Research On Image Retrieval Of Combining Low-level Features And High-level Semantics

With the rapid development of computer and multimedia technology, the number of multimedia image mushrooms. How to retrieval the image you want quickly and accurately in a gigantic image database is a crucial problem of the multimedia technology research. Traditional keywords-based image retrieval technology need manager annotate images by hand, so it not only cost so much human labors, but also have subjectivity, different manager maybe have different label for the same image. The content-based image retrieval technology mainly searches image by visual contents (color, texture, shape and so no). But people understands a image is a process that he uses his knowledge to speculate semantics of the image, thus, it lead a"semantics gap"between low level features and image semantics.To reduce"semantics gap", a method that combines the high level semantic and low level feature is proposed. It utilizes support vector machine (SVM) to transform low level features that are extracted from an image into high level semantics.In this paper, the semantic hierarchy of an image is analyzed and some classical methods of image semantic extracting are presented. On the basis of analysis some methods of extracting color feature, texture feature and shape feature propose, a low level feature extracting method that combines image edges and corners is proposed. It uses moment invariants express image edge and ring-shaped color histogram to express information of corners.This paper focuses on the multi-classification of SVM. In order to overcome the faults of traditional multi-classification of SVM, such as positive sample and negative sample are not balanceable, low recognition rate, train time so long, etc. a new tree structure SVM is proposed. Based on space distribution of the sample images, K-mean clustering is used to analyze space distribution among sample images semantic classification, and Euclidean distances among each clustering center are used as a tool to separate the classes. Firstly, two classes are classified, which have biggest distance of positive sample and negative sample of SVM in the root node of tree structure SVM. Then, the other classes will be classified to the corresponding SVM node if their distances are shorter to the one of two classes. For the other nodes, the classes are classified to two classes again. This step is repeated until only one class in the node. This distribution of positive sample and negative sample of SVM keep the balance of positive sample and negative sample. It classifies the two classes that have biggest distance to avoid disturbing other classification and increases the accuracy of classification. Moreover, it decreases the number of nodes of SVM and the training time of SVM. The experimental results show that the proposed method not only can improve image retrieval accuracy, but also reduce retrieval time.