Multi-Residual Feature Fusion Network for lightweight Single Image Super-Resolution

Recently, single image super-resolution (SISR) methods using deep convolution neural networks (CNNs) have achieved remarkable performance. Especially, lightweight net-works have received unprecedented attention because of their broad application prospects. However, existing methods for lightweight SR lack the adequate utilization of hierarchical features, which weakens the representation ability of the net-work. To alleviate this issue, we propose an effective and accurate multi-residual feature fusion network (MRFFN) for SISR. Specifically, we design a multi-residual block (MRB) to boost the representation ability of the network. By adopting the multi-residual learning (MRL) strategy, MRB can efficiently improve reconstruction results while halving the parameters, compared with the ordinary residual block (RB). To use the hierarchical features sufficiently, we construct a multi-residual fusion block (MRFB) by cascading the MRBs. Finally, we build our MRFFN by densely stacking MRFBs and introduce double-residual learning (DRL) strategy into the network at the global level. Extensive experiments demonstrate that the MRFFN is superior to the state-of-the-art SISR models while taking up less computing resources.