DIMD-DETR: DDQ-DETR With Improved Metric Space for End-to-End Object Detector on Remote Sensing Aircrafts

Abstract

Aircraft target detection in remote sensing images faces numerous challenges, including target size variations, low resolution, and complex backgrounds. To address these challenges, an enhanced end-to-end aircraft detection framework (DIMD-DETR) is developed based on an improved metric space. Initially, a bilayer targeted prediction method is proposed to strengthen gradient interaction across decoder layers, thereby enhancing detection accuracy and sensitivity in complex scenarios. The pyramid structure and self-attention mechanism from pyramid vision transformer V2 are incorporated to enable effective joint learning of both global and local features, which significantly boosts performance for low-resolution targets. To further enhance the model's generalization capabilities, an aircraft-specific data augmentation strategy is meticulously devised, thereby improving the model's adaptability to variations in scale and appearance. In addition, a metric-space-based loss function is developed to optimize the collaborative effects of the modular architecture, enhancing detection performance in complex backgrounds and under varying target conditions. Finally, a dynamic learning rate scheduling strategy is proposed to balance rapid convergence with global exploration, thereby elevating the model's robustness in challenging environments. Compared to current popular networks, our model demonstrated superior detection performance with fewer parameters.