End-to-end infrared radiation sensing technique based on holography-guided visual attention network

Infrared radiation imaging is extensively utilized due to its unique advantages in terms of wavelengths. This paper presents an end-to-end short-wave infrared digital holography system at 1550 nm that integrates hardware and algorithms to reconstruct targets in a lens-free system. To address challenges such as inherent noise, twin image, and proper focusing encountered in traditional holographic process, a holography-guided visual attention network with a Squeeze and Excitation module, named Diffraction Multi-Distance Regulation Attention Network, is employed to resolve the inverse process. The proposed approach achieves noise-free image reconstruction through a well-trained convolutional neural network, eliminating the need for prior knowledge. The effectiveness of the proposed method is verified through comprehensive experiments involving amplitude and phase type samples, simulated and physical objects, as well as trained and unknown targets. Compared to existing infrared band detection techniques, intelligent infrared digital holography enables efficient phase imaging without time-consuming iterations. Furthermore, the net-training for infrared digital holography relies on simulated data rather than on experimental big data, providing promising prospects for future research endeavors.