SPEFormer: A Synergistic Perception-Enhanced Transformer With Density-Guided Dynamic Queries for Remote Sensing Small-Object Detection

Abstract

Small-object detection in remote sensing images is highly challenging due to their limited pixel representation, weak feature expression, and strong interference from complex backgrounds. Moreover, existing query-based detectors typically use a fixed number of object queries for all images, making them poorly adaptive to sparse or dense target distributions, which may lead to missed detections or redundant computations. To address these issues, we propose SPEFormer, a detection framework specifically designed for small objects in remote sensing images. SPEFormer incorporates the selective global–local perception network (SGLPNet), which combines conditional convolution, global branches, and local branches across different stages to extract high-quality multiscale features, and the adaptive synergistic attention fusion (ASAF) module to adaptively fuse cross-level features, effectively enhancing small-object responses while suppressing background interference. In addition, a density-guided dynamic query (DGDQ) mechanism dynamically allocates the number and positions of object queries based on multiscale density maps, enabling the model to handle both sparse and dense target scenarios. Experimental results on VisDrone2019, DIOR, and DOTA-v1.0 show that SPEFormer achieves competitive detection performance. In particular, SPEFormer obtains strong AP50 scores of 54.7%, 87.7%, and 65.3% on the three datasets. Code is available at: https://github.com/SLBY8/SPEFormer