【摘要】 【目的】实现非接触式牛个体身份识别。【方法】提出了一种基于改进YOLOv5s的牛个体身份识别的轻量化模型YOLOv5-LW(Light Weight YOLOv5,YOLOv5-LW)，该模型采用轻量化主干网络ShuffleNetV2替换YOLOv5主干网络以减少原主干网络过多的参数量，采用SimAM注意力机制在不引入过多参数的条件下对目标通道的特征提取能力有提高效果，并采用CARAFE轻量化上采样算子在不引入过多参数的情况下提高采样核的感受野，开发了牛身份识别系统方便可视化检测。【结果】在测试集上，YOLOv5-LW算法目标检测的平均精度mAP为99.4%，参数量为4.22 M，计算量为9.2GFLOPs，模型占用的内存空间小于YOLOv5s原模型，仅为8.5 MB，模型大小降低38.9%，与Faster-rcnn,SSD,YOLOv3等目标检测模型相比在图片推理时间上分别减少28.7 ms、18.8 ms、40.3 ms，精度上分别增长了6.8%、8.9%、1.2%，具备显著的轻量化和准确率优势。与YOLOv5模型相比较参数量减少了40%，计算量减少了43%，满足轻量化的设计需要。【结论】改进后的YOLOv5-LW算法在保证了算法识别效果良好的情况下达到了轻量化设计的目标。更多还原

【Abstract】 【Objective】 In order to achieve non-contact identification of cattle.【Method】 A lightweight bovine identification model,YOLOV5-LW(Lightweight YOLOv5),based on an improved version of YOLOv5s,was proposed. This model uses the lightweight ShuffleNetV2 backbone network instead of the original YOLOv5 backbone network to reduce its excessive number of parameters. The SimAM attention mechanism improves the target channel’s feature extraction capability without introducing too many parameters. The CARAFE lightweight up-sampling operator improves the receptive field of the sampling kernel without introducing excessive parameters. A bovine identification system has been developed to facilitate visual detection.【Result】 In the test set,the YOLOv5-LW algorithm achieved an average precision(mAP) of 99. 4% for target detection,with 4. 22 M and a computation amount of 9. 2 GFLOPs. The model occupied less memory space than the original YOLOv5s model(only 8. 5 MB) and was 38. 9% smaller than FasterRCNN. Compared with SSD,YOLOv3 and other target detection models,image reasoning time was reduced by 28. 7 ms,18. 8 ms and 40. 3 ms respectively,while accuracy increased by 6. 8%,8. 9% and 1. 2% respectively. This gives it significant advantages in terms of both lightness and accuracy. The reduction of 43% meets the design needs of lightweight.【Conclusion】 The improved YOLOv5-LW algorithm achieves a lightweight design while ensuring an effective recognition rate.更多还原