Statistical Model Based Digital Image Retrieval and Fragile Watermarking

Abstract

The objective of this thesis is to acquire abstract image features through statistical modelling in the wavelet domain and then based on the extracted image features, develop an effective content-based image retreival (CBIR) system and a fragile watermarking scheme. In this thesis, we first present a statistical modelling of images in the wavelet domain through a Gaussian mixture model (GMM) and a generalized Gaussian mixture model (GGMM). An Expectation Maximization (EM) algorithm is developed to help estimate the model parameters. A novel similarity measure based on the Kullback-Leibler divergence is also developed to calculate the distance of two distinct model distributions. We then apply the statistical modelling to two application areas: image retrieval and fragile watermarking. In image retrieval, the model parameters are employed as image features to compose the indexing feature space, while the feature distance of two compared images is computed using the novel similarity measure. The new image retrieval method has a better retrieval performance than most conventional methods. In fragile watermarking, the model parameters are utilized for the watermark embedding. The new watermarking scheme achieves a virtually imperceptible embedding of watermarks because it modifies only a few image data and embeds watermarks at image texture edges. A multiscale embedding of fragile watermarks is given to enhance the embeddability rate and on the other hand, to constitute a semi-fragile approach.