【摘要】 生活水平的提升直接导致生活垃圾体量的增加，垃圾分类对于可回收资源的再利用及环境污染问题的治理具有重要意义。针对垃圾种类繁多、体积大小不一、形态各异的特点，提出一种基于YOLOv5s的垃圾识别及分类方法，并从注意力机制、网络结构及损失函数三个维度对算法进行改进。针对YOLOv5s模型精度较低的问题，提出在网络中添加SE注意力机制；为提升模型对不同尺度垃圾的识别效果，将Neck网络中引入高效的BiFPN结构；为提升模型对遮挡情况下的垃圾目标识别效果，提升模型的训练速度，使用SIOU＿Loss作为边界框损失函数。并通过自制的包含2000张、2291个目标框的数据集，对改进模型的性能进行验证。实验结果表明，改进的YOLOv5s网络模型对垃圾的识别精度提高了7.71个百分点，召回率提高了9.69个百分点，研究结果可为垃圾分类机器人识别的相关研究提供参考。更多还原

【Abstract】 The improvement in people’s living standards directly leads to an increase in the volume of household garbage, making the classification of garbage crucial for the reuse of recyclable resources and the management of environmental pollution. This paper proposes a garbage recognition and classification method based on YOLOv5s, and addresses the diverse characteristics of garbage, including its various types, sizes and shapes. The algorithm is enhanced in three key areas: attention mechanism, network structure and loss function. SE attention mechanism is added to the network to address the issue of low accuracy in the YOLOv5s model, while an efficient BiFPN structure is also brought in to improve the model’s recognition capabilities for various scales of garbage.Additionally, SIOU＿Loss is used as the boundary box loss function to enhance the model’s ability to recognize garbage targets in case of occlusion and to accelerate the training speed. The performance of the improved model is verified by a self-made data set containing 2,000 images and 2,291 target boxes. The experimental results indicate that the improved YOLOv5s network model increases recognition precision of garbage by 7.71% and the recall rate by 9.69%, which can provide reference for the related research on the recognition of garbage classification by robot..更多还原