Noise Removal in Traffic Sign Detection Systems

Abstract

The application of Traffic sign detection and recognition is growing in traffic assistant driving systems and automatic driving systems. It helps drivers and automatic driving systems to detect and recognize the traffic signs effectively. However, it is found that it may be difficult for these systems to work in challenging environments like rain, haze, hue, etc. To help the detection systems to have better performance in challenging conditions like rain and haze, we propose the use of a deep learning technique based on a Convolutional Neural Network to process visual data. The processed data could be used in the detection. We are using the NoiseNet model [11], a noise reduction network for our architecture. The model is trained to enhance images in patches instead of as a whole. The training is done using the Challenging Unreal and Real Environment - Traffic Sign Detection Dataset(CURE-TSD) which contains videos of different roads in various challenging situations. The enhanced images obtained are compared using the object detection algorithms YOLO and Faster RCNN. The Mean Absolute Error(MAE) of original and enhanced images are calculated and compared for two classes of images - rain and haze for both the algorithms. The proposed approach achieved an average Peak Signal to Noise Ration(PSNR) of 25.30 and an Structural Similarity(SSIM) of 0.88. The average MAE values of YOLO and Faster RCNN model reduced by 0.11 and 0.30 respectively on using enhanced images.