Single-image super-resolution reconstruction using dark channel regularization network

For single-image super-resolution (SR), deep learning-based approaches have attained superior performance that overshadow all previous approaches. Most recently published deep learning-based single-image SR approaches rely on either deeper or more complex network to achieve further improved results, which are time and space intensive. In this paper, we propose a new method to effectively improve the quality of the final magnified image: a dark channel prior-based network is first designed and then used to regularize the previously existed SR networks. The motivation of this work is an interesting observation that the dark channel of the magnified image contains less dark pixels than that of the original high-resolution image. Since the proposed dark channel prior-based network is a fixed network and does not contain any trainable parameters, the combined hybrid network thus can maintain its original complexity and achieve state-of-the-art results.