Researchers from the Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences, publishing in Scientific Reports, have created a new infrared imaging system that improves the clarity of dim, small targets often obscured by noise.

Infrared imaging faces a major challenge: dim, distant targets often appear as just a few blurry, noisy pixels, making them hard to distinguish from background clutter. Traditional methods to enhance image resolution (super-resolution) rely on sequences of low-resolution (LR) images. These sequences often suffer from time delays between frames and slight camera movements (spatial displacement), introducing errors that limit the quality of the final high-resolution (HR) image. Furthermore, existing super-resolution algorithms typically require pre-existing high-resolution images for training, which aren't available for novel, dim targets in real-world scenarios.

To tackle the source error problem, the team constructed a specialized Reflective Infrared Micro-Scanning Optical (RIMO) system. This hardware solution uses a precisely controlled, high-speed vibrating mirror. "The RIMO system generates multiple, perfectly aligned low-resolution image frames in rapid succession," explained lead researcher CHEN Jian. "It eliminates the time lag and spatial displacement inherent in capturing separate images over time with a standard camera." This provides a clean sequence of LR images as input for the super-resolution process, breaking a fundamental limitation of software-only approaches.

Addressing the algorithmic challenge, the researchers developed a Self-Supervised Super Resolution Restoration (SSRR) algorithm. Unlike traditional methods needing HR examples for training, the SSRR algorithm works independently. "Our key innovation is that the algorithm learns to estimate both the image blur and the potential high-resolution image itself directly from the noisy, low-resolution sequences provided by the RIMO system, without any external high-resolution supervision," Chen noted. It employs a feature alignment network using deformable convolution to intelligently combine information from neighboring frames in the sequence, further enhancing its ability to reconstruct details and suppress noise.

The combined RIMO-SSRR system was rigorously tested. Simulations using both infrared and visible light cameras compared the new method against traditional super-resolution techniques like POCS. Visual results demonstrated a stark difference: images processed with the traditional method showed significant background noise and blurry target outlines, while the SSRR algorithm produced cleaner backgrounds and sharper, more defined targets. Objective measurement using Peak Signal-to-Noise Ratio (PSNR) confirmed the visual improvement, showing the SSRR method achieved an over 16% higher PSNR than the conventional approach, indicating superior image fidelity and noise suppression.

This advancement holds promise for enhancing infrared detection capabilities. By enabling clearer imaging of tiny, dim targets – even in noisy conditions, the RIMO-SSRR system improves target tracking accuracy, and overall situational awareness. The integration of specialized hardware (RIMO) with a powerful, self-taught algorithm (SSRR) represents a robust solution to a persistent challenge in infrared optoelectronics.