Image Super-Resolution Using Very Deep Residual Channel Attention Networks

The paper tackles image super-resolution, where CNN depth is crucial but very deep networks are hard to train, and where low-resolution inputs carry abundant low-frequency information that is treated equally across channels, hindering the network's representational ability. To solve this, the authors propose the very deep residual channel attention network (RCAN). It uses a residual-in-residual (RIR) structure—several residual groups with long skip connections, each containing residual blocks with short skip connections—so abundant low-frequency information can be bypassed through multiple skip connections, letting the main network focus on high-frequency information, while a channel attention mechanism adaptively rescales channel-wise features by modeling interdependencies among channels.

Extensive experiments show that RCAN achieves better accuracy and visual improvements than state-of-the-art super-resolution methods. The combination of a very deep architecture made trainable through nested residual learning and the channel attention mechanism demonstrated that adaptively emphasizing informative frequency components substantially improves reconstruction quality, influencing subsequent attention-based super-resolution designs.