A SIAMESE TRANSFORMER ARCHITECTURE WITH LOCAL WINDOW AND GLOBAL DEFORMABLE ATTENTION FOR CHANGE DETECTION IN IMAGES

Abstract

The task of change detection in images is addressed using the example of processing aerospace monitoring data of the Earth’s surface. To solve this task, a novel siamese deep neural network model called X-ChangeNet is proposed, based on a hierarchical transformer architecture. The model introduces a comprehensive mechanism for matching multi-temporal features, consisting of three key modules that sequentially identify changes from the local to the global level. These modules include: a multi-scale pairwise correlation unit, which captures basic structural changes through pairwise concatenation and multi-scale grouped convolution; a multi-scale window transformer block, designed to extract complex local and regional dependencies via multi-scale window attention; deformable patch transformer block with deformable attention, enabling the model to capture global context and focus on structural changes with significantly lower computational cost compared to traditional global self-attention. Comparative experiments were conducted using standard aerospace datasets LEVIR-CD and CDD Dataset, evaluating the proposed model against existing popular approaches. The results show that X-ChangeNet achieves high and competitive accuracy, outperforming most state-of-the-art models while using significantly fewer trainable parameters. Specifically, the model achieved an F1-Score of 91.91% on LEVIR-CD and 97.81% on the CDD Dataset, with only 5.8 million parameters. These results demonstrate the potential of the proposed model for effective change detection across various application scenarios.