Unified Dictionary Learning and Region Tagging with Hierarchical Sparse Representation

Publication Type:
Journal Article
Computer Vision and Image Understanding, 2013, 117 (8), pp. 934 - 946
Issue Date:
Filename Description Size
1-s2.0-S1077314213000623-main.pdfPublished Version3.06 MB
Adobe PDF
Full metadata record
Image patterns at different spatial levels are well organized, such as regions within one image and feature points within one region. These classes of spatial structures are hierarchical in nature. The appropriate integration and utilization of such relationship are important to improve the performance of region tagging. Inspired by the recent advances of sparse coding methods, we propose an approach, called Unified Dictionary Learning and Region Tagging with Hierarchical Sparse Representation. This approach consists of two steps: region representation and region reconstruction. In the first step, rather than using the ℓ1-norm as it is commonly done in sparse coding, we add a hierarchical structure to the process of sparse coding and form a framework of tree-guided dictionary learning. In this framework, the hierarchical structures among feature points, regions, and images are encoded by forming a tree-guided multi-task learning process. With the learned dictionary, we obtain a better representation of training and testing regions. In the second step, we propose to use a sub-hierarchical structure to guide the sparse reconstruction for testing regions, i.e., the structure between regions and images. Thanks to this hierarchy, the obtained reconstruction coefficients are more discriminate. Finally, tags are propagated to testing regions by the learned reconstruction coefficients. Extensive experiments on three public benchmark image data sets demonstrate that the proposed approach has better performance of region tagging than the current state of the art methods. © 2013 Elsevier Inc. All rights reserved.
Please use this identifier to cite or link to this item: