Adaptive Window Size Estimation in Unsupervised Change Detection

Many problems related to change detection require to compute image features on local windows. Such features usually combine in each pixel locations spectral values (luminance) associated with some spatial properties, such as texture features or more advanced local relationships between pixels. Therefore, as far as local windows are considered, the optimal size selection is a key point for the performance of the algorithm. This paper tackles this issue by proposing an original mean to estimate the size of local windows at each pixel. It uses a stochastic representation of the image grid. By combining some rules of stochastic calculus, we redefine image features functions on specific image grids where the changed areas are modeled. This enables to extract in each location the optimal size on which image feature function should be consistent. For validation´s sake, we propose a simple change detection approach that takes into account basic image features computed on local windows where the sizes are either manually fixed or automatically estimated, both using our approach and existing window size estimation techniques. In addition, some comparisons with state-of-the-art change detection methods are presented. This allows to validate the efficiency of our proposition and to demonstrate that even simple techniques associated with accurate image features can provide interesting results.