Research On The Techniques Of Semantic-Based Image Retrieval

In today's world science and technology are advancing rapidly, and computers have played an ever-growing role in all of our lives. With the rapid development of multimedia, network and storage technology, all walks of life are in great need of pictures. So the size of the image database is growing at an alarming rate. How can we find needed image information from voluminous image database? This requires a kind of technique to find images fast and accurately, which is the so-called image retrieval techniques.Traditional image retrieval techniques, such as text-based image retrieval, can not fully meet the needs of the people. Content-based image retrieval technique analyzes image on the basis of the visual features embedded in image itself. With the feature vector as index, through the image similarity matching, CBIR realizes image retrieval. Many scholars have made higher levels abstraction for image retrieval, namely high-level semantic based image retrieval. Semantic-based image retrieval searches images according to use's understanding. But between low-level visual features which have limited expression ability and users' rich semantic expression, there are great differences, which is the so-called "semantic gap". How to solve the "semantic gap" is the key problem of semantic-based image retrieval. Now widely used solution is establishing a mapping from the low-level features to the high-level semantic of the image. The low-level features include global features and regional features. In order to obtain regional features of the image, we should segment the image into some meaningful regions. From the segmented regions we can get the detail information of the image, and then extract more high-level semantic.According to the above research frame, this paper first introduces the research background and development of the image retrieval techniques. Through the introduction of image semantic model, image semantic description and image semantic extraction for SBIR, the paper discusses the problems existing in SBIR. In order to extract effective color characteristics, this article introduces the concept of color quantization and puts forward a kind of color quantitative algorithm based on QPSO with dimension mutation operator. Experiments prove that the algorithm have a fast convergence speed and a good quantitative effect. On the basis of color quantization, in order to extract high-level semantic features of the image, the paper further proposes a kind of FCM cluster analysis algorithm based on QPSO used for color image segmentation. Experiments show that the algorithm is fast and accurate.Finally, extracting the low-level features of the segmented regions to get feature vectors, according to the similarity of the feature vectors we classify the segmented regions and then calculate weights for every region. Each image of database is described by a semantic vector, each component in vector represents the weight of the corresponding segmented regions. In order to realize SBIR, classify the images of database with FCM cluster analysis algorithm based on QPSO.