Deep Compressed Super-Resolution Imaging with DMD Alignment Error Correction

Abstract

In the field of compressed imaging, many attempts have been made to use the high-resolution digital micromirror array (DMD) in combination with low-resolution detectors to construct imaging systems by collecting low-resolution compressed data to reconstruct high-resolution images. However, the difficulty of achieving micrometer-level alignment between DMD devices and detectors has resulted in significant reconstruction errors. To address this issue, we proposed a joint input generative adversarial network with an error correction function that simulates the degradation of image quality due to alignment errors, designed an optical imaging system, and incorporated prior imaging system knowledge in the data generation process to improve the training efficiency and reconstruction performance. Our network achieved the ability to reconstruct 4× high-resolution images with different alignment errors and performed outstanding reconstruction in real-world scenes. Compared to existing algorithms, our method had a higher peak signal-to-noise ratio (PSNR) and better visualization results, which demonstrates the feasibility of our approach.