MESR: Multistage Enhancement Network for Image Super-Resolution

Recently, the deep-learning-based image super-resolution methods have achieved astounding advancement. Whereas most of these methods utilize features from the low-resolution image space exclusively, and ignore the dependency between contextual features simultaneously, resulting in their limited ability to restore details. To this end, a multi-stage enhancement image network for super-resolution (MESR) is proposed. The network consists of two stages, where the first stage is used to generate a coarse reconstructed image, and the second one is to refine the coarse image, which enhances the super-resolution performance. Specifically, in the first stage, to acquire more abundant features, an effective funnel-like multi-scale feature extractor is proposed, incorporating a channel attention mechanism to boost the feature representation capability. Moreover, an adaptive weighted residual feature fusion block is designed to effectively explore and exploit the dependency between contextual features for generating more beneficial features. In the second stage, a refinement block is proposed to additionally strengthen the details of the reconstructed image by exploring the feature information from the high-resolution image space. Experimental results demonstrate that the proposed method achieves superior performance against the state-of-the-art SR methods in terms of both subjective visual quality and objective quantitative metrics.