Research Of Image Clustering And Classification In Subspace

Nowadays, with the popularity of digital cameras and the rapid development of the network multimedia, the number of internet images is growing at an exponential rate. Therefore, how to effectively manage and retrieve large scale internet images put forth a great challenge. Accordingly, how to improve the performance of image clustering and classification has become a hot research topic in recent year. But Because of the limitation of current image understanding technology, there is the famous semantic gap between high level concept and low level features for an image.In order to solve these difficulties caused by the differences between underlying visual features and high-level semantic and to improve the performance of clustering and classification, this paper presents a semi-supervised projection with spine embedding. The purpose of this algorithm is to learn an projection matrix, making the dispersion between-class of the labeled training data dimensional reduced by the projection matrix as large as possible, and making the dispersion within-class as small as possible. Meanwhile, the algorithm maps the local low-dimensional embedded coordinates into global low-dimensional embedded coordinates by local spine regression, keeping manifold structure of the projection data. This algorithm utilizes the labeled and unlabeled training data efficiently, and then obtains optimized image expression. But the proposed algorithm cannot preserve the metric structure in the subspace due to the nonorthogonal projection matrix. So we extended the original algorithm with orthogonal processing. Then we proposed semi-supervised orthogonal projection with spine embedding and applied it to the image clustering and classification tasks. Our algorithm utilizes a small number of labeled data and a large amount of unlabeled data for training, and tested on several image datasets. The experimental results show the effectiveness of the proposed algorithm.