Points Of Interest Based On A Digital Image Retrieval And Relevance Feedback

In recent years, image is more and more widely used in all walks of life, not only in terms of quantity being produced with immeasurable speed, but also in the availability and importance have been unprecedented growth. There has been intensive research on CBIR technology, but this technology is mainly used in specific and professional gallery, fully automatic and intelligent retrieval process is not implemented. And it has become a challenging problem for us to effectively manage the vast and complicated image database. This paper extracts features from the interest region that can reflect users’ retrieval needs, being based on the interest points of the image. As CBIR system currently exists the problem named "semantic gap", retrieval and feedback processes in image retrieval have been optimized. In this paper, the specific contents include:1. In order to avoid the condition that non-maxima suppression is not easy to set the threshold, an effective method about Harris corner detection is used to obtain corner information of an image, which threshold depends on the image itself. Then characterize the main information of the image by the corner points instead of interest points.2. The center of circumscribed circle of rectangular area that includes all the interest points can be considered the interest region center, and circumscribed circle can be regarded as the interest region. The circumscribed circle is divided into six concentric rings that have the same radius. In each of concentric rings, color and texture features are extracted. The feature of color is extracted by color histogram with36-dimensional. The feature of texture is extracted by Gabor filtering with4frequencies and6directions. Combined with the distribution of interest points and the importance of every annular region in interest space, an improved weighted algorithm for feature vector is used to obtain the color and texture features of the entire interest region. Then the color and texture features are normalized and weighted in order to obtain the integrated feature for image retrieval.3. In order to further narrow the "semantic gap", SVM relevance feedback is added in the image retrieval process, which is adapted to the small sample size problem. According to the original algorithm, and combine with the interest region, an improved sample set is used, which can be continuously updated and expanded. Then we show the effectiveness of the improved algorithm in image retrieval through some experiments.