Real-Time Channel Mixing Net for Mobile Image Super-Resolution

Garas Gendy, Nabil Sabor, Jingchao Hou, Guanghui He

Abstract: Recently, deep learning based image super-resolution (SR) models show a strong performance thanks to the convolution neural network (CNN). However, these CNN-based models mostly need large memory and use a lot of power cost, which limits its use in mobile devices. To solve this problem, we propose a channel mixing Net (CDFM-Mobile) for mobile SR. The idea of the CDFM-Mobile is based on making channel mixing by using a pointwise convolution and deep features extraction by using 3 \(\times \) 3 convolution. In addition, inspired by the prior work in the field, we used anchor-based residual learning and deep feature residual learning, which improved the performance. In addition, we used the quantization-aware training approach to optimize the model performance based on training at 8-bit quantize. Finally, we take part in MAI 2022 for mobile SR, and extensive results are conducted to show the model performance.