Unsupervised Super Resolution Reconstruction of Traffic Surveillance Vehicle Images

Abstract

The surveillance of public transportation is of great significance to improve public safety. However, the low resolution of vehicle images becomes a bottleneck in real scenarios. Since high-low resolution vehicle images pairs are not available in traffic surveillance scenarios, this paper aims to study the problem of unsupervised super-resolution to reconstruct the high quality vehicle image. Most of the existing super-resolution algorithms adopt pre-defined down-sampling methods for paired training, however, the models trained in this pattern cannot achieve the expected results in traffic surveillance scenarios. Therefore, we propose a super-resolution method that does not require paired data, and raise a novel down-sampling network to generate low-resolution images of vehicles close to the real-world data, and then utilize the synthesized pairs for pair-wise training. Our extensive experiments on private real-world dataset Vehicle5k demonstrate the advantages of the proposed approach over baseline approaches.