Context Synergy and Multitask Interaction Network for Semantic Change Detection in High-Resolution Remote Sensing Images

Le résumé fourni par la source

High-resolution remote sensing images (RSIs) are usually characterized by complex backgrounds and significant intra-class and inter-class scale variance, making long-range contextual information (LRCI) essential for semantic change detection (SCD). However, image cropping operations tend to truncate the LRCI inherent in RSIs, that would compromise the integrity of geometric features and disrupts the dependency of semantic correlations. To address these limitations, we propose a context synergy and multitask interaction network (CSMINet). Specifically, an extra context synergy branch is introduced to aggregate LRCI from broader image regions, compensating for the loss of contextual correlations in cropped images. CSMINet comprises three pivotal modules: the context synergy guidance module (CSGM), the local-dominant context-synergy fusion module (LCFM), and the decoder feature-aligned interaction module (DFIM). The CSGM is utilized to filter valuable LRCI that enhances contextual associations. The LCFM achieves fine-grained feature fusion, which is dominated by local features and synergized with long-range contextual information. The DFIM facilitates efficient information interaction across subtasks, bridging the gap between semantic representations and change targets. Extensive testing on two benchmark datasets validates the effectiveness of our proposed method, outperforming six comparative methods across various evaluation metrics and achieving Fscd of 62.13% and 67.83%, respectively.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.