Analysis Of 3D Face Models Based On Multi-scale And Multi-level Feature

With the rapid development of computer graphics,numerous applications and 3d modelshave been studied and explored by researchers.Feature analysis of models has always been a focus of attention.Traditional feature analysis and detection of methods focus on the model's points,lines,or the local characteristics,and recently,more and more new application shows demand for more general and higher level information(semantic or human intention information)as to detection.3D model feature detection is very important in geometric processing field.Because the feature is a link between models and applications.Accurate feature detection contributs to development of computer graphics' s application,such as pattern recognition,mesh segmentation,comprehension,etc.Feature detection of huaman face model has a far-reaching bearing on 3D face reconstruction,matching,recognition and so on.About feature's detection and describing,most of previous work mainly used model's geometric property based on feature regions,such as statistical information of first-order or second-order differential geometric data.But these methods don't integrate globle property with local property very well.Graph wavelets can reflec model's local property and analyze the signal function in frequency domain with multi levels.Most of previous work about graph wavelets is used in graph or whole models,not in local property's feature description.We will introduce a multi-level and multi-scale feature descriptor which we used to detect human faces' feature on face models.Our aim is to get the five sense organs of a face model.The main idea of the descriptor is to convert the signal function from the time domain to the frequency domain based on graph wavelets.It can describe the feature in multi levels.It also can use statistical information based on distance fields to describe the feature in multi scales.We don't use graph wevelets' coefficients as a part of our descriptor.Instead,we use probability distribution to re-organize the graph wavelets' coefficient together.In this way,our descriptor can be adopted on models in various scales.We adopt signal functions which are more suitable to human face models: a self-defined curvature signal and a color signal.These two signals are not sensitive to noise and can highlight features of faces.When measuring descriptors' robustness,we use heat kernel signature as a signal function.As for faces' details,experiment results show that the discrimination of our descriptors is more significant than other descriptors.For example,our descriptors can retrieve the feature of eyes in multiple models accurately.Our descriptors are more suitable for large scale and refined models.In the meantime,our descriptor is not sensitive to noise.Experiments showthat our descriptors have robustness.The extracted facial feature areas are relatively accurate and can be applied to mesh alignment,registration and deformation.