Applying Gaussian mixture model on Discrete Cosine features for image segmentation and classification

Abstract

Content-based image retrieval (CBIR) is the process of searching digital images in a large database based on features, such as colour, texture and shape (so-called visual content) of a given query image. Consequently, retrieved images are the most similar in content to the query image. One effective approach is to segment an image into regions (i.e. clusters) of similar colour and texture to capture its visual content. This paper presents a study that applies an adaptive Expectation-Maximization algorithm on Gaussian Mixture Model (EM/GMM) to segment an image according to local colour and texture features extracted from Discrete Cosine Transform coefficients (DCT). The EM algorithm determines rather than imposes the effective number of clusters from the image's content. This paper evaluates the effectiveness of our method by conducting a number of image classification experiments using the k-nearest neighbor (k-NN) classifier. The experiments have shown a clearly marked improvement in image retrieval accuracy of using EM/GMM over the k-means algorithm. The paper is intended to demonstrate the effectiveness of adaptive GMM in segmenting an image and capturing regions of similar colour and texture within an image.