Research On Image Statistical Part-Based Model

A variety of multimedia application calls for a higher detection precision of Content-Based Image Retrieval (CBIR). Unfortunately, the performance of CBIR is seriously restricted by feature extraction and similarity measure, which are two key factors in CBIR. The statistical part-based model which is proposed by Digital Vedio and Multimedia Lab of Columbia University can solve the problem effectively. In the theme we depend on the model and develop the model with a new method which is based on Attributed Relational Graph (ARG). Moreover, we give the method a realization with MATLAB and C++ programming. In the theme we mainly work in following aspects.Firstly, in the aspect of corner detection, in Harris corner detection we use a new corner response function instead of the old one which could lead to the cluster of corner. The experiment result shows that the new one has a better performance in corner detection.Secondly, in feature extraction, we give a realization of color space transformation by programming which can make transformation from RGB to HSV to be true. This is because compared with RGB HSV is more consistent with human visual perception. The experiment result shows that the performance of retrieval under HSV is more precise than RGB.Thirdly, in image representation, we use a new image representation method ARG which can not only represent image attributes but also have a description of topology structure of an image. It is better than bag of pixels representation which can not reflect topology structure and the experiment result gives a conclusion of this.At last, in similarity measure, we propose a new method which is called similarity method based on ARG instead of the old one. The reason for the change is based on two reasons: one is that Earth Mover's Distance (EMD) used before can lead to a higher rate of error judgment; the other is EMD method is not suitable for the image representation of ARG used in our theme. We can have a conclusion from the experiment result that the method we proposed makes retrieval accuracy have an improvement.It is clearly that the method we used is better than the method used in primitive model in improving the performance of image similarity retrieval. Our research makes the statistical part-based model more perfect and useful. Besides, it can provide well theoretical guidance and practical application value to the research of CBIR.