Convolutional Neural Network Design for Single Image Super-Resolution

Lecture notes in electrical engineering(2023)

引用 0|浏览6
Single image super-resolution (SR) is designed to recover high-resolution (HR) images from a single low-resolution (LR) image, which has important applications in surveillance equipment, satellite imagery, mobile phone cameras and medical images. Skip-connection is widely used in SR networks (e.g., ESRGAN and RDN), and have led to further performance improvements. However, SR models with many skip-connections are often accompanied by extremely high computation, which hinders their deployment on mobile devices. To solve this problem, we propose a connection pruning algorithm that automatically prunes redundant connections to make the network more compact. By connection pruning, a lightweight skip-connection structure is generated, called Residual-in-Residual Sparse Block (RRSB), which improves the performance of the network while reducing the computation. Our proposed lightweight network is evaluated on Set14, BSD100 and Urban100 datasets. The experimental results show that our network performance does not degrade in the case of a 50% reduction in the amount of parameters. The textures generated by our network are close to baseline and are sharper and more natural than other methods.
convolutional neural network,neural network,super-resolution
AI 理解论文
Chat Paper