Feature Analysis And Machine Learning Of Facial Beauty Attractiveness

Beauty is a universal part of human experience, and the perception of facial beauty or attractiveness is one of the most common human activities in our daily life. Facial beauty delights our sight and provokes pleasure in our mind, but, what is beauty? Is there any“beauty code”existing objectively? The facial beauty or attractiveness is easy to recognize but hard to define. However, recent studies in psychology have demonstrated high cross cultural agreement in attractiveness rating of faces of different ethnicities, and this high congruence over ethnicity, social class, age, and sex has led to the belief that perception of facial attractiveness is data driven irrespective of the perceiver and provides theoretical basis for the research on facial attractivenessThe research on facial attractiveness has had a significant part, there are many studies in psychology, anthropology, philosophy, plastic surgery, but relatively less in computational science for several decades. However, in recent years, it has received attention by the scientists of computation science, and the study on facial attractiveness by means of machine learning and image analysis is becoming an emerging topic. In this thesis, we study on this new research topic and aim to depict facial attractiveness quantitatively and objectively. Our research on facial attractiveness focus on three aspects: the feature extraction, machine learning and prediction model development, and the facial beauty composition on the basis of average hypothesis. The main works in our thesis include:(1) We study on whether the notion of facial beauty or attractiveness can be learned by machine. According to the principles of Turing test in artificial intelligence, we propose a hypothesis on the Learnability of facial attractiveness. The experimental results demonstrate that, by means of quantitative feature analysis and presentation and the adoption of appropriate machine learning algorithms, the concept of facial beauty attractiveness can be learned by machine with numeric expressions.(2) A facial attractiveness prediction model based on image feature analysis and machine learning is proposed. Within the framework of this model, we present a new method for facial attractiveness evaluation using our proposed geometric features and popular machine learning methods. The machine learning methods employed include SVM regression, LMS linear regression, Akaike linear regression, Gaussian regression, MLP, KNN regression etc. For geometric feature extraction, we studied 6 sets of geometric features previously used, and based on the Chinese traditional facial beauty standard---San Ting Wu Yan and the current Chinese aesthetic theory, we propose our own geometric feature sets such as 17-dimensional distance feature, normalized-feature-point feature, triangle area feature and two combined feature to exact the geometric information which makes critical contribution to facial attractiveness more completely and accurately. Experimental results show that the 17-dimensional distance feature we proposed has better prediction performance, and its combinations with the other feature sets can further improve performance, the highest correlation of 0.916 with average human ratings is obtained by the combination of all our proposed feature kinds and the subsequent feature filtering using wrapper. The experimental results on the one hand prove the efficiency of our geometric features for the evaluation of facial attractiveness, on the other hand, they also validate the hypothesis of“Facial attractiveness can be learning by machine”we proposed in this thesis.(3) We propose to use 2-dimensional Gabor texture features to predict facial attractiveness, inspired by the close relation between facial texture and facial attractiveness which is supported by cognitive psychology, biology and medicine. On the basis of traditional Gabor filters, we present two new methods, feature points sampling Gabor feature and triangular centroid sampling Gabor feature, to extract facial texture information. We first perform image normalization and preprocessing, then texture features are extracted followed by dimension reduction using principal component analysis (PCA), after that, wrapper feature selection is performed to remove irrelevant features for better performance, and finally, SVM is employed to train the model and predict facial attractiveness. Experimental results prove that the texture feature resulted from our proposed feature point sampling Gabor feature and triangular centroid sampling Gabor feature perform significantly better than the other texture features, and so far the highest correlation of 0.932 with average human ratings is obtained by combining our Gabor texture features with the geometric features. These results clearly demonstrate the importance of texture information and the effectiveness of our feature extraction methods for facial attractiveness.(4) We present a mew method of creating good looking facial composites based on the average hypothesis in psychology and biology. Superior to the simple arithmetic averaging and line drawn representation of the averaged composites in previous research, our method takes into account the averaging of shape features and texture features simultaneously, results in natural and good looking averaged facial composites. The composition process includes the extraction of anatomical landmark points using AAM(Active Appearance Model), mesh generation covering all the landmark feature points by 2-D Delaunay triangulation algorithm, and the apply of Generalized Procrustes Analysis (GPA) for shape alignment in which landmark feature points are used for one-to-one point correspondence, finally, the local affine transformation and weighted averaging for all the images. Experiments show that the obtained averaged facial image always looks beautiful, with suitable proportions of the five sense organs, good-looking layout and skin color, and get above-average human ratings of facial attractiveness. The result is encouraging because it validates the effectiveness of our proposed method for producing attractive facial composites, and shed new light on the averageness hypothesis in psychology. In addition, from another perspective, it indirectly verifies our proposed hypothesis that the machine can learn the notion of facial attractiveness. Interestingly, by introducing weighting mechanism, our approach can make the facial composite more identical to some pre-designated individual image, resulting in many innovative applications such as facial beauty enhancement, facial baby processing and facial aging, etc.In the fields of pattern analysis and image processing, the study of human facial attractiveness is an emerging area. Our work in this thesis is a meaningful attempt to study the intelligent perception of facial attractiveness by machine. The research on facial attractiveness is far from mature. There exists much room for the study of facial attractiveness. We hope the technology of artificial intelligence and image processing can provide advanced tools and strong support for other disciplines'development of facial attractiveness research, and also, help the realization of the machine intelligent perception of facial beauty in the near future.