Vol. 56, Issue 2, pp. 295-312 (2026)

Abstract

Miniaturized spectrometers are gaining significant attention in scientific and industrial fields due to their compact footprint, low cost, and on-site analysis capability. Spectral reconstruction algorithms are crucial for computational spectrometers to achieve accurate signal demixing and recovery. However, these algorithms often struggle under low-light or noisy conditions. While deep learning-based methods show promise, they remain insufficient for computational spectrometers employing band-stop encoding arrays, which offer high optical throughput but introduce strong spectral interference and complex measurement challenges. This work proposes a novel spectral reconstruction framework that explicitly integrates physical priors through a multi-head cross-attention mechanism. This approach dynamically models the interaction between the measurement matrix and observed signals, enabling effective multimodal feature fusion and enhanced extraction of discriminative spectral features. An experimental band-stop computational micro-spectrometer system was built to acquire real data, and data augmentation strategies were applied to improve model generalization. Comparative experiments with various algorithms on both datasets validate the effectiveness of fusing physical prior information via the multi-head cross-attention mechanism. This study presents a new paradigm for spectral reconstruction in band-stop computational micro-spectrometers, offering an effective solution for interference-resistant spectral analysis in computational spectrometry.