Graph-guided fusion penalty based sparse coding for image classification

Publisher:
Springer
Publication Type:
Conference Proceeding
Citation:
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 2013, 8294 pp. 475 - 484
Issue Date:
2013-01-01
Full metadata record
Files in This Item:
Filename Description Size
10.1007%2F978-3-319-03731-8_44.pdfPublished version404.01 kB
Adobe PDF
In image classification, conventional sparse coding only encodes local features independently. As a result, the similar local features may be encoded into code vectors with large discrepancy. This sensitiveness has became the bottleneck of the traditional sparse coding based image classification methods. In this paper, we propose a novel graph-guided fusion penalty based sparse coding method. To alleviate the sensitiveness of the traditional sparse coding, our approach constrains that the similar local features are encoded into similar code vectors. To achieve this goal, we add the popular graph-guided fusion penalty term into the traditional l1-regularized sparse coding formulation. Finally, we adopt the multi-task form of the smoothing proximal gradient method to solve our optimization problem efficiently. Experimental results on 3 benchmark datasets demonstrate the effectiveness of our improved sparse coding method in image classification. © Springer International Publishing Switzerland 2013.
Please use this identifier to cite or link to this item: