CNN-based Super-resolution Reconstruction for Traffic Sign Detection

Automatic identification for traffic signs is an important part of intelligent driving and traffic safety. Deep learning has already made a great achievement in traffic sign detection. However, the camera on a car may capture a low resolution and blurry image in certain environments, which make it inaccurate for traffic sign detection. Therefore, we propose a method based on image super-resolution reconstruction for improving the detection rate of traffic signs. Firstly, a low-resolution image is transformed by CNN-based super-resolution network into a high-resolution one. Then, to meet the requirements of on-line processing, we use the generated super-resolution image as input for the detection network with 16 filters in this layer. At last, we separately trained two CNNs for super-resolution reconstruction and traffic sign detection, which reduce the processing time. Experimental results demonstrate that our model can achieve better performance than the existing methods for traffic sign detection.