Efficient Screen Content Image Super-Resolution with Weighted Multi-Branch Convolution and Attention

In recent years, with the rapid advancement of smart media technologies, screen content images have become increasingly-prevalent in various aspects of daily life. Applications such as screen content projection and online classrooms have heightened the demand for real-time performance on image super-resolution task. In response to these developments, and considering the characteristic abundance of sharp edges and rich details in screen content images, we propose an efficient model with Weighted Multi-Branch Convolution and Attention(WMBCA) module, composed of Weighted Multi-Branch Convolution Blocks (WMBCB) and Contrast-aware Channel Attention (CCA). The WMBCB and CCA within the module are respectively utilized for extracting information from low-resolution images and integrating information across channels. Our proposed model enhances the representation of the abundant edge information in screen content images. In the experiments, our model demonstrated superior performance compared to other lightweight SR models in screen content image super-resolution task. Furthermore, the structure effectively balances inference speed and results, offering high practicality in real-time screen content image restoration task.