Multiple vector representations of images and robust dictionary learning

摘要

In this paper novel multiple vector representations of images are proposed and a robust dictionary learning method is designed. The multiple vector representation scheme enables an image to be observed with multiple views. Moreover, multiple vector representations are directly generated from the original image via a simple and efficient way whereas multi-view data usually have a high acquiring cost. The proposed method applies the same dictionary learning algorithm to the multiple vector representations and designs a very reasonable weighted logarithmic sum scheme to integrate classification scores of all vector representations. Main merits of this work are in the following points. First, it offers a quite novel viewpoint to take insight into representation of objects. It for the first time reveals that rows and columns of images can be viewed as two different sequences and pixel arrangements in terms of rows and columns of the image allow the object to be observed with two different angles of view. Second, it shows that when conventional dictionary learning algorithms are combined with the proposed multiple vector representations and weighted logarithmic sum scheme, very robust and accurate classification results can be obtained. This also partially means that diversity of vector representations of the image can be further consolidated by matrix decomposition in dictionary learning, so the resultant complementary information can be better exploited. Third, differing from conventional research routines, our study tells us that to fully dig and exploit possible representation diversity might be a better way to lead to potential various appearance and high classification accuracy of the image. The code of the proposed method is accessible at http://www.yongxu.org/lunwen.html. (C) 2019 Elsevier B.V. All rights reserved.