Abstract:To address error accumulation caused by detection-association decoupling and instance confusion induced by dense feature fusion in multi-class aircraft change detection, a candidate-based association-reorganization end-to-end method is proposed. The method takes bitemporal candidate instances as the basic reasoning unit and constructs a three-module framework comprising candidate generation, candidate association, and candidate reorganization with joint prediction. The candidate generation module employs a shared-weight oriented object detector to extract aircraft candidates from bi-temporal images; the candidate association module fuses visual features, class probabilities, and geometric relationships, and explicitly models unchanged, disappeared, and appeared relationships via background-augmented global one-to-one assignment; the candidate reorganization and prediction module, guided by association results, unifies matched and unmatched candidates into semantically complete fused features, and jointly outputs aircraft class, change states, and oriented bounding boxes. On the ACD-v1.5 dataset, the method achieves 70.6% mean average precision (mAP), outperforming the best baseline by 3.0 percentage points. The results support practical applications including aircraft inventory at airport aprons, dynamic tracking of aerial targets, and situational awareness in key air-space regions.
