Detection And Recognition Of Traffic Signs In Low Illumination Environment

With the rapid development of artificial intelligence,traffic sign recognition has been becoming an essential part of the intelligent transportation system.In recent years,although there have been a lot of research achievements on traffic sign detection and recognition,the existing intelligent traffic sign recognition system is mostly used in normal light.However,the lack of light due to weather or night driving makes it difficult for drivers or on-board cameras to effectively judge the information about traffic signs on the road ahead.Traffic sign recognition in low-light environments has gradually become important research.Firstly,the paper analyzes the characteristics of low illuminance images and the shortcomings of traditional image enhancement algorithms,and then studies the basic theory of Retinex low illuminance image enhancement methods and the imaging principle of color images,and analyzes and evaluates the principle and algorithm flow of the surrounding Retinex method and Retinex-Net convolution neural network.Considering the adverse effects such as noise and color distortion caused by Retinex-Net during the image enhancement,this paper put forward a modified Retinex-Net algorithm based on the Decom-Net and Enhance-Net structures of Retinex-Net.In the convolution layer,an efficient channel attention mechanism ECA modules avoiding dimension reduction are engaged to guide the network to reduce noise and improve the brightness of the image.The attention connection module DCA-Net is used to establish the jump connection between attention modules and improve the learning ability of the network.The denoising loss and color loss are introduced into the loss function to suppress noise and improve the robustness of image color restoration.The experimental results show that the image quality by the method in this paper is further improved,and the visual effect and objective evaluation indicators also achieve better results.Secondly,by analyzing the requirements of low illumination traffic sign recognition and the principle and structure of PP-YOLO network,this paper uses deformations convolution and depthwise separable convolution to balance the parameter quantity,uses data augmentation to improve the generalization ability of the network model,and removes optimization strategies that will bring extra delay to improve the network running speed.The experiment shows that the improved strategy in this paper greatly improves the running speed of the original network without much decrease in m AP value,and basically meets the real-time requirement of traffic sign recognition.Finally,the human-computer interaction interface of the system is designed by Py Qt software to achieve the requirements,which realizes the visualization of low illumination traffic sign enhancement and recognition process.The actual measurement shows that the platform can effectively demonstrate the enhancement effect of traffic sign images and intuitively output the recognition results.