Research Into Generation Of Realistic Procedural Texture Based On Visual Perception

Generally, texture has been known as a repeated pattern containing some elements or textons according to certain rules. There are two main aspects about research on textures. One is image processing, including texture analysis, texture segmentation, texture recognition, texture classification, texture retrieval, etc.. The other one is computer graph, including texture generation, texture synthesis, texture mapping, etc.. Traditional methods of image processing always compute features by mathematic approaches and then gain significant conclusions. Since the development of machine learning and data mining, such methods have been improved with self-study and self-adaptive abilities. Recently, researchers in the field of computer vision open up a new research area in which we can analyse textures from the view of visual perception. Some related works not only comprise theoretic research on visual perceptual dimension such as directionality and roughness, but also comprise application research based on visual perception such as texture retrieval and texture defects detection.So far, texture generation and texture synthesis have made great progresses. Especially a kind of texture generation algorithm named procedural texture arises great attention. The skill of procedural texture generation utilizes a set of software precompiled to produce textures. They are realizations of some mathematic models or some programs written in accordance with certain rules or flows. Users can gain different expected textures by give different parameters. Because the models of procedural texture generation are usually defined by mathematics, it is very difficult for common users to understand what the parameters mean.This thesis focus on finding a proper way to map the values of visual perception expressing textures used by human visual system to corresponding values of procedural texture generation models. Users can describe textures by familiar visual perception. Once they determine the values of visual perception, they could derive a group of values of relevant generation model and then driving such model to generate textures.This thesis contains seven chapters. Chapter one introduces the background, goal, content and innovations of our research. Chapter two makes a survey of procedural texture generation and demonstrates several popular algorithms of procedural texture generation. Chapter three introduces the progress of research on texture visual perception, analyses the main obstacles confronted by computer vision and vision science. Chapter four proposes the general requirements of psychophysical experiments in this thesis and the detailed requirements of the sample procedural textures, then prepares the sample procedural textures strictly according to the former requirements, designs and implements psychophysical experiments related to visual perception of procedural textures, including recording data about visual realness, perceptual grouping and rank of visual perceptual dimension individually. Chapter five analyses the data gained from psychophysical experiments, finds out the models which can generate textures possessing visual realness, sums up some interesting conclusions demonstrating human perception of procedural textures is affected by some characteristics such as gender, age and education. In the condition of limited samples, we also finds out the most important dimensions of visual perception by which subjects could describe the sample textures enough in Chinese context. Chapter six discusses a group of representative machine learning algorithms, at last chooses artificial neural network as our mapping tool. We compare BP and RBF neural networks and then choose RBF neural network to map values of visual perception to values of proper procedural texture generation model’s parameters, verify the validity of such mapping. Chapter seven summarizes the whole thesis, points out the problems and disadvantages in our work and analyses the content of future work.