UH-TSD: a lightweight model for traffic sign detection and recognition under hazy environment

Traffic sign detection and recognition, as an essential part of intelligent driving systems, has received increasing attention and application in recent years. However, in the real world, it is susceptible to extreme weather. Haze, as common weather, makes traffic signs blurred and distorted, leading to a significant decrease in the detection effect of the model. In this paper, based on the detection model YOLOv5s, we design a wavelet-based up-and-down sampling module to alleviate the information loss problem caused by traditional up-and-down sampling. Meanwhile, we propose a feature extraction module incorporating a multi-scale attention mechanism to enhance the feature extraction capability for hazy images. Moreover, we adopt the paradigm of joint training of the image dehaze network and detection network to optimize the detection performance in the hazy environment without increasing the parameters. Extensive experiments on the TT100K and CURE-TSD datasets demonstrate the validity of our proposed model.