基于通道变换轻量化Yolov5s交通标志识别算法

Lightweight Yolov5s Traffic Sign Recognition Algorithm Based on Channel Transformation

为满足车载终端设备对于交通标志识别模型占用内存的需求,本文针对计算资源受限的嵌入式平台,提出一种轻量交通标志识别算法Yolov5s-lite。通过在Yolov5s中引入Fire Module结构进行通道变换、降低残差模块数量,减少了其模型占用内存。在TT100K数据集上实验结果表明,Yolov5s-lite相比于Yolov5s,模型参数量下降22.8%、计算量下降27.9%、实际模型内存下降21.7%、mAP仅下降0.5%,在检测准确率相当的前提下有效压缩了模型大小。

In order to meet the memory requirements of the vehicle terminal equipment for the traffic sign recognition model, this paper proposes a lightweight traffic sign recognition algorithm Yolov5s-lite for the embedded platform with limited computing resources. By introducing Fire Module structure in Yolov5s to transform channels and reduce the number of residual modules, the memory occupied by the model is reduced. The experimental results on the TT100K data set show that compared with Yolov5s, Yolov5s-lite reduces the number of model parameters by 22.8%, the amount of computation by 27.9%, the actual model memory by 21.7%, and the mAP by only 0.5%. On the premise of the same detection accuracy, this effectively compresses the model size.