基于改进YOLOv8n的钢材缺陷检测算法

Steel Defect Detection Algorithm Based on Improved YOLOv8n

摘要: 针对目前钢材表面缺陷检测方法中存在的算法复杂度高、模型计算量大等问题，本文选择YOLOv8n为基准模型，提出了一种基于YOLOv8n改进的跨尺度特征融合的轻量化的钢材表面缺陷检测方法，并命名为YOLOv8n-FMR。首先，新模型利用多尺度卷积模型MSC (Multi-Scale Conv)对模型进行改进，有效减少因堆叠不同尺寸的特征提取层而产生的精度损失问题；其次，为了降低融合MSC模块所带来的召回率和准确率减少的影响，引入了Focal Modulation焦点调制模块对Neck网络中的SPPF模块进行改进。最后，为了提高模型整体的检测性能，引入了应用于遥感领域的RFD模块，有效增加了模型对小物体的识别能力。实验结果表明，改进后的YOLOv8n-FMR模型在NEU-DET数据集上的mAP达到了72.8%，比原模型提高了2.7%，并且模型的参数量也比原模型下降了0.1 M。改进后的模型计算量仅为9.5 GFLOPs，提出的模型不仅实现了轻量化的设计，而且提升了检测效果。

Abstract: In response to the high algorithm complexity and large model computation of current steel surface defect detection methods, this paper selected YOLOv8n as the benchmark model and proposed a lightweight steel surface defect detection method based on YOLOv8n improved cross scale feature fusion, named YOLOv8n FMR. First, the new model utilized the multi-scale convolution model (MSC) to improve the model, effectively reducing the accuracy loss caused by stacking different size feature extraction layers. Second, to reduce the impact of the recall rate and accuracy reduction caused by the fusion of the MSC module, the Focal Modulation (FoM) module was introduced to improve the SPPF module in the Neck network. Finally, to improve the overall detection performance of the model, the RFD module, which is commonly used in remote sensing, was introduced to effectively increase the model’s ability to recognize small objects. The experimental results show that the mAP of the improved YOLOv8n-FMR model on the NEU-DET dataset reached 72.8%, an improvement of 2.7% over the original model, and the model’s parameter count was reduced by 0.1 M. At the same time, the model’s computation volume was only 9.5 GFLOPs. The proposed model not only achieves a lightweight design, but also improves detection performance.