Computerized Learning And Classification Of Traditional Chinese IWPs (Ink And Wash Paintings)

Along with the advancement of Internet and digital technologies, more and moreChinese traditional paintings are becoming available on the Internet. As a result,computerized indexing and classification of given Chinese paintings emerge to be oneof the focused research areas over recent years. As traditional Chinese paintings relyon the special drawing tools to illustrate the artistic styles, it distinguishes fromwestern paintings in terms of strokes, contours, color and textures etc. Additionally,drawing lines play important roles in most traditional Chinese paintings. Consequently,the existing research on Western paintings is normally not applicable to traditionalChinese paintings. In addition, color-based approaches are also not applicable astraditional Chinese paintings mostly rely on gray scale texture to express their artstyles and content.This thesis reports my intensive research programme on computerizedclassification and automated learning and analyzing techniques for traditional Chinesepaintings, in which a number of novel research and ideas are developed and putforward for style-based classification as well as its related theories and new conceptintroductions. My own novel contribution can be highlighted in three areas, including(1) pixel domain classification of traditional Chinese paintings;(2) wavelets domainclassification of traditional Chinese paintings; and (3) machine learning, especiallyone-class SVM (OCSVM) based classification of traditional Chinese paintings.In pixel domain, a new histogram-based global feature and local featureextraction algorithm has been developed to capture the artistic styles in terms of bothglobal perspective and local perspective, on top of which a novel entropy-balancedfusion scheme is also proposed to integrate the classification results. Experimentalresults support that the entropy of each classification effectively indicates theorientation of artistic styles and hence helps to improve the results even further.In wavelets domain, a number of new local features have been proposed toexploit the advantage of decomposition from the input art works and characterizingthe artistic styles across different sub-bands. The research also paves the way forfurther work on direct classification of traditional Chinese paintings in compresseddomain, where compression technology is based on wavelets transforms.Finally, the one-class SVM technology is revised to introduce a supervisedlearning element and arranged into a parallel OCSVM classifier. Based on thestatistics features, a new concept of enforced learning has been introduced to remove the false positives at each learning cycle together with a new scheme of adaptiveupgrading of decision parameters. Extensive experimental results support that theproposed new classifier achieves significant improvements in comparison withexisting representative techniques.