An Ensemble Model to Extract Discriminative Features for Semantic Image Classification in Large Datasets

Abstract

AbstractAn efficient image representation and extracting discriminative features in compressed domain is attracting researchers in computer vision and pattern recognition to develop efficient algorithms to classify and annotate images in large datasets. In this paper, an ensemble model combining DCT, Wavelet and HOG is developed to represent an image in compresses domain. DCT is useful to find low frequency coefficients of an image which express visual features, further subjected to multi-resolution analysis using wavelets with an advantage of developing robust and geometrically invariant structured object visual features through spectral analysis and finally PCA is used to map lower dimensional feature space with the transformational matrix in which a set of observed data is infused highlighting the edge orientation along with histogram of oriented gradients as its feature vectors. For classification purpose, different distance measure techniques and machine learning algorithm is used to obtain average classification rate. Proposed ensemble model is demonstrated on Caltech-101 and Caltech-256 datasets and compared the results with several benchmarking techniques in literature. KeywordsDiscrete cosine transforms (DCT)Distance metrics (DM)Histogram of oriented gradients (HOG)Image classificationLogistic regression (LR)Wavelets