With the increasing frequency of natural disasters and health emergencies, wearable infrared thermal imaging devices have gained widespread use in the firefighting and medical fields. However, such devices tend to have poor imaging performance and often suffer from low contrast, dark areas, high noise and blurred boundaries, which greatly hinder practical applications.
To address these issues, the research team from University of Science and Technology of China and AVIC Huadong Photoelectric published their latest findings on 15 January 2026 in Frontiers of Computer Science , a journal co-published by Higher Education Press and Springer Nature.
The team proposed a novel unsupervised lightweight 3D convolutional network (UL3DCN) specifically designed for infrared image enhancement on wearable devices. In this framework, the task of infrared image enhancement is conceptualized as generating high dynamic range infrared images from the corresponding temperature sequences during thermal equilibrium. Comprehensive experimental results demonstrate that the method achieves excellent image enhancement effect and good real-time performance. Additionally, the proposed UL3DCN model has been successfully integrated into a wearable infrared firefighting mask.
In this research, they draw inspiration from HDR imaging technology, and formulate the task of infrared image enhancement as the generation of HDR infrared images from corresponding infrared images at different temperatures. The whole process is a non-linear spatio-temporal transformation process.
Firstly, the researchers regard an infrared image as an image corresponding to a specific temperature difference during the thermal equilibrium process. As the thermal progresses, the infrared image changes accordingly, forming a series of image sequences. To simulate this process, a dynamic filter generation component has been developed. This component generates different filters, which are understood as different temperature difference conditions in the thermal equilibrium process, and the filtered image sequences can be regarded as the corresponding infrared image sequences under a specific temperature difference. Second, by constructing a 3D deep learning network, the spatio-temporal correlation of the infrared image sequences can be learned. Again, based on the idea of Zero-DCE, the spatio-temporal correlation of the infrared image sequences is used to construct a pixel-level and high-order curve. Finally, the enhanced infrared images are predicted based on this curve.
Future work can focus on addressing issues related to image noise, jitter, and other factors that arise during the enhancement process, while also paying attention to emerging ideas and methods for image enhancement in the context of lightweight models and model distillation.
Frontiers of Computer Science
Experimental study
Not applicable
Unsupervised lightweight 3D convolutional network for enhanced infrared imaging in wearable devices
15-Jan-2026