DAL-DETR: Dynamic Adaptive Lightweight Detection Transformer for Small Object Detection in Low-Altitude Remote Sensing Imagery

Small object detection in low-altitude remote sensing imagery plays a pivotal role in urban planning, disaster assessment, and related applications. However, existing approaches commonly suffer from excessive model complexity and substantial computational overhead, making it challenging to achieve an optimal balance between lightweight design and detection accuracy on resource-constrained edge computing devices. To address these limitations, we propose dynamic adaptive lightweight DETR (DAL-DETR). To overcome the constraints of fixed receptive fields in conventional convolutions—which struggle to simultaneously capture local details and global contextual semantics of small objects—we design a dynamic adaptive mixer backbone (DAMB). This module leverages dynamic multibranch convolutions (DMCs) to achieve adaptive fusion of multiscale receptive fields, improving multiscale feature representation while significantly reducing parameter count. To mitigate computational redundancy introduced by feature concatenation, we introduce a lightweight dynamic fusion neck (LDFN), which replaces traditional concatenation operations with deformable modulation fusion, further reducing model complexity and computational cost. In addition, we incorporate an auxiliary decoupled geometric loss (ADGL) that employs an auxiliary bounding box generation mechanism with decoupled geometric constraints, improving localization precision for small objects. Extensive experiments on multiple public datasets demonstrate that DAL-DETR achieves a 45.2% reduction in parameters and a 34.0% decrease in computational cost compared to the baseline, while still delivering consistent accuracy improvements and outperforming existing state-of-the-art detection methods.