Research On Image Classification Based On Outlier Detection

Multimedia technology is widely used, and Internet has become an importantpart of our life, they together bring lots of digital images. Image classification wasproposed and developed in such an environment. Assigning the correct category forimages according to its visual content can not only improve the accuracy of imageretrieval, but also be helpful to effective image organization and management. Howto classify image scientifically and effectively is an important research topic inmultimedia field.There are many kinds of digital images, including color images, gray images,infrared images and so on. We focus on the classification of gray images and colorimages in this paper. In recent years, both image feature extraction techniques andmachine learning algorithms develop rapidly. But, as the difference betweenclassification models and the disparity between visual contents of image data set,there is not a feature or a classifier which fits all image classification problems.Thus, image classification techniques based on classifier fusion and feature fusionhave been used widely. This paper adopts outlier detection into image classification.We first detect images that may be classified falsely in the initial classification, andthen offer them to the classifier in the post-processing module. Classificationperformance can be improved through complementarity of classifiers and effectiveoutlier detection. The main work of this paper is as follows.1. We implement the initial image classification based on SVM algorithm.Many kinds of image visual features are analyzed and compared, then we choosetypical combination of shape, texture and SIFT feature to represent images. Usingmulti-category SVM algorithm we achieve initial image classification results.Experiments show that our method is effective.2. We apply the angle-based high-dimensional outlier detection algorithm intofinding misclassified images. In outlier detection field, low-dimensional data isoften less than twenty dimension, others are treat as high-dimensional data. We usehigh-dimensional outlier detection algorithm to detect the possible misclassifiedimages and offer them to the post-processing module.3. We proposed image classification results optimization methods based onoutlier detection. We use maximum entropy algorithm to obtain the category ofpossible misclassified images and use the new category to replace the initialcategory obtained by the initial classification algorithm. We use thecomplementarity of classifiers and effective outlier detection to improve theperformance of classification system. 4. We design and realize the image classification experiment system based onoutlier detection. The system can handle classification problems of color images andgray images. After initial image classification, we do classification resultsrefinement based on outlier detection. We extract visual features, and then learnSVM classification model and maximum entropy model from the training set.Focusing on the test set, we firstly extract visual features. Then we use SVM modelto do initial classification, and use outlier detection algorithm to detect outliers.Finally, maximum entropy model is used to modify the category of outliers.Experiments show that our method has achieved effective classification on test data,and can satisfy the demand of using outlier detection to improve image classificationperformance.