Texture Synthesis And Classification Based On Support Vector Machines And Ensemble Learning

Texture synthesis and classification are active research topics during the past few years, they are widely used in image processing, computer vision and pattern recognition field.The aim of texture synthesis is to produce a new image which is larger than the sample image, and the new one has the same texture characteristics with the sample image. Current techniques of 2D texture synthesis are nearly mature. Unlike 2D texture,real-world 3D surface textures normally comprise rough geometry surface and various reflectance properties. The same surface texture can vary dramatically with different illumination directions and view conditions. However, 2D texture synthesis methods ignore the effect of illumination and view conditions. So the research on 3D surface texture synthesis is more realistic. As a very important branch of pattern recognition field, the aim of texture classification is to distinguish the difference of all objects with different texture characteristics.This thesis introduces three steps including texture synthesis, texture classification and texture feature selection.In this thesis, several methods about 2D texture synthesis, feature extraction, feature selection and classification are first introduced, and then the research on texture synthesis and classification is discussed. Unlike texture synthesis using support vector machines, which can only synthesize highly structured texture, the proposed method based on support vector machines and multiple seed-blocks can successfully synthesize both random and structured textures, and it is further extended to 3D surface texture synthesis. Because training different texture data can obtain different models, a method using SVM twice for feature extraction and texture classification is proposed on the basis of the character. In texture synthesis and classification experiments, training a SVM and prediction (synthesis) require high processing time. To resolve the problem, a new efficient feature selection method based on Bagging ensemble learning method is introduced to reduce the number of characteristics. The experiments show the effectiveness of the proposed texture synthesis, classification and feature selection methods.