Residual Adaptive Dense Weight Attention Network for Single Image Super-Resolution

Recently, with the rise and progress of convolutional neural networks (CNNs), CNN-based single image super-resolution (SISR) methods have gained considerable advancement and showed great power for image reconstruction tasks. Never-theless, existing methods cannot dynamically adjust the network according with the input, which greatly impairs the practical performance of the network. To address this issue, a novel residual adaptive dense weight attention network (RADWAN) is proposed consisted of several adaptive residual groups (ARGs) to enhance the generalization performance of the network. Specifically, each ARG contains several adaptive dense weight attention blocks (ADWAB). It generates dense connection coefficients dynamically using an adaptive dense weight block (ADWB) for more accurate feature extraction. Besides, an avg-std channel attention block is further presented to maximize the potential of RADWAN to make the model focus on critical information. To adjustably utilize additional features from shallow and intermediate layers, we introduce adaptive short connections (ASC) and adaptive long connections (ALC) to effectively integrate abundant hier-archical features. Extensive experiments on several datasets have demonstrated the superiority of RADWAN over the state-of-the-art methods in aspects of both quantitative metrics and visual quality.