AsymCD: A change detection method for dual asymmetric scenarios

Change Detection (CD) is crucial for natural disaster assessment, urban construction management, ecological monitoring, etc. Nevertheless, most existing pixel-level change detection models assume that bi-temporal images are spatially aligned and possess consistent imaging quality, thus failing to address dual asymmetries in spatial range and information granularity in practical applications. To address these asymmetries, this paper proposes AsymCD, a change detection method for asymmetric scenarios, which comprises the Query-Adaptive Localization (QAL) module and the Cross-Granularity Structural Consistency (CGSC) module. By integrating registration into the change detection model, the method enables effective joint learning through end-to-end optimization. Specifically, the QAL module transforms image alignment into a dynamic query-guided feature anchoring process to resolve macroscopic regional misalignments caused by spatial range discrepancies. Furthermore, the CGSC module performs asymmetric feature verification using a unidirectional gating mechanism. This addresses microscopic detail mismatches caused by information granularity asymmetry, accurately filtering out false positive responses. Extensive experiments on asymmetric evaluation benchmarks constructed from LEVIR-CD, WHU-CD, and S2Looking datasets demonstrate that AsymCD achieves state-of-the-art performance. These experimental results fully validate its effectiveness and robustness in challenging asymmetric scenarios.