HiCD: Change Detection in Quality-Varied Images via Hierarchical Correlation Distillation

TL;DR

This work tackles change detection when bi-temporal remote sensing images differ in quality due to imaging conditions and platforms. It introduces HiCD, a teacher–student framework where a model trained on high-quality pairs guides learning on degraded pairs, augmented by hierarchical correlation distillation that transfers self-, cross-, and global correlations rather than exact features. Semantic feature distillation and change feature distillation modules enable robust representation learning and effective change localization under degradation. Extensive experiments on LEVIR-CD, BANDON, and SV-CD demonstrate state-of-the-art performance across resolution-difference, single-degradation, and multi-degradation settings, highlighting the practical impact of transferring deep priors without heavy restoration.

Abstract

Advanced change detection techniques primarily target image pairs of equal and high quality. However, variations in imaging conditions and platforms frequently lead to image pairs with distinct qualities: one image being high-quality, while the other being low-quality. These disparities in image quality present significant challenges for understanding image pairs semantically and extracting change features, ultimately resulting in a notable decline in performance. To tackle this challenge, we introduce an innovative training strategy grounded in knowledge distillation. The core idea revolves around leveraging task knowledge acquired from high-quality image pairs to guide the model's learning process when dealing with image pairs that exhibit differences in quality. Additionally, we develop a hierarchical correlation distillation approach (involving self-correlation, cross-correlation, and global correlation). This approach compels the student model to replicate the correlations inherent in the teacher model, rather than focusing solely on individual features. This ensures effective knowledge transfer while maintaining the student model's training flexibility.

Figures (8)

• Figure 1: The performance of the excellent CD methods on the multi-degradation LEVIR-CD dataset. * indicates the method trained on quality-varied image pairs, otherwise on the equal high-quality image pairs.
• Figure 2: The overview of our proposed method. (a) Knowledge distillation-based training strategy. (b) Hierarchical correlation distillation. During the training of the student, the teacher takes the corresponding high-quality image pairs as input, to generate bi-temporal representations and the change feature for guiding the student. Based on the novel feature distillation for change detection, i.e., hierarchical correlation distillation, we devise the semantic feature distillation module and the change feature distillation module, to transfer different kinds of knowledge acquired by the teacher. $\mathcal{L}_\mathrm{ce}$, $\mathcal{L}_\mathrm{sfd}$ and $\mathcal{L}_\mathrm{cfd}$ are cross-entropy loss, distillation losses of the SFD and CFD modules, respectively.
• Figure 3: The building change detection is affected by various irrelevant changes.
• Figure 4: Example of the LEVIR-CD dataset with 12 types of degradations.
• Figure 5: Qualitative results on the BANDON dataset with the setting of only resolution difference. GT, FC-SC, and FC-SD are short for the ground truth, FC-Siam-conc daudt2018fully and FC-Siam-diff daudt2018fully, respectively. Different colors, i.e., white, black, red, and green, indicate true positive, true negative, false positive, and false negative.