Feature representation for large scale classification
Date of Issue2016-03-15
School of Computer Engineering
Centre for Computational Intelligence
Feature representation plays the center role for classification. The extraction of knowledge in data, whether through pre-defined functions or procedures, or through learned projection matrices or neural networks, is crucial for the success of a large scale classification system. In this dissertation, three different feature embedding methods are considered to improve the efficiency and/or effectiveness of the classification of three different types of data. Particularly, we propose linear regression support vector machine (LR-SVM) for bag-of-visual-words (BOV) data, feature pair selection (FPS) for data with multiplicative correlation between its features, and a multi-view multi-instance convolution neural network based system for raw image data with multiple objects. Experimental results demonstrate that our methods are efficient to handle big data, and they well exploit their respective data to achieve high classification accuracy.