Study On The Technology Of Portrait Caricature Robot Based Machine Vision

To generate caricaturing portrait by using computer is one of front research field of machine vision. In theory, the technology of portrait caricaturing builds up valid method for evaluation of facial features. The technology sums up these facial features as a mathematic description, and expresses these features by results of quantification. So, the technology provides a kind of new theory for face recognition. The theory will greatly help with facial sorting and ID identification in theory. As for practice, portrait caricaturing robot can be used for scientific education and entertainment. The tech of evaluation of face can be employed to search criminals, security verification, VISA card verification, net meeting and so on. Moreover, the relevant methods for caricaturing can contribute to generating special effect of films, MTV and cartoon.The core problem of auto portrait caricature robot is how to caricature the facial features of face. By communicating with caricaturist and reading some caricature books, the paper combines the skill of caricature and computer technology, and divides the caricature technology into two layers: intuition layer and art layer. The intuition layer includes feature quantification and feature exaggeration. It describes some objective traits, and these methods of the layer can generate the caricature. The art layer includes texture feature, art style and individual connotation. These aspects represent some subjective feeling and art abstraction which can provide the caricature better expressive force.The paper focuses on the intuition layer of caricature, and figure out the quantification and exaggeration problems of facial features. In the art layer, the paper research on recognition of double-edges eyelid of texture feature. For verifying these theories and approaches mentioned in the paper, we build up a platform of portrait caricature robot. As for the paper, the technology of portrait caricature robot is to realize following procedures: firstly, to take a picture of front face. And then, to comprehend and get some features to generate portrait by computer. For creating a caricature, these individual facial features must be evaluated. After getting these features, the original portrait can be exaggerated by corresponding features. Finally, vector the caricaturing face to lines, and the robot can draw it on a paper.Through researching, in intuition layer, we can find that the features quantification and exaggeration would be divided into several key points: composition and feature selection eye location feature points recognition feature quantification feature exaggeration. Composition and feature selection will solve the problem of which features can express enough individual traits of objective face; the eye location is to provide reference and base for composition and feature selection; feature points recognition is the basis of feature quantification; feature quantification achieve the mathematic description of features; feature exaggeration is to exaggerate these facial feature to proper directions and grades. In art layer, the paper had discussed how to recognition double-edges eyelid which enhance the expressive force of portrait caricature. So, aiming at these key technique described above, the paper extract several unfathomed research points. We propose ourselves way to solve these problems which is the creative points of the paper:1. Present the method for selecting of features and feature points based on caricature composition rules.Before this paper, these are no papers proposed how to select feature points. Different papers select different feature points. So, maybe these feature points are not good enough to present facial features totally. Maybe many feature points had been selected, but the objective features don’t be expressed adequately. Thereby, the paper aims at the problem discussed above, the composition rules of painter are proposed. We employed mathematic language to express the composition rules. To build up models of facial features and feature points will provide solid theory foundation of selecting features and feature points. Further, the method accords with basic skill of drawing caricature and portrait. In fact, the method can be used in wider field of face recognition, and provide practical guiding.2. Present the eye location algorithm based on segment textures.According to the composition rules, for getting these features and feature points, the eye location must be spotted firstly. The paper analyzes deeply the frequency distribution of image transformed by Fourier Transform. By experiments and researching, we find that the frequency distributions of eyes, skin and edges take on different features respectively. By analyzing global and local textures, and finding out some stable features of frequencies that can indicate their individual features. According to these features the location of eyes can be found out. The paper uses AR database and private database to achieve experiments, and compare with other existed methods. we found that the algorithm takes high accuracy and robustness. The eye location can provide base of recognition and researching of feature points and qualification of features. Because the method is based on Fourier Transform, it can give some referenced proof and practical algorithm for feature analysis in frequency domain.3. The recognition of double-edges eyelid based on wavelet transformTo employ wavelet transform to process image at different layers, we can get some kind of textures at different layers. Wavelet transform can deform details of textures quickly. And during the transforming, the wavelet transform can express not only the directions of textures but also their strengths and positions. The paper presents that these feature points can indicate some potential features of details. For instance, the double-edge must emerge above the top eye line. To analyze the textures of the position, and calculate some normal directions of top eye line, we can get a 3D distribution plot by using the textures expand method discussed in the paper. By calculating the mean value and moments of 3D plot, these extra feature points indicated double-edge eyelid can be gained which can be used to generate caricatures. The algorithm can provide basic research method for analysis of detailed textures.4. Proposed the caricature quantification method for facial feature based on self-reference model.Aiming at the demerits of mean face model, the paper proposes self-reference model. The conception of self-reference model is to seek for a reference in a face to express other features’parameters. The method ignores these error and complex computation load derived from normalizing. It can calculate these features’parameters directly. By employing statistic method, use a lot of data and image to analyze the self-reference model, we can find that these facial features based on self-reference are complied with normal distributions. And then, by using the testing of normal distribution, we prove the normal distribution of them, and get the expectation and deviation. The expectation indicates the trend of a facial feature; the deviation reflects the range of difference of a facial feature. Additionally, by comparing the facial feature calculated by self-reference of caricatures drawn by caricaturist with the feature parameters of self-reference, we find they are accordant. So, the comparison proves the validation of self-reference model in theory. Finally, the caricatures generated by self-reference are demonstrated, it is very comply with intuitionistic feeling. So, the self-reference model provides an efficient and simple method for evaluating facial features which can express the difference between objective face and normal face at proper way.At last, the paper employs B spline curves to generate portraits and caricatures. According to given features scales, to move corresponding feature points will generate caricatures.The paper solves these key problems discussed above step by step. These lines of vector can be sent to robot and draw the caricature at a paper. The paper provided theoretic base and practical method for facial feature evaluation. Practice proves that the method is efficient, reasonable and practical. Moreover, the system had been exhibited at Science Museum of DongGuan, Science Museum of Shanghai and CIIF 2006 (China Industry International Fair 2006)...