Multi-Scale Residual Attention Network for Image Super-Resolution
Shuiping Ni, Shijie Wang, Huifang Li, Pengkun Li · 2023
Single image super-resolution (SISR) reconstruction algorithms mostly use a single scale convolution kernel to extract the features of low-resolution images, which will easily cause a large number of details loss. Usually, the depth of the model network is increased to achieve better image reconstruction effect. With the increase of the depth of the network, the model gradient disappears, and increase network parameters, making the training difficult. A multi-scale residual attention model is proposed to achieve single image super-resolution reconstruction. The model consists of shallow feature extraction, multi-scale residual attention network and reconstruction. The convolution kernels of different scales are used for feature extraction of low-resolution image convolution. Residual attention network can eliminate gradient disappearance and network degradation. Adding attention mechanism in residual network can make the high-frequency region characteristics of network channel get more attention, and enhance the texture and detail information of reconstructed image.