Research On Key Technologies Of Image Object Recognition Based On Information Geometry

Image object recognition is an important area of study in the field of computer vision and image analysis. In this article, three main important tasks of object recognition are researched: image denoising and rigid point set registration and object classification based on shape. The main work and research results are as follows:1. The current algorithm for Gaussian and Poisson denoising using only the pixel gray value to determine the similarity weights between pixels is not accurate enough. The information geometry principles applied to the image denoising, using the difference pixel neighborhood statistics to determine the similarity weights, then a Gaussian denoising and a Poisson denoising algorithms are proposed based on geodesic. For Gaussian noise, based on Gaussian statistical modeling, the geodesic distance between two points on the manifold represents the similarity between pixels, Geodesic represents the differences between two pixel neighborhoods' average gray intensity and richness of detail. For Poisson noise, using non-local maximum likelihood estimation of the parameters of the Poisson distribution model, which will transformed pixels onto statistical manifolds, geodesic distance between two points is calculated to determinate the weights of similarity between pixels, the non-local means algorithm is used for Poisson denoising. The experimental results shows that Gaussian denoising algorithm based on geodesic on a flat area can achieve better denoising effect, and can effectively preserve edge details. In terms of Poisson noise, Poisson denoising algorithm based on geodesic has better performance than the non-local means algorithm.2. During robust points set registration algorithm based on hybrid model, build GMM of point sets, the use of similarity between the two models to determine the registration parameters. According to the basic idea of the information geometry, using of KL(Kullback-Leibler) divergence distance between model express their similarity, which can be seen as a point on the manifold projection to its submanifolds. Then a robust rigid point sets registration algorithm based on Kullback-Leibler is proposed, which has a clear geometric interpretation. By finding the minimum distance KL divergence between the two GMM which determines the final registration parameters. The experimental shows that robust rigid point set registration algorithm based on Kullback-Leibler has good robust to noise and outliers.3. During target classification, based on the Peter's basis research work, we have a preliminary study based on statistics manifold of target classification. First, the use of model learning methods, selecting the same amount key points automatically based on the minimum description length. Then, by modeling of GMM, the key point set of target shape is transformed into a statistical manifold, the target point set is point on manifold, the geodesic distance between two points represents the magnitude of deformation between shapes. Experiments show the feasibility of the proposed algorithm in terms of target classification.