A study of subspace mixture models with different classifiers for very large object classification

Since Gaussian Mixture Models (GMM) captures complex densities of the data and has become one of the most significant methods for clustering in unsupervised context; we study and explore the idea of mixture models for image categorization. In this regard, we first segment all image categories in hybrid color space (HCbCr - LUV) to identify the color homogeneity between the neighboring pixels and then k-means technique is applied for partitioning image pixels into its coordinated clusters. Further, transformation matrix for each of the clusters is obtained by applying subspace methods such as Principal Component Analysis (PCA) & Fisher´s Linear Discriminant (FLD) to all segmented classes. These clusters are viewed as mixture of several Gaussian classes (latent variables) and Expectation Maximization (EM) algorithm is applied to these Gaussian mixtures giving best maximum likelihood estimators and thereby obtaining highly discriminative features in reduced feature space. For subsequent classification, we use diverse Distance Measures (DM) and Probabilistic Neural Network (PNN). The results obtained is evident that the proposed model exhibits highly discriminative image representation that leads to the improved classification rates to the state-of-the-art on standard benchmark datasets such as Caltech-101 & Caltech-256.