Linear feature vector compression using Kullback-Leibler distance

Since multimedia contents became digital lossless and lossy compression techniques have been more and more relevant to store and classify such signals. Some of the linear transformation algorithm used for compression such as DCT PCA or LDA are known for decades and are successfully used for image, audio and video compression or in the field of multimedia content classification. In this paper a new linear algorithm for lossy feature vector compression is introduced. It can be used to simplify a dataset by reducing the number of dimensions of feature vectors (hopefully) without loss of information to enable a faster less memory consuming classification. The algorithm bases its compression strategy on the Kullback-Leibler distance.