Multi-Scale Image Super-Resolution via Hierarchical Filter Groups

摘要

The development of deep convolutional neural networks (DNN) has progressed the recent research interest on super-resolution (SR). The existing DNNs benefited from residual learning and achieved improved performance. In this paper, we propose a discrete wavelet-based hierarchical filter groups (HFG) for multi-scale image SR. Discrete wavelet transform (DWT) produces four sub-bands with different frequency components. Three detail sub-bands (i.e., horizontal, vertical and diagonal) contain sparse wavelet coefficients whereas, approximation sub-band contains average spatial image information. The reconstruction accuracy is improved by exploiting the information in all the wavelet sub-bands simultaneously. Furthermore, hierarchical filter groups are adopted for creating computationally efficient DNN without compromising performance. Moreover, symmetrical skip connections are used for feature reuse in different layers of DNN to improve reconstruction accuracy. In addition, the problem of vanishing of gradients and feature redundancy are mitigated by short skip paths created using symmetrical skip connections. The proposed HFG method achieves superior performance over the other recent competitive methods on benchmark datasets.