SUMMARY
Efficient multimedia image extraction with high precision compatible with diverse image datasets is an implicit requirement of current image retrieval systems. In this paper, a multimedia image descriptor is introduced to achieve high performance along with high accuracy. For this, Histograms of Oriented Gradients (HOG) are extracted from a dense grid partitioned image by taking edge intensity based orientation histograms as primitive feature vectors. We depleted these massive redundant candidates to linearly uncorrelated variables by applying orthogonal transformation to achieve Principal Components (PC) where succeeding component’s constraint dependent orthogonal variance based local descriptors are compact and robust to deformation. A distinctness of our proposed approach is the selection of a single coefficient having largest variance as image descriptor out of returned dimensionally reduced vectors which results in higher performance and less space and time consumption. Supervised learning using Support Vector Machine (SVM) is then applied on non-probabilistic binary linear classification of images. The experimental results show higher precision, low memory consumption and sufficient performance gain.