Title :
Compression of pre-computed per-pixel texture features using MDS
Author :
Pang, Wai-Man ; Wong, Hon-Cheng
Author_Institution :
Comput. Arts Lab., Univ. of Aizu, Aizu-Wakamatsu, Japan
Abstract :
There are many successful experiences employing texture analysis to improve the accuracy and robustness of image segmentation. Usually, per-pixel based texture analysis is required, which involves intensive computation especially for large images. Precomputation and storing of the texture features involves large file space which is not cost effective. To adopt to these novel needs, we propose the use of multidimensional scaling (MDS) technique to reduce the size of per-pixel texture features of an image while preserving the textural discrminiability for segmentation. Per-pixel texture features will create very large dissimilarity matrix, making the solving of MDS intractable. A sampling-based MDS is therefore introduced to tackle the problem with a divide-and-conquer approach. A compression ratio of 1:24 can be achieved with an average error rate lower than 7%. Preliminary experiments on segmentation using the compressed data show satisfactory results as good as using the uncompressed features. We foresee that such a method will allow texture features to be stored and transferred more efficiently on low processing power devices or embedded systems like mobile phones.
Keywords :
divide and conquer methods; image coding; image segmentation; image texture; matrix algebra; dissimilarity matrix; divide-and-conquer approach; embedded systems; image segmentation; low processing power devices; mobile phones; multidimensional scaling technique; precomputed per-pixel texture feature compression; sampling-based MDS; textural discriminiability; Compressed texture features; Compression; Gabor wavelet transform; Multidimensional scaling;
Conference_Titel :
Picture Coding Symposium (PCS), 2010
Conference_Location :
Nagoya
Print_ISBN :
978-1-4244-7134-8
DOI :
10.1109/PCS.2010.5702517