Anti-drift Preserving Network with UAV ImageSuper-resolution

UAV Image Super-Resolution has become increasingly prominent in recent times, owing to the widespread use of atmosphere monitoring, tracking, and surveillance, which are critical areas of interest. The primary focus of this work revolves around enhancing the resolution of UAV images, utilizing an anti-drift network as a key component. Past endeavours in this domain faced difficult hurdles, primarily attributed to the dynamic characteristics of the target data. This work accelerated the process by conducting experiments on benchmark datasets such as VisDrone and UAVid. These datasets showcase densely populated scenes featuring objects such as people and vehicles, which often undergo significant variations in density; identifying where these density changes occur over varying time intervals presents a complex challenge. To tackle these challenges, we propose a Context-reasoning Swin-transformer Graph Attention Network that employs the Swin transformer,semantic reasoning and context-aware graph attention to modulate context descriptors by semantic reasoning along spatial-channel interaction tensor to extract more global contextual information for UAV SR. Our proposed approach has been thoroughly evaluated through extensive experimental results and ablation studies, showcasing its effectiveness and efficiency. It consistently surpasses state-of-the-art methods by a notable margin, achieving more than 0.5 dB improvements.