Research On Traffic Sign Detection Based On Improved Lightweight Model

In recent years,unmanned systems have become a research area of great interest.Among them,fast and accurate detection of traffic signs is a key research component in unmanned systems.Current algorithms based on deep learning can effectively solve the problem of target detection class,but the smaller the target size and high resolution in the images of traffic sign dataset,the longer the time consumed for inference,which does not meet the requirement of real-time for autonomous driving in practical situations.Moreover,the limited resources of the in-vehicle hardware platform do not work well for image algorithms that require larger computation,and the accuracy of image recognition is not guaranteed.In order to enable the target detection algorithm to operate in scenarios such as embedded and mobile devices with weak computing power,this paper focuses on the following three aspects.1.improving the CycleGAN data enhancement algorithm.In this paper,by introducing the attention mechanism ACmix into the CycleGAN generator,the convolution is enabled to select features purposefully so that effective features can be reused.Meanwhile,the bulldozer distance is used to improve the stability of GAN training,solve the problems of distribution without overlap and zero derivative of JS dispersion,and improve the generation of samples2.lightweight target detection algorithm C₅-M-YOLOX.this paper optimizes the backbone network of YOLOX,introduces the Mobilenet V3 backbone network based on the YOLOX network structure,and optimizes the linear fitting of the activation function.A new mobile network attention mechanism Coordinate Attention is introduced,in which the channel attention contains location information.3.Knowledge distillation algorithm based on channel space attention.In this paper,we use an adaptive module to improve the model accuracy through a channel space attention mechanism to accommodate different layers of features with different feature map sizes and number of feature channels.Experiments prove that the algorithm can effectively improve the distillation effect,so that the recognition accuracy of the lightweight model can be effectively improved.