Deep recursive multi-scale residual network for single image super-resolution reconstruction

Due to weather, camera technology, compression transmission, etc., people are not always able to obtain highresolution images that meet the needs of specific scenes. A deep recursive multi-scale residual network (DRMRN) for single image super-resolution is possessed to solve this problem. The proposed multi-scale residual block is based on the basic residual unit and different branches consist of convolution kernels of different sizes. Adaptive image feature extraction is beneficial to fully extract and utilize LR image features. In addition, our network has both local residual learning and global residual learning, which reduces the difficulty of training and solves the problem of gradient disappearance and gradient explosion in the training process. Compared with other models, the proposed model improves the reconstruction effect, and the reconstruction results have higher PSNR and SSIM values.