A Single Residual Network with ESA Modules and Distillation

Abstract

Although there are many methods based on deep learning that have superior performance on single image super-resolution (SISR), it is difficult to run in real time on devices with limited computing power. Some recent studies have found that simply relying on reducing parameters or reducing the theoretical FLOPs of the model does not speed up the inference time of the network in a practical sense. Actual speed on the device is probably a better measure than FLOPs. In this work, we propose a new single residual network (SRN). On the one hand, we try to introduce and optimize an attention mechanism module to improve the performance of the network with a relatively small speed loss. On the other hand, we find that residuals in residual blocks do not have a positive impact on networks with adjusted ESA. Therefore, the residual of the network residual block is removed, which not only improves the speed of the network, but also improves the performance of the network. Finally, we reduced the number of channels and the number of residual blocks of the classic model EDSR, and removed the last convolution before the long residual. We set this tuned EDSR as the teacher model and our newly proposed SRN as the student model. Under the joint effect of the original loss and the distillation loss, the performance of the network can be improved without losing the inference time. Combining the above strategies, our proposed model runs much faster than similarly performing models. As an example, we built a Fast and Efficient Network (SRN) and its small version SRN-S, which run 30%-37% faster than the state-of-the-art EISR model: a paper champion RLFN. Furthermore, the shallow version of SRN-S achieves the second-shortest inference time as well as the second-smallest number of activations in the NTIRE2023 challenge. Code will be available at https://github.com/wnxbwyc/SRN.