Effectiveness of Super-Resolution Methods in Surveillance Systems: A Comparative Study on Satellite, Aerial, and Roadside Images

Recent advancements in deep learning have facilitated the development of numerous super-resolution (SR) techniques. This study investigates the effectiveness of these techniques in enhancing image fidelity for various surveillance applications. Focusing on three distinct scenarios with varying capture distances (satellite, aerial, and roadside) and their inherent image characteristics, the study compares the traditional bicubic interpolation with several prominent SR methods (SRCNN, FSRCNN, SRResNet, EDSR). Findings demonstrate superior performance across all SR methods, with Enhanced Deep Super-Resolution (EDSR) achieving the most significant improvement. Furthermore, the research explores the impact of training data specificity on SR efficacy. Models trained on scenario-specific datasets consistently outperform those trained on mixed or general-purpose datasets. This work underscores the potential of deep learning SR for surveillance image quality enhancement, highlighting the importance of tailored training data selection for optimal performance across varied surveillance contexts.