Research On Image Classification Represented By Histogram Features

With the development of digital technology, the reduction of hardware cost, the popularization of intelligent devices and the evolution of Web, people can share their photograph on the Internet more easily. How to classify images more efficiently and effectively is becoming an urgent problem. Many image classification methods depend on the feature representation of image and histogram feature is one of the most important feature types in image classification. However, in this paper, we figure out there are several issues should be clarified in histogram feature based image classification and following are main contributions:1. Directly calculating distance on distance on histogram features would mislead the distance based classification on inappropriate quantization cases; we proposed a non-Mahalanobis metric based on Expected Hitting Time, which considers high-order relationship between histogram features. Besides we provide algorithm LED and its speed-up edition LED-SGD. At last, experiment results reveal the superiority of LED.2. Image ambiguity becomes obvious and intractable when images are represented by histogram features, we introduced three popular multi-label algorithms and two multi-label algorithms designed for histogram features representation. We compare the performance, trainTime, testTime and modelSize of these algorithms on image datasets.3. Based on the comparisons made above, we focus on histogram feature based image classification with resource budgets. We developed a multi-label image classification system. With collection of training dataset, we obtained the corresponding labels by crowdsourcing. Different type of histogram features and refined multi-label methods were adopted in the system. At last, the system got good performance on accuracy, memory consuming, responding time and was authorized for commercial use.