From Grid to Ring: A New Convolutional Paradigm for Arbitrary-Scale UAV Thermal Super-Resolution

Abstract

Arbitrary-scale thermal super-resolution (SR) has emerged as a promising solution to overcome the limitations posed by the low resolution of thermal images, particularly in applications that demand flexible resolution enhancement on unmanned aerial vehicle (UAV) platforms. Due to the absence of semantic and color cues in thermal images, the model ability to extract texture representation is crucial for recovering fine details in thermal SR. Considering that thermal images captured by UAV exhibit rich and ring-like texture structures at object boundaries, we introduce Ring Convolution (RingConv), a novel ring-shaped convolution specifically designed to capture such peripheral and texture-dominant patterns. Unlike conventional convolutions that treat all spatial positions equally, RingConv selectively emphasizes precise texture regions to better enhance structural fidelity. Built upon RingConv, we further propose RingSRNet, a new backbone tailored for arbitrary-scale thermal SR that achieves accurate image reconstruction, eliminating the need for complex attention modules. Moreover, a thermal SR dataset (TSR100K) is constructed with a thermal imaging-equipped UAV, comprising 100,000 images collected across a wide range of real-world environments. Finally, extensive experiments conducted on the TSR100K and public datasets demonstrate that RingSRNet consistently outperforms state-of-the-art methods across various scales, delivering superior results in both visual quality and quantitative metrics.