Unsupervised Global Difference Modeling for Image Change Detection

In change detection, impact of non-intrinsic changes such as those caused by illumination, season, and viewing angle variances are common in practice but also a great challenge for change detection methods. In this paper, we propose a novel unsupervised image change detection method by modeling global difference information to deal with such non-intrinsic changes. Comparing global features can mitigate the impact of them due to the global consistency of them in the same scene at the same time. But global modeling for change detection also faces the challenges of feature learning with limited data and difficulty in generating pixel-wise changed regions. To overcome the challenges, firstly, we use a backbone network to capture the global features of bi-temporal images. Then an energy function is designed with a masked difference between the two features and a margin-aware constraint in order to align the global features and meanwhile maintain detail information. To train the network with only two images, we propose an adversarial learning method by introducing a generalization network that consecutively generates two images that can minimize the energy. Then a new loss function is derived to alternately train the feature learning network and generalization network. Secondly, after learning with bi-temporal images, it is also important to generate the pixel-wise changed regions. Then we design a difference mapping method that maps the changed regions from global difference. Experiments on different types of data by comparing with both supervised and unsupervised methods demonstrate the effectiveness of the proposed method