Edge‐guided Dual Alignment Framework for Semi‐Supervised Scene Change Detection

ABSTRACT Scene Change Detection (SCD) aims to identify changes in the same street scene captured at different times, and is crucial for applications such as automatic map updating. However, most existing methods rely on costly pixel‐level labels and inadequately model change boundaries, which leads to false positives (FPs) and false negatives (FNs) in boundary regions. To utilize unlabeled data to achieve accurate change boundary segmentation, this article proposes an Edge‐guided Dual Alignment (EDA) framework for semi‐supervised change detection (SSCD). First, an edge‐enhanced dilated spatial pyramid (EDSP) module is designed. This module integrates prior boundary information with multi‐scale contextual features to enhance the perception and localization of change boundaries. Second, during the semi‐supervised learning (SSL) process, a pixel‐wise prediction alignment (PPA) strategy based on consistency regularization is introduced to learn prediction invariance under data augmentations. In addition, an edge‐guided feature alignment (EFA) strategy is proposed for accurate change boundary segmentation, based on the EDSP module. Under the guidance of pseudo‐label boundary priors, boundary features extracted by the EDSP module under different perturbations are aligned, thereby enhancing the consistency and stability of change boundary features. Experimental results on the VL‐CMU‐CD and PCD datasets demonstrate that the proposed EDA framework performs better than the selected baselines using only limited amounts of labeled data, particularly excelling at detecting change boundaries.