基于低秩分解的YOLO轻量化目标检测模型

Lightweight YOLO models for object detection based on low-rank decomposition

随着列车智能化程度的不断提高,许多研究探索了车载设备目标检测模型的轻量化技术,以满足在资源有限情况下的高效计算。针对当前YOLO(You Only Look Once)系列目标检测模型轻量化方法通用性不够强的问题,文章提出了一种针对YOLO系列的低秩分解参数压缩算法。首先通过预设的低秩比例系数和卷积单元的输入/输出通道数量计算低秩,然后通过对目标结构的卷积层进行Tucker分解,得到新的卷积序列,最后融合新的卷积序列,取代原有卷积层。使用公开数据集对所提出的基于低秩分解的参数压缩方法进行试验,选用了YOLOv5-l、YOLOv8-x和YOLOX-x这3种模型,在保证低秩分解后的模型检测平均精度为原模型96%的前提下,模型参数量和浮点计算量均减少了约40%,同时图像检测速度能达到原模型的150%左右。此外,可视化结果显示,该方法压缩过的模型与原模型在相同图像上的关注区域基本相同。试验结果表明,文章提出的方法可以有效地对单阶段YOLO系列目标检测模型进行轻量化压缩,提高模型在车载设备上的可用性;同时,所做工作对轨道交通领域自动驾驶场景下的其他模型的轻量化处理也具有重要的借鉴意义。

As train intelligence continues to advance,numerous studies have emerged to explore lightweight techniques for object detection models on onboard equipment,for the purpose of improving calculation efficiency amidst limited resources.This paper proposed a parameter compression algorithm based on low-rank decomposition for the You Only Look Once(YOLO)series of object detection models,aiming to overcome the limited versatility of current lightweight treatment methods for these models.Initially,calculations were conducted to determine a low rank using the preset low-rank ratio coefficient along with the number of input and output channels of convolution units.Subsequently,a new convolution sequence was obtained by performing the Tucker decomposition to the convolutional layer of the target structure.Lastly,a new convolution sequence was fused to replace the original convolutional layer.Experiments were conducted on the proposed parameter compression method based on low-rank decomposition,using public datasets and three models,i.e.,YOLOv5-l,YOLOv8-x,and YOLOX-x.While ensuring an average detection accuracy of the models based on low-rank decomposition over 96%compared to the original models,the number of parameters and float-point calculations of these models reduced by about 40%.The experiments resulted in a nearly 50%improvement in image detection speed.Further visual displays illustrated similar receptive fields on the same images between the compressed and original models.The experimental results show that the proposed method is effective in lightweight compression of single-stage object detection models of the YOLO series and can enhance their usability for onboard equipment.In addition,this study serves as a valuable reference for the lightweight treatment of other models in the automatic train operation scenarios of rail transit field.