Perceptional Texture Classification Technology Based On Visual Patches Extraction

Perceptional texture is a very important research topic in computer vision andpattern recognition. In this thesis, we analyzed perceptual texture classification andconcluded that textures contain some visible “features” that humans may use forperceptual classification. To find visible feature patches, we started with featureextraction and clustering analysis. Optimization algorithm and classifiers were used toobtain visual patches for describing visual texture features.Most texture descriptions are under supervision, too narrow to show the details ofvisual features. However, we extracted texture patches that contain abundant visualinformation under unsupervised learning. The distinctive patches were used to explorethe relationship between feature patches and perceptual texture classification.For feature extraction, we proposed an optimized K-means clustering algorithmusing SVM (support vector machine). The SVM classified and predicted the oldpatches with the help of the existing clusters. Indistinctive and duplicative clusterswere abandoned, while new and distinguishable patches were included into newclusters. This process ended until all new clusters were built up. In contrast withtraditional texture classification scheme, our resulting visual vocabulary of clusteranalysis was compared with perceptional texture classification experiment.