Deep Image Translation With an Affinity-Based Change Prior for Unsupervised Multimodal Change Detection

Image translation with convolutional neural networks has recently been used as an approach to multimodal change detection. Existing approaches train the by exploiting supervised information of areas, which, however, is not always available. A main challenge in unsupervised problem setting avoid that pixels affect learning function. We propose two new network architectures trained loss functions weighted priors reduce impact on objective. The prior derived fashion from relational pixel captured domain-specific affinity matrices. Specifically, we use vertex degrees associated absolute difference matrix and demonstrate their utility combination cycle consistency adversarial training. proposed are compared state-of-the-art algorithms. Experiments conducted three real datasets show effectiveness our methodology.