Svm-based Relevance Feedback In Image Retrieval Research

The development of multimedia and internet techniques has brought on a rapid increase of the number of digital images. The traditional text-based image retrieval faces some problems when dealing with large scale image databases. It is necessary to extract more comprehensive, more general and more objective image features to conduct image retrieval. Content-based image retrieval (CBIR) was proposed to meet such a requirement. It's a technique that retrieves relevant images based on image content, such as color, texture and shape features. This paper focuses on SVM-based relevance feedback in CBIR. The main work of this paper is summarized as follows:1. The SVM-based relevance feedback techniques are studied and compared with the re-weighting method used in the MARS system. To get around the small sample issue, the feedback examples are accumulated after each iteration. A new feedback method using multi-SVM is proposed to improve retrieval performance. The experimental results show that the SVM-based techniques have wonderful performance in image retrieval using the features and image database chosen by this paper.2. A new HSV color space quantization method is proposed. In this method, when intensity value or saturation value is smaller than a certain threshold, color is approximated by a gray value; otherwise color is represented by its hue and saturation, thus the influence of intensity is ignored. This quantization method reduces the dimension of color histogram and experiments have shown its good performance in image retrieval.3. Multi-feature based image retrieval is conducted. Color, texture and shape features are extracted from images and combined to retrieve relevant images. Experiments show that using multi-feature achieves much better retrieval performance than using a single image feature.4. A content-based image retrieval system is designed. This system allows user to select the image features before conducting the retrieval. Users can choose color, texture, or shape, or any combination of these three features. The system also supports relevance feedback. Two relevance feedback methods are available: SVM-based and re-weighting.