带钢表面缺陷检测领域，深度学习网络已经取得了显著的进步，但是通用算法复杂度高、计算量庞大，检测算法部署困难。本文基于YOLOv8提出一种轻量级的带钢表面缺陷检测模型YOLOv8-VCRA。该模型引入轻量级网络VanillaNet构建全新的骨干网络，通过减少不必要的分支结构降低模型的复杂度。此外，设计一个新模块CBAM-RFB，增强特征提取网络的感知场，有效改善目标易漏检的问题。最后将AFPN融入Neck结构，在融合过程中提取更多有用的信息。实验结果表明，改进后的YOLOv8-VCRA网络模型有着良好的检测性能，检测精度相比原始YOLOv8网络提升了1.2%，模型参数量降低了0.95 M，模型计算量降低了3.1 G，在保持原高精度少量提升的情况下，能够快速且实时地对钢材表面缺陷进行检测。

Abstract: The emphasis of steel surface defect detection is rapid identification and accurate location. At present, in the field of strip surface defect detection, deep learning networks have made remarkable progress, but the general algorithm is of high complexity and huge computation, and the detection algorithm deployment is difficult. In this paper, a lightweight strip surface defect detection model, YOLOV8-VCRA, is proposed based on YOLOv8. This model introduces lightweight network VanillaNet to build a new backbone network, and reduces the complexity of the model by reducing unnecessary branch structure. In addition, a new module CBAM-RFB is designed to enhance the perception field of the feature extraction network and effectively improve the problem that the target is easy to miss. Finally, AFPN is integrated into Neck structure to extract more useful information in the fusion process. The experimental results show that the improved YOLOV8-VCRA network model has good detection performance, the detection accuracy is increased by 1.2% compared with the original YOLOv8 network, the number of model parameters is reduced by 0.95 M, and the calculation amount of the model is reduced by 3.1 G. It can detect steel surface defects quickly and in real time.