MixerSR: A New Feature Extraction Paradigm for Single Image Super-Resolution.

Single-image super-resolution (SR) tasks have achieved fancy success in recent years by leveraging deep convolution neural network (CNN). Although CNNs obtain powerful representation capabilities of reconstructing a high-resolution (HR) image from its corresponding low-resolution (LR) observation, it is hard to apply these accomplishments to real world for its large number of parameters that bring about huge computing resource cost. To solve this problem, many researchers turn their directions to lightening the network by delicate designs while keeping the performance in the high level. In this paper, we propose a new method to further reduce the parameters by using the new feature extraction paradigm called Mixer which only contains Multi-Layer Perceptron (MLP). Compared to convolutional operations, Mixer operation has the same representation capabilities with fewer parameters. In this paper, three typical lightweight networks, where we replace the convolutional operation with Mixer, are used to show the excellent ability. Our experimental results demonstrate that Mixer can help these lightweight networks further reduce the number of parameters by up to 38% while keep their performance in the same level.