Citation

Liu CH, Lin JK. PLoS One 2015; 10(11): e0142820.

Affiliation

Department of Medical Imaging, The 174th Hospital of Chnese PLA, The Succeed Hospital affiliated to Xiamen University, Xiamen, Fujian Province, China.

Copyright

(Copyright © 2015, Public Library of Science)

DOI

10.1371/journal.pone.0142820

PMID

26569486

Abstract

In this paper, we give a systematic study to report several deep insights into the HOG, one of the most widely used features in the modern computer vision and image processing applications. We first show that, its magnitudes of gradient can be randomly projected with random matrix. To handle over-fitting, an integral histogram based on the differences of randomly selected blocks is proposed. The experiments show that both the random projection and integral histogram outperform the HOG feature obviously. Finally, the two ideas are combined into a new descriptor termed IHRP, which outperforms the HOG feature with less dimensions and higher speed.

Language: en