Adversarial Training for Boosting Super-Resolution by Combination with License Plate Recognition

Super-resolution (SR) and license plate recognition (LPR) are popular tasks in image analysis. SR is to produce higher resolution images from low-resolution ones. Meanwhile, LPR produces license numbers in the text format. It is well-known that SR can produce higher resolution images; so it can cause LPR achieving better performance. On the other hand, this paper presents a multi-task deep-neural-network by exploring an adversarial property between SR and LPR to enhance the quality both of the tasks. We also present a new method for training such kind of multi-task network, which is based on adversarial learning. For each step in the proposed training algorithm, SR improves its performance and helps LPR doing better; and, LPR, which uses the most recent updated model of SR (in the previous training step), should improve itself and can force SR being more accurate in the next training step. The experiments in the paper show that the proposed multi-task network in combination with the new (adversarial) training outperforms most of recent researches in super-resolution.