A Deep Dual-Branch Networks for Joint Blind Motion Deblurring and Super-Resolution

Abstract

Image super-resolution is a fundamental pre-processing technique for the machine vision applications of robotics and other mobile platforms. Inevitably, images captured by the mobile camera tend to emerge severe motion blur and this degradation will deteriorate the performance of current state-of-the-art super-resolution methods. In this paper, we propose a deep dual-branch convolution neural network (CNN) to generate a clear high-resolution image from a single natural image with severe blurs. Compared to off-the-shelf methods, our method, called DB-SRN, can remove the complex non-uniform motion blurs and restore useful texture details simultaneously. By sharing the features from modified residual blocks (ResBlocks), the dual-branch design can promote the performances of both tasks other while retaining network simplicity. Extensive experiments demonstrate that our method produces remarkable deblurred and super-resolved images in terms of quality and quantity with high computational efficiency.