Residual multi-branch distillation network for efficient image super-resolution

A Residual Multi-branch Distillation Network (RMDN) is proposed and implemented for efficient image super-resolution (SR) by reducing the redundancies in the convolution operations which exists in current information distillation network-based models. To enhance the model ability, an attention module is designed to fuse the distilled features. By reducing the redundancies and focusing on the distilled features, the network can improve the reconstruction performance with low computation costs. Specifically, the Multi-branch Aware Convolution (MAConv) is introduced in the Residual Multi-branch Distillation Block (RMDB), which is the basic building block of RMDN. MAConv utilizes multiple kernel sizes and dilation rates in depth-wise separable convolutions to reduce redundancies and introduce hierarchical receptive fields to extract multi-level features. Meanwhile, a Residual Hybrid Attention Module (RHAM) integrates the reduction-and-expansion strategy and the Residual Hybrid Attention (RHA) in RMDB to efficiently fuse and enhance the aggregated distilled features. To improve the training stability, pixel normalization is adopted after RHAM. Extensive experiments show that the proposed RMDN performs favorably against other state-of-the-art image SR methods in terms of computation costs and image reconstruction quality.