Coordinate aware implicit neural representation for UAV small object detection

Abstract

Small object detection in unmanned aerial vehicle (UAV) imagery remains challenging due to low resolution, complex backgrounds, and limited pixel occupancy. Conventional CNN-based detectors rely on fixed receptive fields, which often fail to capture fine-grained spatial details and implicit geometric relationships. To address these limitations, this paper proposes a Coordinate-Aware Implicit Neural Representation Enhanced C2INR module that integrates coordinate-based representations into convolutional feature extraction. The Multi-Scale Coordinate Encoder (MSCE) constructs frequency-aware positional embeddings through sinusoidal encoding to enhance spatial continuity at multiple scales. The INR Feature Enhancer (IFE) further fuses encoded coordinates with visual features via lightweight MLP modulation, improving sensitivity to small-scale variations. Additionally, a Small Object Attention mechanism combines global context, local detail, and high-frequency cues to strengthen responses to tiny targets. Experiments on three UAV benchmarks—AI-TOD, UAVDT, and VisDrone—demonstrate consistent improvements over existing methods with minimal computational overhead. Further evaluation on PASCAL VOC verifies strong cross-domain generalization. These findings confirm that coordinate-aware implicit representation provides an effective and broadly applicable solution for improving spatial continuity, geometric fidelity, and localization precision in small-object detection.