Ανάπτυξη αποδοτικών παραμετρικών τεχνικών αντιστοίχισης εικόνων με εφαρμογή στην υπολογιστική όραση

Abstract

Computer Vision has been recently one of the most active research areas in computer society, since it constitutes the fusion of Artificial Intelligence and Image Processing areas and aims at developing smart systems in order to recover critical information from real images. Many modern computermachine vision and medical imaging applications, such as Super-Resolution, 3D Teleconference, Robot Navigation, Satellite Imaging and MRI-based diagnosis require the solution of the well known image registration problem. Image registration consist in finding correspondences between two or more images, which are projections of the same (usually physical) scene. Sampling of digital images, perspective projection of the scene through pinhole camera model into the image plane and camera or scene motion are some of the factors that make the problem, in a wide sense, too difficult. The majority of image registration algorithms recommend parametric techniques. These techniques adopt a specific parametric transformation model, which is applied to one image, thus providing an approach of the other one. The succes of this approach is quantified through a similarity criterion, the optimization of which requires the optimum estimation of the parameter values. Parametric techniques register image regions or features and, based on the optimization strategy, are broadly classified into two categories; the full or exhaustive search methods and the differential or gradient based methods. In this dissertation, the Stereo Correspondence and the general problem of Image Alignment are considered . In stereo case, the goal of correspondence is the construction of the disparity map based on the reference image. The ability of a stereo algorithm to produce a disparity map with sub-pixel accuracy as well as to provide reliable results under non uniform illumination of a scene in a real application are two necessary properties. Taking these algorithm properties into consideration, a local differential algorithm is proposed which involves a new similarity criterion, the Enhanced Correlation Coefficient (ECC). This criterion is invariant to any linear photometric distortions and results from the incorporation of a single parameter model (1D interpolation kernel) into the classical correlation coefficient, defining thus a continuous objective function. Although the objective function is nonlinear in translation parameter, its maximization results in a closed form solution, saving thus much computational burden. The proposed algorithm provides accurate results even under non- linear photometric distortions