In this master thesis, we study the problem of structural change analysis and understanding from two multi-temporal very high resolution optical images. We propose a new general approach, based on features classes and features distribution learning, for an efficient detection of areas of change. The major contributions of this work are at two levels: from an application point of view, we borrow some ideas from the field of object recognition, and adapt them to the question of change/non change recognition ; from a methodological point of view, we introduce a new model and framework based on probabilistic modeling. In the present manuscript, after a short introduction in which we state the problem we are dealing with, we will give a brief review of the existing change detection methods, introduce some works on related research topics. The following chapters (3,4) then present in details our methodology, implementation, and results: 1. In chapter 3, we propose a structural change detection method based on invariant appearance features. We first segment the images into structurally consistent regions. For this purpose, we build a relation graph based on straight-lines extracted from the image, and partition this graph using Normalized Cut. Then, SIFT-like features are computed at keypoints and clustered to form a ``dictionary'' of ``words'' (characterized by the SIFT cluster centers). They are used as the basic elements to capture the structure of each region. Histograms of the ``words'' in each region are created. By estimating the distance between the histograms associated to pairwise regions taken from each of the two images, the structural change is quantified. 2. In chapter 4, we introduce an MRF based approach and label patches of the images as change/non-change with a MAP estimation framework. The appearance features used in chapter 3 are now completed by shape features, characterized by straight-line direction histograms. Following a supervised approach, we estimate the parameters of the distribution of these features for change and no-change areas separately. The density functions are then computed from these distributions. To complete the model, we define the prior term with a Gibbs model.
修改评论