Researches On Image Features

Researches on Computer Vision has been booming since late 1970's. Many new fields have emerged, such as Image Segmentation, Stereo Vision, Visual Motion Analysis, Image Based Modeling and Rendering, Modeling, Range Image Analysis and etc. All these researches focus on one topic, i.e., how to retrieve most important information from huge amount of signal data and then use these information to analyze and understand image signals. This information is called Image Feature.This thesis focuses on the study of Image Features by analyzing them in different applications, which include: Blind Image Restoration, Object Tracking in Video, Human Expression Mapping and Human Face Synthesis.Chapter III discusses Blind Image Restoration in detail. An algorithm named RIBIM (Ratio Image Based Iteration Method) is introduced. The algorithm suppresses noise by analyzing texture features in image and estimates the PSF (Point Spread Function). Finally, Ratio Image is used to generate initial image of recursion for restoration.Chapter IV focuses on Object Recognition and Tracking in complicated scenes. A 360 degree camera is used in the application which enables non-blind-spot supervision. A training method named TALENT (TemplAte LEarNing based Training) uses clustering to improve searching speed. Training based recognition and status transition map based tracking enhances supervision results.Chapter V discusses Human Expression Mapping by Ratio Image technology. After introducing current expression synthesis techniques, an improvement of expression mapping algorithm is presented, which greatly reduces computational costs and suppresses noise.Chapter VI demonstrates a framework for Human Face Synthesis: RATES (Region based feAture Tracking and Expression Synthesis). By training 497 different faces of different ages, genders and face shapes, RATES can automatically recognize, locate, align and morph human faces in images. RATES uses bottom-up method to precisely locate feature points on human face by 13 training data sets. In the texture morphing phase, GPU technology is implemented for fast synthesis. Skin color realignment and special treatments on edge features greatly improve the morphing results.