Skip to content
Open access

PDQ-DETR: Parallel Multiscale Perception With Density-Guided Dynamic Query Detection Transformer for Small-Object Detection in Remote Sensing Imagery

2026 · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · Vol 19, pp. 35529-35548 · 0 citations · 58 references

Abstract

Detection Transformer (DETR) and its variants have made significant progress in small-object detection in remote sensing through end-to-end set prediction and global context modeling. However, existing DETR-based methods struggle to simultaneously accommodate the distinct feature-modeling requirements of small objects in satellite and uncrewed aerial vehicle (UAV) imagery and the different query-allocation requirements of sparse, medium-density, and dense small-object scenes. Specifically, small objects in satellite imagery typically occupy only a few pixels and provide limited fine-grained cues, such as textures and edges. These cues are further attenuated by successive downsampling, making it difficult for deep features to preserve sufficient small-object details. Small objects in UAV imagery, by contrast, exhibit more pronounced scale variations. Moreover, the number and spatial density of small objects differ substantially among the three scene types in both forms of imagery. To address these challenges, we propose parallel multiscale perception with density-guided dynamic query detection Transformer (PDQ-DETR) for small-object detection in satellite and UAV imagery. First, a Parallel Multiscale Dual-Gated Feature Fusion module extracts features for small objects at different scales and selectively fuses shallow high-resolution features with deep features, thereby preserving small-object details while adapting to small-object scale variations. Second, a density-guided dynamic query mechanism uses density features and a location map to highlight potential small-object regions and suppress background responses. It dynamically adjusts the number of queries according to the number and spatial density of small objects in each image and guides initial query position selection. Finally, a density-adaptive gaussian map generation strategy provides a supervision signal with more compact responses in dense regions for location-map learning, reducing the interference of overlapping responses from adjacent small objects on initial query position selection. PDQ-DETR achieves $\text{mAP}$ values of 39.4%, 34.2%, and 47.9% on VisDrone, AI-TOD v2, and DOTA 2.0, respectively, ranking among the leading methods in comparisons with existing advanced approaches. These results validate its small-object detection performance in both satellite and UAV remote sensing scenarios.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.