End-to-End Image Demosaicking via Region-Level Non-Local Modeling and Residual Aggregation

Image demosaicking aims to reconstruct a full-resolution color image from spatially sparse and interleaved color filter array observations. Despite the significant progress achieved by deep learning-based methods, existing approaches have not fully addressed the sampling-structure-constrained nature of demosaicking. In particular, four-channel half-resolution packing may disrupt the CFA spatial phase relationships, while local convolutions and global non-local matching struggle to model reconstruction-relevant cross-position dependencies. To address these issues, this paper proposes an end-to-end image demosaicking network with region-level non-local modeling and residual aggregation (RNRA-Net). Instead of packing Bayer RAW data into a four-channel half-resolution representation, RNRA-Net decomposes the original mosaic image into a three-channel representation at the original resolution, thereby preserving the spatial arrangement of CFA sampling. To capture structurally related information, a region-level non-local module is introduced to compute feature correlations within spatially bounded regions, enabling aggregation of reconstruction-relevant contextual information. In addition, a residual aggregation module is developed to explicitly collect and refine early residual compensation features, facilitating the recovery of edges, textures, and high-frequency details. Extensive experiments on benchmark and high-resolution datasets demonstrate the effectiveness of RNRA-Net.