Application of a CNN-Based Reconstruction Method in a Dual-Dispersive Coded Aperture Spectral Linear-Polarization Imager

Based on computational optical imaging theory, coded aperture compressive imaging captures multidimensional target data by introducing coded apertures into the optical imaging system. This approach allows for the approximate reconstruction of the original multidimensional data from a single or a few compressive measurements, which would typically require many more measurements in conventional imaging systems. As a result, it can significantly enhance the measurement capacity and efficiency of detection devices. In this article, we utilize a coded aperture spectral polarization imaging prototype as the research object and apply a classical convolutional neural network (CNN) to reconstruct the target’s 2-D spatial, 1-D spectral, and 1-D polarization information from a snapshot 2-D measurement. Experiments are conducted using both simulated and real data. Inspired by the anisotropic dispersion properties within the optical system, it is found that using anisotropic nonsquare convolution kernels correspondingly at specific network layers significantly improves reconstruction performance in both spatial and spectral domains, without adding extra network parameters or calculation cost. Additionally, the robustness and nonlinear response characteristics of polarization detection are analyzed. In real-data experiments, the results show that the proposed network model, utilizing strip-type convolution kernels, demonstrates effectiveness and efficiency in spatial, spectral, and polarization reconstruction from a single compressive measurement. The peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) remain above 21 dB and 0.77, respectively. The spectral angle between the curves reconstructed and referenced is controlled within 0.15 rad, and the absolute error of polarization degree is within 0.004.