DnCAT: Learning Residual Denoising and Channel Attention for Accurate RGB-to-HSI Reconstruction

Hyperspectral imaging (HSI) has shown great potential in areas ranging from remote sensing to medical imaging due to its ability to capture detailed spectral information across hundreds of bands. However, direct HSI acquisition remains costly and challenging. In contrast, RGB imaging is inexpensive but limited to three bands, restricting its applications. Since both modalities have complementary strengths, reconstructing HSI from RGB offers a cost-effective and practical alternative, making spectral analysis more accessible. Early approaches relied on convolutional neural networks (CNNs), but these often fail to capture long-range dependencies and nonlocal repeating patterns that are essential for accurate reconstruction. Attention-based models have improved results, yet state-of-the-art methods typically suffer from large parameter counts, complex architectures, and limited interpretability. Moreover, attention-based models commonly use multi-scale resolution strategies, which can reduce reconstruction accuracy by losing subtle spatial details. To address these challenges, we propose DnCAT, a transformer in U-shape, designed for HSI reconstruction. DnCAT integrates a transformer encoder-decoder architecture with a bridge module that connects the feature encoder and feature decoder. Its multi-head spectral attention leverages the spatial sparsity and spectral self-similarity of HSI data, improving spectral modeling while significantly reducing parameters. Unlike existing models, DnCAT avoids feature downscaling, preserving fine-grained details crucial for accurate reconstruction. Experiments demonstrate that DnCAT achieves strong performance, with an MRAE of 0.257 and a PSNR of 31.32 representing at least a 2.18% reduction in MRAE compared to previous state of the art methods. These results establish DnCAT as an efficient and reliable benchmark for HSI reconstruction.