Fabrication-friendly filtering channel optimization strategy for hyperspectral reconstruction

Abstract

Hyperspectral image recovery has attracted much attention due to its cost-effectiveness. Existing approaches mainly focus on enhancing reconstruction accuracy, either by optimizing network architectures from single RGB image or by optimizing customized filtering channels. However, reconstruction from an RGB image is limited due to their inherent channels. More critically, current methods for optimizing customized channels always neglect critical fabrication feasibility. Here, we propose a fabrication-friendly filtering channel optimization strategy for hyperspectral reconstruction. To the best of our knowledge, it is the first strategy which achieves the balance between high reconstruction accuracy and fabrication feasibility. This is realized by primarily combining the target feature matching and channel correlation-based optimization method with the film design optimization method accounting for manufacturing errors. Our proposed strategy has been validated using both synthetic datasets and real-world scenarios. Experimental results on synthetic datasets demonstrate the proposed method outperforms existing channel optimization methods in both reconstruction accuracy and fabrication friendliness. In real-world testing, our method improves reconstruction accuracy compared to conventional RGB channels. Specifically, it achieves a peak signal-to-noise ratio improvement of over 18.2% while reducing the root mean square error by at least 45.6%, the mean relative absolute error by no less than 50.2 %, and the spectral angle mapper by a minimum of 20.8%. Furthermore, our strategy can integrate filtering channels with existing multispectral systems, especially in which filtering wheels and multispectral filtering arrays setups, making it particularly suitable for weight-constrained, real-time applications like aerial surveillance or mobile sensing.