Multi-stage feature purification for robust object detection in haze-degraded aerial images

Robust object detection in haze-degraded unmanned aerial vehicle imagery remains challenging because atmospheric corruption progressively weakens neural representations throughout the detection pipeline, while the aerial viewpoint further intensifies scale variation and small-object ambiguity. Existing solutions usually depend on image dehazing, external priors, or heavy multimodal systems, which often suffer from object-restoration mismatch, limited robustness, or excessive complexity. In this paper, we present the Multi-stage Purified Representation Network (MPRNet), a lightweight end-to-end detection framework that improves degraded-scene representation learning through architecture-level feature purification. The proposed network performs semantic purification in deep representations and multi-scale purification during cross-scale feature interaction, thereby reducing haze-related noise propagation and improving the recovery of weak small-object cues. Experimental results on three hazy-scene benchmarks show that MPRNet achieves competitive detection accuracy and consistent performance across both synthetic and real hazy scenes. On the main benchmark, MPRNet achieves 55.3% mean average precision, improving the strongest competing method by 3.3 percentage points while requiring only 10.3 million parameters and 30.9 billion floating-point operations. Edge-side evaluation on Jetson Xavier NX further reports 33.5-36.1 milliseconds of inference latency together with low memory occupancy, providing supplementary evidence that the proposed model offers a favorable accuracy-efficiency trade-off and is suitable for edge deployment.