Research On Semantic-based Image Retrieval Technology

With the development of Multimedia technology and Internet, the application of digital image has been increasingly widespread and it leads to the image retrieval technology becoming more important. Since the last decade, people's need of image retrieval has been increased. In order to retrieve these huge image data effectively, Content-based image retrieval technology become popular.The most existed Content-based image retrieval systems use traditional low-level features such as color, texture and shape to describe the image content, which are usually represented by statistic data. Actually, there are big differences between these statistic data and the image content which people understand. Because people's understanding of image content isn't based on statistic and the image content is fuzzy, it couldn't be represented by simple vectors. These problems lead to deflection of the image representation and people's understanding, which called semantic gap. Usually, we couldn't get satisfied result if we only use the low-level feature to retrieve.So, how to describe image and to make it coincide with people's understanding become the key point of improving retrieval accuracy. In the point of cognition, people's understanding and description of image content is on semantic level. How to reduce 'semantic gap', how to accurately represent content semantic of image and retrieval intention of people becomes important and critical.Based on image feature retrieval, this paper introduces a fuzzy semantic classification and relevant feedback technology. By pay attention to the mapping of image feature and semantic concept, we explore a map method for low-level feature and semantic under special scenes. The main studies are in the following aspects:1. Use feature extract algorithm to extract color, texture and shape feature of image, which can support semantic mapping. Image content understanding of Computer is .based on image features. Superior extract algorithm can increase accuracy of semantic mapping. Furthermore, it can improve the retrieval effect finally.2. Use fuzzy classification to establish map of low-level feature and semantic. Capturing semantic concept of image is the first step of semantic-based image retrieval. Under the present computer technology and pattern recognition technology, capturing semantic concept should be based on image object recognition. Recognizing image object and connecting it to semantic through pattern recognition method can map low-level features to semantic concept. This paper introduces fuzzy classification technology to establish mapping from low-level feature to semantic concept by choosing the best training sample. Meanwhile, fuzzy set can categorize one image into different semantic category. This is more consistent with people's cognition.3. Use relevant feedback based on semantic classification to bring in people's understanding of semantic concept. According to the distinction between anticipated result and system retrieval result, users can select two kinds of images, the one which could represent retrieval intention accurately and the one which couldn't, and then doing feedback. Based on this, users can remedy the mapping between images and their semantic concept. So, users' understanding of semantic concept could be better represented and the subjective of training sample selector could be reduced. At the same time, relevant feedback can bring down subsidiary degree of semantic concept, of which mapping image is similar to feature but irrelevant to semantic. It can get better retrieval result.