Hierarchical dense recursive network for image super-resolution

Abstract Image super-resolution (SR) techniques with deep convolutional network (CNN) have achieved significant improvements compared to previous shallow-learning-based methods. Especially for dense connection based networks, these methods have yielded unprecedented achievements but bring the higher complexity and more parameters. To this end, this paper considers both reconstruction performance and efficiency, and advocates a novel hierarchical dense connection network (HDN) for image SR. First of all, we construct a hierarchical dense residual block (HDB) to promote the feature representation while saving the memory footprint with a hierarchical matrix structure design. In this way, it can provide additional interleaved pathways for information fusion and gradient optimization but with a shallower depth compare to the previous networks. In particular, a group of convolutional layers with small size (1 × 1) are embedded in HDB, releasing the computational burden and parameters by rescaling the feature dimensions. Furthermore, HDBs are connected to each other in a sharing manner, thereby allowing the network to fuse the features in different stages. At the final, the multi-scale features from these HDBs are integrated into global fusion module (GFM) for a global fusion and representation, and then the final profile-enriched residual map is obtained by realigning and sub-pixel upsampling the fusion maps. Extensive experimental results on benchmark datasets and really degraded images show that our model outperforms the state-of-the-art methods in terms of quantitative indicators and realistic visual effects, as well as enjoys a fast and accurate reconstruction.