A Nonlinear Feature Extraction Method And Its Applications In Image Processing And Face Recognition

Face recognition is a relevant subject in image processing, pattern recognition and machine intelligence. Over the last decades, face recognition remains as an unsolved problem and a demanded technology which can be widely used in national defense, public safety, civil application and economic construction, as well as computer science and many other fields. For a long time, most classical face-feature selection methods use the discriminant rule based on geometric distance as the feature evaluation criterion. Owing to the changing of facial expression and variations in gesture, and the affection of various factors such as illumination and background, the intra-class difference of the features extracted by these methods is usually greater than their inter-class difference, so that using the inter-class difference to classify different faces become very difficult. In addition, the traditional recognition methods (e.g. eigenface and fisherface) are linear methods. They cannot extract the nonlinear structural information of the face image, so they have the essential limitations.In order to effectively extract the key features of the face images for accurate classification, we study the pattern representation, training and classification problems of the pattern recognition system in depth. Then, using the self-organization and information theory, combining with statistical analysis and stochastic approximation method, we derive a new Probabilistic Self-Organizing Network (PSON) model. The PSON has the outstanding property of probability density estimation. It can extract the key features from the complex data correctly and classify the non-linear data accurately. On this basis, we put forward a variety of effective face recognition and detection methods based on the probability separability criterion. The main research work and contributions of this dissertation are as follows:1. Study the feature selection and statistical classification methods in depth, then closely integrate the information theory, maximum likelihood estimation, stochastic approximation, and Self-Organizing neural network to derive the Probabilistic Self-Organizing Network model. The PSON can be truly applied to inhomogeneous mixtures, so it can greatly improve the estimation accuracy. And since it is stochastic gradient descent, it can effectively overcome the local minima and achieve global optimal (or approximate global optimal) solution. Besides, the PSON is a nonlinear feature extraction method. It clusters the input data but does not compress their dimension, therefore has distinct advantages in terms of feature extraction. Our researches establish a well-defined objective function for the PSON and provide a general proof of convergence for the algorithm. We also provide a new interpretation for the competition and cooperation mechanisms of the Self-Organizing Map from the statistical point of view. The validity of the PSON is fully proved through a wide range of probability estimation, pattern classification, and topology mapping experiments. Experiments also show that the PSON is far superior to the expectation-maximization (EM) algorithm in terms of estimation accuracy and computational efficiency. We give this a reasonable explanation by using the stochastic approximation and neural network theory. In artificial intelligence and machine learning, the EM algorithm is the basis of many clustering algorithm, but it has a high possibility of being trapped in local optima and is very computationally expensive with large data sets. Experimental results show that when used to estimate large datasets, the PSON is30-80times faster than the EM algorithm at least. In many cases, it can completely replace the EM algorithm.2. Using the PSON to model the skin and non-skin pixels accurately and effectively, and then put forward a skin detection method based on the Bayesian decision rule to classify the pixels correctly. This detection method can adapt to the changes in the lighting and the viewing environments, greatly reduce the storage requirements and the computational complexity, and effectively improve the detection performance. An image database which contains a great part of challenging images with diversity of human races, variable ambient lights, confusing backgrounds, and also various resolutions and visual quality is employed to train and evaluate the method objectively. The typical detection results are True Positive Rate (TPR)=80%with False Positive Rate (FPR)=8.2%, or TPR=90%with FPR=14.2%. These results are comparable with histogram model-based classifiers and outperform the GMM-based classifiers, but our method is much superior in storage requirements and computation performance.3. Propose a new face recognition method based on probabilistic separability criterion. This method uses the PSON to accurately estimate the intrinsic distribution of each face image from its limited training pixels, and puts forward several fast algorithms to calculate the overlapping degree between different probability functions. It effectively solves the discriminant issues based on probability features. The proposed method has the following characteristics:(1) since the probability distribution reflects the intrinsic property of patterns, this method needs not conduct explicit feature extraction or transformation, and needs less prior knowledge of the underlying categories;(2) in practical applications, attribute measurement will always introduce some random errors, so using the probability theory to solve the recognition problem will be more reasonable and reliable;(3) its recognition result has a direct connection to the probability of classification error, so it can obtain statistical optimal solutions according certain decision criteria;(4) using the probabilistic difference to measure the separability between different categories can overcome the limitations of the Euclidean distance effectively;(5) its density estimation algorithm has implicitly fulfilled the performance of non-linear transformation, so it can take full advantage of the higher-order correlation of the original data to improve the separability of each categories or the accuracy of classification significantly. In strict accordance with the test principle of FERET, a large number of experiments are performed on the FR Data database to fully verify the feasibility and effectiveness of the proposed method. Experimental results show that usually the top1match rates of this method are up to95%, and the first five cumulative match scores will reach100%. These results are much better than those of eigenface. Various experiments prove that this method exhibits high robustness against variant expressions, partial occlusions, rotations, and scaling. It also has excellent performance in one image per person problem.