The deficiency is primarily attributed to traditional Convolutional Neural Network (CNNs) and Transformers employing only single-pass forward propagation, leading to inadequate utilization of feature information. Additionally, the commonly used CD information extraction module further contribute to information loss, resulting in inaccurate edge detection. To cope with this challenge, we propose a Transformer-based Conditional Generative Diffusion Change Detection (T-GDCD) approach to iteratively integrate various information and produce fine-grained CD maps.
a Spectral Harmonic Cross-Attention Transformer (SHCAT) mechanism and a Change Detail Enhancer (CDE) are introduced to inherently integrate the high-dimensional and low-dimensional CD conditions into the Denoising Diffusion Probabilistic Model (DDPM). Such integration process effectively minimizes information loss and facilitates the diffusion model in enhancing the CD map through numerous sampling iterations, contributing to the generation of high-quality CD maps. Moreover, the T-GDCD framework comprises a pioneering CD information extraction module, termed Composite Differential Integration (CDI). This module is specifically engineered to circumvent the information loss commonly associated with rudimentary subtraction techniques, enabling the extraction of enriched CD information.
