RCAn

RCAN, in the context of computer vision, most commonly refers to the Residual Channel Attention Network, a deep convolutional network designed for single-image super-resolution (SISR). The architecture emphasizes channel-wise feature recalibration to enhance representational capacity, enabling more accurate reconstruction of high-resolution images from low-resolution inputs. RCAN is built to be very deep while remaining trainable, leveraging a residual-in-residual (RiR) structure and multiple skip connections to promote information flow and ease optimization.

The network is organized into multiple residual groups (RGs). Each RG contains several residual channel attention blocks (CABs). Within a CAB, feature maps are refined by a channel attention mechanism that first aggregates spatial information via global average pooling to form channel descriptors, then passes these through a small two-layer perceptron to produce channel-wise weights. Applying these weights to the original features amplifies informative channels and suppresses less useful ones, enabling the network to focus on relevant textures and details. The RiR configuration provides hierarchical skip connections across blocks and groups, allowing very deep models to be trained effectively without vanishing gradients.

RCAN is typically trained on paired low- and high-resolution image datasets, using loss functions such as L1 or perceptual losses. When introduced, RCAN demonstrated state-of-the-art performance on standard SR benchmarks, offering improvements in PSNR and SSIM and producing sharper textures in reconstructed images compared with many prior methods. Variants of RCAN explore different depths and channel counts, including deeper models with hundreds of layers to push performance further.

Other uses of the RCAN acronym exist in different domains, but in the ML literature RCAN most commonly denotes the Residual Channel Attention Network for image super-resolution.