Abstract:

To address the issues of low accuracy and efficiency in existing object detection methods for surface defect detection of strip steel， a GS-YOLO （Gather-and-distribute-Squeeze-YOLO） model was proposed for surface defect detection. Firstly， in the backbone network， SE （Squeeze-and-Excitation） attention mechanism was incorporated to enhance the model’s capability in recognizing and locating defect features significantly. Then， the traditional convolutions in the original C3 module were replaced with Ghost convolutions， thereby reducing the model’s parameters and computational cost effectively. Finally， the GD （Gather-and-Distribute） feature fusion module was introduced in the neck part of the model to replace the traditional Path Aggregation Network （PAN） and Feature Pyramid Network （FPN） architectures， so as to ensure the continuity of feature fusion and improve the recognition accuracy of objects with different scales. Experimental results demonstrate that， compared to the original YOLOv5s， the proposed model increases the precision， recall and mAP@0.5 by 1.32， 5.18 and 2.56 percentage points respectively， and reduces the computational cost by 0.4 GFLOPs. The above verifies that the proposed method increases the detection accuracy and decreases the computational cost of the model at the same time.

Key words: YOLOv5, strip steel surface defect, attention mechanism, feature fusion, lightweight structure, object detection

摘要：

为解决现有目标检测方法对带钢表面缺陷检测精度不高、效率低下的问题，提出一种GS-YOLO（Gather-and-distribute-Squeeze-YOLO）模型检测表面缺陷。首先，在骨干网络中，引入SE（Squeeze-and-Excitation）注意力机制，以显著增强模型对缺陷特征的识别与定位能力；然后，将原始C3模块中的传统卷积替换为Ghost卷积，从而有效降低模型的参数量与计算量；最后，在模型颈部引入GD（Gather-and-Distribute）特征融合模块取代传统路径聚合网络（PAN）和特征金字塔网络（FPN）架构，从而确保特征融合的连续性，并提高不同规模目标的识别准确度。实验结果表明，与原始的YOLOv5s相比，所提模型的精确率、召回率和mAP@0.5分别提升了1.32、5.18和2.56个百分点，而计算量减少了0.4 GFLOPs，充分表明所提方法在兼顾检测精度提高的同时，降低了模型的计算量。

关键词: YOLOv5, 带钢表面缺陷, 注意力机制, 特征融合, 轻量化结构, 目标检测