Natural Image Retrieval Based On Semantics

With the rapid development of computer, multimedia and Internet, a great many images have been produced, which makes the effective and rapid search for a certain image from a large database increasingly important. Content-Based Image Retrieval (CBIR) mainly deals with this problem. To make CBIR efficient and effective, semantic information should be fully used. Founded on CBIR, Semantic-Based Image Retrieval extracts the semantic information of images in a multi-channel way, searches a certain image according to the semantic information, to meets people's demands further.This paper mainly focuses on the research of extracting and retrieval of semantic information for images, and provides a high-efficient and practical Semantic-Based Image Retrieval system for natural images. The main work of this paper:1. Proposed a new SVM learning algorithm. According to "semisupervised learning", both labeled and unlabeled data were used to train classifier. Combining this idea with standard SVM classifier and adding a mixed data sets near the interface. It was used for classification of small data sets;2. Provided a simple yet rapid searching framework based semantic information. Firstly, it classified the segmented natural images, extracted their features and built a features database. After that, it applied multi-class classifier training to each class to get a class model. After classifying and recognizing the images, semantic information of each image was finally extracted. While searching, it only checked whether the image's semantic keyword matches users' demand to determine whether it was a hit. This, of course, saved searching time greatly while improving searching effectives and maked users easy to use as well;3. Provided an improved algorithm for color features extraction. Considering the segmented image mainly included one object, it only extracted the major factor of color histogram as the color feature for matching. This algorithm reduced calculation and storing volume. While extracting, a threshold should be set, and the mean of color histogram was choosed;4. Applied the JSEG algorithm to segment the natural images. It considered the spatial information of images (which was often neglected), and segmented an image following two steps: color quantization and spatial segmentation.