Applying the extended mass-constraint EM algorithm to image retrieval

We extend the mass-constraint data clustering and vector quantization algorithm to estimate Gaussian Mixture Models (GMMs) as image features applying to the image retrieval problems. The GMM feature is an alternative method to histograms to represent data density distributions. Histograms are well known for their advantages including rotation invariance, low calculation load, and so on. The GMM maintains the rotation invariance properties; moreover, it addresses the high-dimensional problems due to which histograms usually suffer inefficiency problems. The extended mass-constraint (EMass) GMM estimation algorithm is compared with the typical Expectation–Maximization(EM) algorithm, and the deterministic annealing EM (DAEM) algorithm. The three algorithms are applied to train a GMM for a set of simulation data, and compared with the log-likelihood values. From the comparison results, we know that DAEM still has strong dependence on initial data point selection, which is the main problem we need to solve by taking advantage of the deterministic annealing methods. Thus the DAEM algorithm is not chosen to estimate GMM density functions for image retrieval. The EM and EMass algorithms are then applied to train GMMs from image RGB color features for the purpose of image retrieval. Finally the GMM features are combined with the Local Binary Pattern (LBP) features to achieve higher precision retrieval. After we compare the precision/recall curves and mean average precisions achieved by two algorithms, we conclude that the extended mass-constraint algorithm is a better solution for GMM estimation, and combining the GMM and Local Binary Pattern (LBP) provides a new promising feature for image retrieval.