Automatic sign recognition for low quality images using modified MIRNET

Abstract

Traffic sign board recognition helps self-driving cars and autonomous driving systems to make an appropriate decision on the movement of the vehicle automatically even at higher frame rates. However, traffic sign images captured at night or in bad weather conditions are either blurred, faded, damaged, and with low contrast which makes real-time traffic sign recognition tasks, especially on fine-grained traffic sign categories more challenging. In this thesis work, we try to address the issues of traffic sign images with low contrast, noise, and blurs using a two-step approach. The first step involves the enhancement of traffic sign images by applying a modified MIRNet model to produce enhanced traffic sign images. In the second step, the Yolov4 model is used to recognize the traffic signs in an unconstrained environment. We achieve better prediction results on the enhanced quality traffic sign images. We compare the class-wise results of traffic sign before and after the enhancement. We attain an increment of 5.40% Yolov4 mAP@0.5 on low quality images and achieve an overall mAP@0.5 of 96.75% on the GTSRB dataset. We achieve a mAP@0.5 of 100% on the GTSDB dataset for the broad categories which is comparable with the state-of-the-art work. Moreover, results shown in images with the predicted class and bounding box with their confidence score before and after the image enhancement establish the effectiveness of the MIRNet model. The image prediction results shown on the images in unconstrained environment validate the use of enhancement model. Finally, the comparison of our results with existing methods for predicting traffic signs in fine-grained classes in GTSRB dataset shows that the proposed method outperforms other approaches.