Dimensionality Reduction In Image Retrieval Applications

With the constantly development of information technology, the number of the image data has been showing a geometric progression of growth trend. How can the users search for the multimedia data which they really need from the vast multimedia database has become a hot topic.Content-based image retrieval (CBIR) technology is a study in this research field. The paper based on CBIR mainly focuses on low-level image extraction, in which p layer quadtree decomposition strategy is used. It proposes a new region-based feature extraction method. The method not only extracts the image color feature but also fully takes the spatial information into account. After the dimensionality is reduced by LLE, the new feature extraction method has a better retrieval performance. The work mainly engaged in is as follows:First, every image in the database is divided according to p layer quadtree decomposition strategy, then for each block, we calculate the low-level color moment and threshold value as its image feature. A comprehensive feature of each block to describe the general characteristics of the image is obtained.Second, responding to the method of extracted image features which has the defects of high dimension, some improvements are made. LLE-a kind of non-linear dimensionality reduction method is applied, then it will reduce both spatial and operation complexity.At last, numeral experiments are made. Results show that the quality of image retrieval is in good accordance with the theory target.