Abstract:  In computer vision and information retrieval fields, many applications, such as appearance-based
image recognition, often confront high-dimensional data samples. The curse of high dimensionality is usually a major cause of limitations of many machine learning algorithms. Hence, it is desired to consider methods of feature extraction (or dimensionality reduction) which are able to find the low-dimensional and compact representations for the high-dimensional data points. The subspace learning algorithms are one of the most popular feature extraction methods. Supervised subspace learning algorithms usually achieve better performances than unsupervised ones. And supervised subspace learning algorithms can be divided into two categories, the global structures based discriminator, such as linear discriminative analysis (LDA), and the local structures based methods, such as marginal Fisher analysis (MFA). From the experiments on image recognition, we can find that the global structures based discriminator and the local structures based dis-criminator are suitable for different feature extraction tasks. Hence, we hope to seek a discriminative analysis method which can combine the global structures and local structures of data sets together. In this paper, a new supervised extraction method, called adaptive integrated global and local discriminant analysis (AIGLD), is proposed. The AIGLD algorithm combines the global structure based discriminator (namely, Linear Discriminant Analysis, LDA) with a proposed local structure based discriminator together so that it can use both of the global and the local discriminant information of data sets simultaneously. Compared with LDA and the existing local structure based discriminators, AIGLD is capable of feature extraction for different types of data sets. Moreover, an adaptive method for choosing the parameter which is used the balance the effect of local and global discriminators has also been proposed. This method is much more efficient than the classical method for parameters selections, namely cross validation. The efficiency of the proposed algorithm is demonstrated by extensive experiments using UCI data sets and benchmark face image data sets including ORL database and CMU PIE database. And the experimental results show that IGLD outperforms other classical and state of art algorithms.