Research On Image Retrieval Based On Textual Semantic And Visual Content

With the rapid development of digital imaging, data storage, network and other technologies, all kinds of images are increasing at an alarming rate. The increasingly richness of images resource make it difficult for users to find the information they really need from the vast database, so that efficient image retrieval technology becomes a hot spot of research community in recent years.Nowadays, image retrieval technologies mainly include text-based image retrieval and content-based image retrieval. The former largely relies on the annotated textual information to retrieval, however, it would cost heavy burden of manual annotation when facing billions of images, so that the retrieval method gradually cannot meet the requirement for reality application; the latter largely relies on the technology of feature extraction and high-dimensional indexing, however, the images with similar feature may have unrelated semantic because of the semantic gap, so that the retrieval method cannot meet the information needs of users in many cases. In order to give full pay to the two technologies’strength and reduce their defects, many researchers devised various methods to combine the two technologies to retrieval which dramatically improve the performance of image retrieval. On these basics, the thesis makes some research works on how to fuse the visual content and textual semantic to retrieval images; it mainly divides into following three parts:1The thesis studies the method of describing image’s visual content. In order to overcome the shortcomings of slow extract speed and so on in traditional SIFT algorithm, the method combines an improved DSIFT feature extraction algorithm with BOW model to describe the image’s visual content by the way of constructing visual words.2The thesis researches a method to automatically acquire semantic information according to the image’s visual contents. In order to overcome the shortcoming that semantic acquired by manually annotate would cost expensively, the method proposes an automatically acquire image’s semantic algorithm based on PLSA model. The algorithm regards an image as the fusion of some textual information topics and visual information topics, and it first studies the textual information and visual information by fitting the two PLSA models, then it uses an adaptive asymmetric learning approach to make the association between the two types of information, so that it can automatically acquire semantic information according to the visual content through the association.3The paper researches an image retrieval method which combined the textual information with visual information to improve the precision of retrieval. The method proposes the idea of layered retrieval, it first preliminarily searches and selects the semantically related images from database according to the semantic information automatically acquired, then it particularly retrievals at the second level to sort the filtered images according to the similarity of the extracted visual features.Experimental results in Corel1K database on the self-developed retrieval prototype system demonstrate that the research works of the thesis really improve the performance of image retrieval.