Graph-Modulated Attention and Centrality-Enhanced Network for Remote Sensing Object Counting

Object counting in remote sensing images remains highly challenging due to the dense distribution and high similarity of targets. Existing Transformer-based counting models have achieved remarkable success, but they often suffer from the attention map response convergence issue, where attention maps across different regions become overly similar. To address this limitation, we propose Graph-Modulated Attention and Centrality-Enhanced Network (GACENet), a novel architecture that rethinks attention mechanisms through the lens of graph-structured diversity enhancement. Our framework introduces three groundbreaking components: the Anti-Similarity Attention Graph (ASAG) module, the Centrality-Aware Feature Enhancement (CAFE) mechanism, and a composite loss function. First, the ASAG module dynamically learns a graph that explicitly models semantic dissimilarity between regions, forcing attention heads to focus on complementary features rather than converging to redundant patterns. Second, the CAFE module is designed to mitigate weak responses to key target regions caused by attention homogeneity. It is a topological prior-driven module that identifies and amplifies critical nodes via graph centrality analysis, ensuring robust representation of key targets even under extreme density conditions. Finally, a composite loss function that combines a structural loss and a diversity loss, which encourages response variation across spatial locations. Extensive experiments on multiple remote sensing counting datasets demonstrate that GACENet significantly outperforms other state-of-the-art remote counting methods in terms of MAE and MSE.