CAEM-DETR: Small Aerial Target Detection via Contrastive Attention-Enhanced Multidomain Fidelity Fusion

Small objects in Unmanned Aerial Vehicle (UAV) imagery, often vanishing amidst limited pixels, dynamic views, and complex backgrounds, present a critical detection challenge. State-of-the-art Transformer-based detectors, despite their strong general performance, frequently struggle to detect these small aerial targets, with their effectiveness noticeably decreasing. This paper introduces CAEM-DETR (Contrastive Attention Enhanced Multi-domain Detection Transformer), a novel framework derived from RT-DETR, specifically designed to master these intricacies and deliver significantly improved small object detection accuracy from aerial perspectives. CAEM-DETR incorporates several key innovations: (1) DLCA (Dynamic Linear Contrastive Attention), a mechanism that distinctively leverages positive and negative feature polarities, modeling their diverse interactions with linear complexity to enhance feature contrast and refine semantic representations for subsequent fusion; (2) MDFF (Multi-Domain Fidelity Fusion), a module that synergizes high-fidelity downsampling with dual-domain (spatial and frequency) analysis, integrating DLCA-enhanced polarity information to recover critical high-frequency details from a comprehensive four-dimensional perspective; (3) GRACE-Net, a backbone employing synergistic multi-attention and selective channel modulation to capture fine-grained multi-scale details; and (4) Wise-FocalerIoU, a novel composite loss function that provides robust, difficulty-aware gradient guidance for precise small object localization. Experimental results on four diverse datasets (VisDrone, AI-TOD, SIMD, and DOTA) demonstrate CAEM-DETR’s effectiveness and generalization. On the VisDrone dataset, CAEM-DETR achieves a relative improvement of 13.7% in AP and 11.6% in AP₅₀ over the baseline.