Statistical modeling for automatic image indexing and retrieval

Abstract

In this paper, a statistical modeling algorithm is developed to achieve automatic detection of object classes and image concepts via partial similarity matching. For a given image, its statistical image model is automatically learned by using a finite mixture model to approximate the distribution of its image pixels in the 10-dimensional feature space. Such statistical image modeling process can also achieve automatic image segmentation implicitly. To achieve more precise matching between the mixture components and the local distributions of the relevant image pixels, an adaptive EM algorithm is developed to simultaneously select the model structure (i.e., the optimal number of mixture components) and estimate the model parameters (i.e., locations and statistical properties of the mixture components) according to the local distributions of the relevant image pixels. For a given image concept or object class of interest, its statistical concept model is automatically learned from the statistical image models for the labeled training images. Finally, similarity matching for automatic detection of object classes and image concepts is treated as a partial model matching problem, i.e., matching between the statistical image model for a given test image and all the statistical concept models for the object classes and image concepts of interest. Our experimental results have demonstrated that our statistical modeling algorithm can achieve very competitive results on both automatic image segmentation and classification.

Content-Based Image Retrieval: Survey

“…Zhang et al [7] presented a feature of 10 dimensional in length which consists of 3 colour R, G, and B, 3 colour deviations of R, G, and B using 5x5 block, 2 gradients of luminance, and 2 locations (x, y). After feature extraction, a finite mixture model was calculated using EM algorithm.…”

“…For a set of hash functions, mapping all the objects data to binary codes as: (7) The next step can made Approximate Nearest Neighbor (ANN) search in Hamming space, where the computation cost is reduced. Hamming distance between two binary codes c i and c j as follows:…”