Super-Resolution-Assisted Feature Refined Extraction for Small Objects in Remote Sensing Images

Despite achieving impressive results in object detection in natural scenes, the task of object detection in remote sensing images is still full of challenges due to the large number of small objects in remote sensing images caused by the dense object distribution, complex backgrounds, and diverse scale variations. We propose a Super-Resolution-Assisted Feature Refined Extraction (SRRE) approach to address the difficulties of detecting small objects. Firstly, we employ a deeper level of feature fusion to effectively harness deep semantic information and shallow detailed information. Secondly, in the feature extraction process, a Feature Refined Extraction Module (FREM) is introduced to capture a wider range of contextual information, enhancing the global perceptual capability of features. Lastly, we introduce Super-Resolution (SR) branches at various feature layers to better integrate local textures and contextual information. We compared our method against commonly used approaches in remote sensing image object detection, including state-of-the-art (SOTA) methods. Our approach outperforms these methods and achieves superior results on the DOTA-v1.0, DIOR, and SODA-A datasets.