Hyperspectral image reconstruction algorithms for broadband spectral modulation: a comprehensive review

Spectral-modulation-based hyperspectral imaging is an important approach for acquiring highdimensional hyperspectral data via computational reconstruction.This method addresses limitations of traditional scanning systems in terms of hardware complexity, acquisition speed, and data storage.This review systematically summarizes the development and current status of reconstruction algorithms for spectral-modulation-based hyperspectral imaging.Existing methods are categorized into three main types: prior-constraint approaches based on mathematical models, learning-based methods using deep neural networks, and hybrid frameworks that combine model-based priors with data-driven strategies.Priorconstraint methods, such as those using sparsity, low-rank, or regularization constraints, offer solid theoretical foundations and interpretability, but they often face challenges related to parameter selection and computational efficiency.Deep learning approaches, including convolutional neural networks and transformer architectures, can achieve high reconstruction accuracy and real-time performance, but they typically require large amounts of training data and may have limited generalization to diverse imaging conditions.Hybrid methods, especially deep unfolding networks, seek to balance theoretical interpretability and adaptive performance.Key challenges include balancing reconstruction accuracy and computational cost, integrating physical knowledge into network design, and establishing standardized evaluation frameworks.This review aims to provide a reference for the development and application of spectralmodulation-based hyperspectral imaging in areas such as remote sensing, biomedical diagnostics, and industrial inspection.