Asymmetric image translation network for unsupervised heterogeneous change detection
Jie Lan, Tao Zhan · 2025
In the remote sensing community, change detection (CD) based on heterogeneous images plays an important role in monitoring Earth’s dynamic changes. However, the differences in visual appearance and statistical characteristics between heterogeneous images make it difficult to identify changes through direct comparison. To overcome this challenge, we propose an unsupervised asymmetric image translation network (AITNet) for CD in heterogeneous images. First, an asymmetric multi-level detail-preserving generator is developed to generate high-fidelity translated images by integrating multiscale structural information and deep semantic information. Then, domain-specific difference information can be captured by comparing the semantic feature similarity between images with the same modality. Following that, an adaptive fusion strategy is designed to integrate complementary difference information from different modalities, ultimately generating a precise change map through threshold analysis. Experimental results over two typical heterogeneous datasets demonstrate that the proposed framework outperforms other state-of-the-art CD methods, verifying its effectiveness and superiority.