A Multi-Scale Layer-Channel Attention Network for Image Super-Resolution

Recently, image super-resolution (SR) methods widely employ deep convolutional neural networks (CNNs) to improve the performance of reconstructed images through extracting informative features. However, the most of existing SR networks ignore the interdependencies among layers from different modules, which might hinder the further improvement of representational ability. In this paper, a newly designed multi-scale layer-channel attention network (MSLCAN) is proposed for single image super resolution (SISR) task. Our contributions are twofold. First, the layer-channel attention module (LCAM) can adaptively emphasize the correlations among different channels and interrelationships among multi-scale layers. Second, the multi-scale feature extraction module (MSFEM) effectively extracts adequate image information and enhances network ability of feature representation. In the experiments on five standard benchmarks, we have observed that our MSLCAN outperforms other SISR methods on quantitative and perceptual quality.