Multi-level continuous encoding and decoding based on dilation convolution for super-resolution

Deep neural networks have shown better effects for super-resolution in recent years. However, it is difficult to extract multi-level features of low-resolution (LR) images to reconstruct more clear images. Most of the existing mainstream methods use encoding and decoding frameworks, which are still difficult to extract multi-level features from low resolution images, and this process is essential for the reconstruction of more clear images. To overcome these limitations, we present a multi-level continuous encoding and decoding based on dilation convolution for super-resolution (MEDSR). Specifically, we first construct a multi-level continuous encoding and decoding module, which can obtain more easy-to-extract features, complex-to-extract features, and difficult-to-extract features of LR images. Then we construct dilated attention modules based on different dilated rates to capture multi-level regional information of different respective fields and focus on each level information of multi-level regional information to extract multi-level deep features. These dilated attention modules are designed to incorporate varying levels of contextual information by dilating the receptive field of the attention module. This allows the module to attend to a larger area of the input while maintaining a constant memory footprint. MEDSR uses multi-level deep features of LR images to reconstruct better SR images, the values of PSNR and SSIM of our method on Set5 dataset reach 32.65 dB and 0.9005 respectively when the scale factor is ×4. Extensive experimental results demonstrate that our proposed MEDSR outperforms that of some state-of-the-art super-resolution methods.