A multi-scale feature enhancement and context-aware convolutional network for small object detection in remote sensing images

Abstract

In remote sensing imagery, detecting extremely small objects is inherently challenging due to severe scale imbalance, sparse pixel representation, and complex background interference. In high-resolution aerial scenes, targets often occupy only a few pixels, which weakens feature responses and leads to unstable optimization. Although multi-scale detection architectures partially alleviate this issue, they often lack mechanisms for structural enhancement and scale-aware supervision. To address these challenges, CEMF-Net is proposed, a unified detection framework that integrates frequency-guided multi-scale modeling, context-selective feature modulation, and scale-consistent label assignment. By enhancing high-frequency structural cues and incorporating scale alignment into the supervision process, the proposed framework improves feature representation and localization stability for tiny objects in complex aerial environments. Extensive experiments on AI-TOD, DOTA-v1.5, and VisDrone demonstrate consistent performance gains across diverse benchmarks. On AI-TOD, CEMF-Net achieves 67.3% [email protected] and 54.6% $A{P}_{small}$, highlighting its effectiveness for detecting extremely small objects. These results demonstrate the effectiveness of CEMF-Net as a unified framework for remote sensing small object detection, with potential value for practical applications such as UAV traffic monitoring, maritime surveillance, and emergency response.