An unsupervised change detection technique robust to registration noise

In this paper, we propose a technique for reducing the effects of registration noise in unsupervised change-detection. Such a technique represents a significant improvement of the approach proposed in [1]. It is composed of three main phases. The first phase aims at identifying the direction of the residual misregistration between multitemporal images by an iterative procedure applied to the 2-dimensional spatial domain of images. The second phase, given the direction of misregistration detected in the previous one, estimates the distribution of registration noise in the module-phase (M-P) domain of the difference image. Finally, the third phase generates the change-detection map by taking into account the estimated registration-noise distribution. Experimental results, obtained on a real multitemporal data set, confirm the effectiveness of the proposed approach.