【摘要】 针对现有钢材表面缺陷检测算法参数多、计算量大、精度低和不易部署到资源有限的嵌入式设备中等问题，基于YOLOv8n目标检测算法，提出一种轻量级钢材检测算法YOLO-LGM。通过设计轻量级网络LCNet-C重构YOLOv8n的主干特征提取网络，以降低网络的参数量和计算量；将YOLOv8n网络模型Neck层中的Conv模块替换为GSConv模块，在减轻模型的计算量同时提升模型的精度；最后，将多尺度注意力机制(EMA)融入到C2f模块中，构建C2f-EMA模块，将Neck层中所有C2f模块替换为融合注意力机制后的C2f-EMA模块，进一步提升模型精度。实验结果表明：在NEU-DET数据集上，YOLO-LGM的模型大小为3.5 MB,参数量为1 642 622,GFLOPs为5.0,均值平均精度为76.4%;与YOLOv8n相比，模型大小减少了43.5%,参数量减少了45.4%,GFLOPs减少了38.3%,均值平均精度提升了1.6%。改进后的算法在检测钢材缺陷方面的效果有所提升，且模型更加轻量，适合部署在嵌入式设备中。更多还原

【Abstract】 Aiming at the challenges faced by existing steel surface defect detection algorithms, such as excessive parameters, high computational complexity, low accuracy, and difficulty in deployment on resource-limited embedded devices, a lightweight steel detection algorithm YOLO-LGM was proposed based on the YOLOv8n object detection algorithm.The backbone feature extraction network of YOLOv8n was reconstructed by designing a lightweight network LCNet-C to reduce the parameters number and computation.The Neck layer Conv module of the YOLOv8n network model was replaced with GSConv module to reduce computation and improve accuracy.Efficient multi-scale attention(EMA) was integrated into the C2f module to construct the C2f-EMA module.By replacing all C2f modules in the neck layer with C2f-EMA modules after integrating attention mechanism, the model accuracy was further enhanced.Experimental results demonstrate that YOLO-LGM has a model size of 3.5 MB with 1 642 622 parameters and 5.0 GFLOPs while achieving a mean average accuracy of 76.4% on NEU-DET dataset.Comparing with YOLOv8n, the model size of the proposed method is reduced by 43.5%,the parameter number is reduced by 45.4%,the GFLOPs is reduced by 38.3%,and mean average accuracy is improved by 1.6%.The improved algorithm is effective in detecting steel defects, and the model is more lightweight and suitable for deployment in embedded devices.更多还原