A coupling translation network for change detection in heterogeneous images

Based on the images acquired through different sensors, change detection is much more challenging than those based on homogeneous images. The main reason behind it is that the heterogeneous image-pair cannot be directly compared in original observation space due to their distinct statistical properties. In order to detect the changes, we establish a coupling variational autoencoder (VAE) to transform the heterogeneous images into a shared-latent space, where they have more consistent representations and hence the prior changed regions can be highlighted by direct comparison. And based on the shared space, we build coupled generative adversarial networks (GANs) associated with the coupling VAE to translate the heterogeneous images into homogeneous, from which more accurate change detection results can be obtained in their common observation spaces. The proposed framework is totally unsupervised, and the experimental results on real heterogeneous data sets demonstrate its superiority over some other existing algorithms.