Sellf-Adaptive Weighted Skip Connections for Image Super-Resolution

Recently, the introduction of deep convolutional neural networks has achieved great performance in low-level vision tasks like single image super-resolution. However, deeper networks tend to have larger number of parameters and be more difficult to be trained. Considering massive low-frequency information in low-resolution inputs, we propose a self-adaptive weighted skip connections (SAWSC) structure to make full use of both the low-level and high-level features in order to better the representation ability of super-resolution networks. In our proposed network, we follow a coarse-to-fine strategy, which reconstructs high-resolution images progressively based on the Laplacian pyramid. At each upscale level, feature maps of each block are connected to subsequent blocks with self-adaptive weights. During each block, several residual channel attention layers are cascaded. Evaluations on five public benchmark datasets show that this algorithm achieves better performance than some other existing methods.