Research On Basic Technics Of Semantic-Based Image Retrieval System

Image retrieval has become an active research area since1970s. It forms two researchdirections along with its development. But those applications which are developed bythese two methods can not meet people’s needs of image retrieval. And there is still abig problem of these two approaches—“semantic gaps” that is spanning between imagelow-level features and high-level semantic features—when computer tries to understandimages. Therefore, many experts and scholars worked on how to establish a mappingbetween image low-level features to high-level semantics to span this gap.In this paper, we propose a semantic image retrieval system, using an image semanticmapping approach based on image classification. In our method, image is represented invisual semantics space, and then been classified and annotated. We have four maincontributions in this thesis:①Image low level features based on locationAfter being segmented by self-adaptive image segmentation approach, the image isbeen divided into segments. Some of them have similar low level features such assegments of sky and sea, but they have totally different high-level semantics. They havea high probality to be represented to be an identical visual word when they are projectedinto the visual semantics space. But their location information can distinguish them well.So we employ it as a novel feature.②Weighted visual semantics space mappingMapping a word into semantics space with its weight is widely used to balance theinfluence of other words when dealing with text information. In image processing, thesize of the image area has the same contribution when expressing the semantics of theimage. So the proportion of the area to the whole image is used as a weight whenmapping the image into the visual semantics space.③Using the distance from samples to SVM’s boundary as image’s indexAll the results need an index value to sort samples in information retrievalproblems. The output of SVM about one sample has only two values:+1and-1. It cannot decide the retrieval results’ order. But the distance between the sample and SVM’sboundary can be used as an index value. We use its function as image’s membershipdegree of a semantics class.④Experimental verification We have verificated our approach by the experiments. It shows that:(a) Themethod has a nice ability to distinguish those regions which has a similar low-levelfeature but a different high-level semantics feature after introducing the locationinformation.(b) The algorithm of weighted mapping makes the performance of theregion with a large area. The big region represent the image more than small ones.(c)Using the distance between image and SVM’s boundary as the image’s index value, theresult of the retrieval system can be sorted reasonably. Thus our system can recommendthose images with a high index value to the users.Our semantic image retrieval approach has some practical value of application andbusiness to multimedia’s management and retrieval. For example, the management ofimages libraries, image searching engine.