AeroTinyNet: an attention-enhanced framework for tiny object detection in remote sensing imagery

Detecting tiny and small objects in remote sensing (RS) imagery has long been a challenging task due to their minimal spatial information, weak feature representations, and dense distributions across complex backgrounds. Despite numerous efforts devoted, mainstream detectors still underperform in such scenarios. To bridge this gap, we introduce AeroTinyNet, a detection framework for tiny and small object detection in various RS scenarios. The core design of AeroTinyNet is a spatial-channel hybrid attention (SCHA) mechanism. It models attention along the channel and spatial dimensions, where both dimensions integrate global and local contextual features to enhance the tiny object representation and suppress background interference. Moreover, SCHA uses lightweight pooling and convolution operations, allowing seamless integration into different stages of the detection pipeline for feature refinement. Comprehensive experiments conducted on the public RS dataset AI-TOD demonstrate that AeroTinyNet improves AP and (Formula presented) (Formula presented) by 4.0% and 6.5% over the best compared method, respectively, and VisDrone shows that our model achieves a 1.7% improvement in (Formula presented) (Formula presented) over recent high-performing detectors. Evaluations on DIOR and RS-STOD benchmark dataset further validate its favorable detection performance in diverse RS scenarios. These results indicate that the proposed network provides a competitive and effective approach for tiny object detection on the evaluated RS benchmarks.