关键词: 缺陷检测, YOLOv8 算法, 可变形卷积, 注意力机制, 损失函数

Abstract:

At the traditional image processing algorithms for the detection of steel surface defects, there are problems such as low recognition efficiency and a high false detection rate of leakage. The YOLOv8-DSG (Deformable Convolution Network Squeeze and Excitation Network Generalized Intersection over Union) steel surface defect detection algorithm is proposed. Based on the traditional YOLOv8 algorithm, several improvements are made. Firstly, the DCN ( Deformable Convolutional Network) is embedded in the C2f ( Convolution to Feature) module of the Backbone network, which enhances the feature extraction ability of the model under

complex background conditions. Secondly, the SE ( Squeeze and Excitation network ) attention module is introduced into the Neck network, which highlights the important feature information of the steel surface and enhances the richness of the feature fusion. Lastly, the GIOU ( Generalized Intersection Over Union) loss function is used instead of the original CIOU(Complete Intersection Over Union). Compared with CIOU, GIOU introduces the minimum enclosing frame area ratio, which can more accurately measure the overlapping area of the frames. The experimental results show that the YOLOv8-DSG algorithm achieves an average accuracy mAP of

80% on the NEU-DET dataset, which is 3. 3% higher compared to the original YOLOv8 algorithm. And it has a low rate of misdetection and omission, demonstrating higher detection accuracy and arithmetic efficiency. This algorithm can play an important role in quality inspection.

Key words: defect detection, YOLOv8, deformable convolution, attention mechanism, loss function