Research On Semantic-Based Image Classification

With the rapid development of computer technology, communication technology and multimedia technology and wide application of digital images in all aspects of social life, technology of classification and retrieval of digital images develop. Traditional keyword-based technology of classification and retrieval of digital image can no longer meet the needs of people for its low efficiency and non-standardization; Content-based technology of image classification and retrieval achieve understanding of the image through measurement of the similarity from the bottom visual features. However, considerable differences exist between the bottom visual features and human understanding of images. Content-based technology of image classification and retrieval does not link up the bottom visual features with high-level semantics, cross the "semantic gap"; Semantics-based classification and retrieval technology of digital image is becoming a research hotspot.The thesis introduces the layer model of semantics and several methods to extract image semantics, focusing on the machine learning-based extraction method. Three bottom features have been extracted including color, texture and shape. Feature selection could speed up and improve the performance and efficiency of data mining. An improved sequential forward floating search feature selection algorithm is studied, and the time efficiency and outcome of sequential forward search algorithm, sequential forward floating search algorithm and improved sequential forward floating search algorithm are compared by experiment based on bottom features.Support vector machine is a new type machine learning method based on statistical learning theory. It has good generalization ability and could achieve nice classification results on small sample. The thesis chooses support vector machine as a classifier of image semantics, and studies classification performances of support vector machine on different feature subsets. The experiment result shows that the feature selection is effective and the incremental learning based on support vector machine is significant in application. It is also notes that the support vector machine based on structural risk minimization principle would not cause problems usually caused by traditional methods such as over-learning and local minimum. This thesis compares two incremental SVM learning algorithm: support vector set with misclassified samples and historical training data with misclassified samples. Experiment results show that the latter is more accurate and stable.