Design And Implementation Of Image Retrieval System Based On Feature Synthesis

Due to the huge increase in the amount and kinds of digital images available in the"Internet era", making efficient Content-Based Image Retrieval (CBIR) has become one of the major endeavors. Human beings are the main body of modern information society, so the images of them are the most commonly used types of images in the modern business and daily life, and widely existed in many huge image databases. This paper constructed an image retrieval system based on feature synthesis, using face detection technology which is applied to Content-Based Image Retrieval field.The general characteristics and domain related characteristics are combined in the feature selection. And the domain related characteristics are dedicated to the retrieval of human images, improving the image retrieval rate.In the system design, image filtering is applied based on the traditional structure of the CBIR system. The human images and unhuman images are categorized by the skin feature firstly, and then only those which meet the conditions for filtering selection to participate in the process of image similarity calculation. The results show that the design method can reduce the time of similarity calculation of the human images and greatly improve the speed of retrieval.In feature extraction phase, a method based on skin detection and cell-divided verifying is used to extract skin feature and face features. The skin feature is as filtering feature, face number, size percentage of the face, gained in the face detection process, are taken as the face features. General characteristics are extracted based on the region including color features, shape features, text features and spatial features. All of these characteristics have the features of zoom, rotation, translation invariant, therefore the system can search the images through zooming, paning and rotating more accurately.In feature matching phase, we combine the general vision feature together to calculate the region similarity fristly. An integrated region matching algorithm is used to calculate the image similarity secondly. If the retrieved image is image of human being, the face features paticipate in the conculation of similarity. Finally return the retrieval results sorted according to the image similarity. The final results indicate that the effectiveness of our retrieval system, and proved that our system searches images of human beings more rapidly and accurately.