A multi-level detection strategy via anomaly detection for more accurate target recognition in remote sensing images

Remote sensing object detection presents unique challenges due to cluttered backgrounds, high inter-class contextual correlation, and the presence of densely distributed small targets. To address these issues, we propose CADOR-Net, a novel object detection framework that integrates anomaly detection into the recognition pipeline through a posterior probability-based refinement mechanism. Specifically, CADOR-Net introduces a shared backbone and a convolutional autoencoder (CAE)-based anomaly detection module to jointly model the semantic consistency between objects and their surrounding backgrounds. By leveraging pixel-wise reconstruction error maps, the model suppresses false positives in semantically inconsistent regions and improves classification confidence for contextually coherent targets. We further design a Segment Pooling module to map region proposals to their corresponding error regions, enabling fine-grained anomaly-aware classification correction. Extensive experiments on four benchmark remote sensing datasets—HRSC2016, DOTA-v1.0, DOTA-v1.5, and DIOR-R—demonstrate that CADOR-Net outperforms state-of-the-art detectors in terms of mean Average Precision (mAP), while maintaining lower computational complexity and parameter count. Ablation studies confirm the scalability and robustness of the integrated anomaly detection mechanism.