DualFocus-CapNet: A Dual-Stream Network for Real Change and Interscale Relationship-Aware Change Captioning in Remote Sensing Images

Remote sensing image change captioning (RSICC) aims to generate textual descriptions of changes between bitemporal images. However, accurately describing fine-grained changes while capturing interscale relationships as well as distinguishing real changes from spurious changes (e.g., illumination, seasonal variations) are still major challenges for current methods. To address these issues, we propose DualFocus-CapNet, a novel model tailored for RSICC. DualFocus-CapNet employs a dual-stream architecture, where each stream is dedicated to processing a distinct pair of bitemporal features. Crucially, we introduce a scale-wise progressive convolution (ScalePro Conv) that employs a progressive scale-specific approach to decompose remote sensing features into pixel-level variations, regional continuities, and linear structures. Unlike conventional parallel multiscale processing methods, ScalePro Conv adopts a serial progressive structure to establish interscale relationships, thereby avoiding the fragmentation of feature information. Then, the bi-directional difference guided transformer (BDiGTrans) is proposed to eliminate interference from spurious changes by dynamically masking invariant regions and extracting bidirectional differential features. Furthermore, the cross-temporal adaptive fusion module (CTAF) is introduced to dynamically balance bitemporal features using learnable gating to enhance change discrimination and robust caption generation. Comprehensive experiments on the benchmark datasets LEVIR-CC and WHU-CDC show that our DualFocus-CapNet surpasses state-of-the-art change captioning methods in various evaluation metrics.