Two-stage framework for reconstructing spectral images from diffraction-blurred images

Abstract

In recent years, encoding and decoding-based computational spectral imaging systems have garnered substantial interest. Among these, phase-encoded spectral imaging systems employ diffractive optical elements to encode the phase of the incident light, resulting in wavelength-dependent point spread functions. These point spread functions encode implicit spectral information into the diffraction-blurred images, enabling convolutional neural network-based spectral reconstruction algorithms to extract spectral information effectively. However, the large size of the point spread function leads to considerable degradation. Traditional one-stage reconstruction frameworks attempt to address both deblurring and spectral unmixing simultaneously, and the quality of the reconstruction results requires further improvement. In this paper, we propose an innovative two-stage reconstruction framework. In the first stage, we focus on restoring high spatial resolution deblurred images. In the second stage, we use the deblurred images along with the original diffraction-blurred images to perform spectral reconstruction. Experimental results demonstrate that our proposed two-stage reconstruction framework, alongside the feature extraction and fusion block, significantly enhance the quality of the reconstruction results. Our framework can also be extended to other reconstruction tasks based on point spread function encoding. Moreover, we provide a relative spectral reflectance dataset covering a spectral range from $400nm$ to $900nm$, with a spectral interval of $10nm$, which is conducive to improving the generalization of the spectral reconstruction networks in diverse scenes.