Research On Multi-Feature Based Image Retrieval

Content-based image retrieval system mainly makes use of visible features, such as color, texture, shape and spatial position relationship and so on, as image content features to undertake match and find. In order to solve probably inaccurate retrieval results caused by only one visual feature, we present an image retrieval method based on multi-feature in this paper. The method proposed consists of two parts: the first retrieval and the second retrieval, respectively employed based on color feature and interest points feature.Color is one of the most common used visual features. Classical color-based histogram is easy to compute, insensitive to translation, rotation and scale, and quite resistant to noises as well. But its drawback is lack of spatial information, prone to yield false hits when distinguishing images in large database with similar color composition but different spatial distribution. The color-based first retrieval combines local color feature and spatial information to denote image content. partition images in the training set into fixed size cells and, for each cell, extract a local color histogram as the color invariant feature of the cell. All of the color invariant features are clustered into a number of patterns. Thus all the images in the database can be regarded as a collection of those patterns and transformed into a text-image. Thereby the well-developed text retrieval method can be applied for image query and index through such symbolic descriptions.The color-based image retrieval contains spatial information to a certain extent, compared with classical color histogram methods. But the inclusive spatial information is still not enough to reflect exact image structure, dissatisfied to retrieval in large database. So the paper puts forward the second retrieval method based on interest points, and implements more precise retrieval from the related images acquired by the first one. We adopt the wavelet transformation to get the points of interest. Wavelet transformation could filter noises and increase the correlation between pixels; the obtained characteristic points have multiple resolution, can reflect actual image features. In the second image retrieval, image content information is hidden in those interest points. proper vectors is used to abstract the spatial information of point distribution. The similarity of two images can be determined bycomputing the distance between vectors.Experimental results show that the new method is effective to extract color feature with spatial information, better in retrieval effect than the color-based histogram methods with or without weights proposed previously, and robust to the retrieval of images with domain-free scenes.