Mamba Guide Hyperspectral Image Reconstruction and Defocus Map Estimation

Hyperspectral images(HSIs) defocus deblurring and defocus map estimation(DME) are crucial and challenging, especially when dealing with large-scale data. Existing convolutional neural network-based methods exhibit inherent limitations in modeling long-range dependencies, while Transformer-based approaches suffer from high computational complexity when capturing global contextual correlations. Although existing Mamba-based HSI methods (such as SSUMamba and 3DSS-Mamba) have shown promise, they primarily focus on additive noise removal or classification tasks, and their scanning strategies often disrupt the spatial-spectral intrinsic continuity required for modeling spatially varying blurring. To address these challenges, Joint Defocus Map Estimation and deblurring Reconstruction (JDMEDR) framework with the spatial-spectral Transformer Mamba(SSFMamba) modules network is proposed for enhancing the reconstruction of defocused HSIs effectively. The SSFMamba, constructed by hybridizing Mamba context interaction modules with a Spatial-Spectral Layer-by-Layer Scanning (SSLS) strategy, is designed to balance long-range perception and reduction of computation complexity as state-space model. In addition, a DME network is proposed based on reconstructed HSIs adjusted through the defocus generation mechanism by ResNet, enabling the learning of a direct mapping from clean input images to blurred HSIs, thereby further evaluating the effectiveness of the defocused HSI reconstruction algorithm. Extensive experiments conducted on both synthetic and real HSIs datasets demonstrate that the proposed SSFMamba achieves state-of-the-art reconstruction performance and defocus estimation accuracy compared with other state-of-the-art methods.