Translator Disclaimer
30 October 2009 Boosted distance based on local and global dissimilarity representation
Author Affiliations +
Proceedings Volume 7496, MIPPR 2009: Pattern Recognition and Computer Vision; 74960X (2009) https://doi.org/10.1117/12.833006
Event: Sixth International Symposium on Multispectral Image Processing and Pattern Recognition, 2009, Yichang, China
Abstract
We propose a new distance estimation technique by boosting and apply it to enhance the effectiveness of classifier when the training set is insufficient. The proposed method is called Boosted Distance based on local and global dissimilarity representation (BDLGDR). It is a modified method of Boosted Distance. Rather than simply differentiating the feature vectors, we calculate a new dissimilarity representation of each couple of feature vectors. This new dissimilarity representation contains two parts: local dissimilarity representation part and global dissimilarity representation part. The proposed method does not only achieve high classification accuracy when the training set is insufficient but when the number of training set is sufficient it also can achieve as high accuracy as AdaBoost. The method has been thoroughly tested on several databases of high-resolution (1.25m) Terra-SAR images. In the first experiment, we decreased the number of the training sample per class from 10 to 1. The result showed that the proposed method outperformed both Boosted Distance and AdaBoost. In the second experiment, we used sufficient training samples. The experimental result illuminated that the proposed method performed at least as well as AdaBoost and needed fewer iteration rounds to converge than Boosted Distance.
© (2009) COPYRIGHT Society of Photo-Optical Instrumentation Engineers (SPIE). Downloading of the abstract is permitted for personal use only.
H. Yin, Y. F. Cao, and H. Sun "Boosted distance based on local and global dissimilarity representation", Proc. SPIE 7496, MIPPR 2009: Pattern Recognition and Computer Vision, 74960X (30 October 2009); https://doi.org/10.1117/12.833006
PROCEEDINGS
8 PAGES


SHARE
Advertisement
Advertisement
Back to Top