Face recognition (Facial RecogniTIon) is to obtain the facial image of the user through the video capture device, and then use the core algorithm to calculate and analyze the facial features, face and angle of the face, and then compare with the existing templates in the database. Right, then determine the true identity of the user. Principle of face recognition algorithm: The system input is generally a series or a series of face images containing unidentified identities, as well as face images of certain known identities in the face database or corresponding codes, and the output is A series of similarity scores indicating the identity of the face to be identified.
Face Recognition Algorithm Analysis 1. Geometric Feature Based Method
The face is composed of eyes, nose, mouth, chin and other components. Because of the differences in shape, size and structure of these components, each face in the world is very different, so the geometric description of the shape and structure relationship of these components Can be used as an important feature of face recognition. The geometric feature was first used for the description and recognition of the side profile of the face. First, several significant points were determined according to the side profile curve, and a set of feature metrics such as distance, angle, etc. for identification were derived from these significant points. Jia et al. simulated the side profile from the integral projection near the center line of the front grayscale image is a very innovative approach.
The use of geometric features for frontal face recognition is generally based on the extraction of important features of the human eye, mouth, nose and other important organ geometry as the classification features, but Roder's accuracy of geometric feature extraction is experimental. The results of the study are not optimistic.
The deformable templating method can be regarded as an improvement of the geometric feature method. The basic idea is to design an organ model with adjustable parameters (ie, deformable template), define an energy function, and minimize the energy function by adjusting the model parameters. The model parameters at this time are taken as the geometric features of the organ.
This method is very good, but there are two problems. First, the weighting coefficients of various costs in the energy function can only be determined by experience, which is difficult to generalize. Second, the energy function optimization process is very time consuming and difficult to apply. Parameter-based face representation can achieve an efficient description of the salient features of the face, but it requires a lot of pre-processing and fine parameter selection. At the same time, the general geometric features only describe the basic shape and structure relationship of the components, ignoring the local fine features, resulting in the loss of some information, more suitable for rough classification, and the existing feature point detection technology in the accuracy rate Far from meeting the requirements, the amount of calculation is also large.
Face Recognition Algorithm Analysis 2. Local Feature Analysis (Local Face Analysis)
The representation of the principal subspace is compact, the feature dimension is greatly reduced, but it is non-localized, the support of its kernel function is extended in the entire coordinate space, and it is non-topological, the point adjacent to an axis projection. It has nothing to do with the proximity of points in the original image space, and locality and topologicality are ideal characteristics for pattern analysis and segmentation. It seems that this is more in line with the mechanism of neural information processing, so it is very important to find expressions with this characteristic. Based on this consideration, ATIck proposed a face feature extraction and recognition method based on local features. This method has achieved good results in practical applications, and it forms the basis of FaceIt's face recognition software.
Face Recognition Algorithm Analysis 3. Feature Face Method (Eigenface or PCA)
The feature face method is one of the most popular algorithms proposed by Turk and Pentland in the early 1990s. It has simple and effective features, also called face recognition method based on principal component analysis (PCA).
The basic idea of ​​eigenface technology is to find the basic elements of the face image distribution, that is, the feature vector of the covariance matrix of the face image sample set, from the statistical point of view, to approximate the face image. These feature vectors are called Eigenfaces.
In fact, the eigenface reflects the information that is implicit in the set of face samples and the structural relationship of the face. The feature vectors of the sample set covariance matrix of the eyes, cheeks, and lower jaws are called feature eyes, feature jaws, and feature lips, collectively referred to as feature face faces. The feature face generates a subspace in the corresponding image space, called a child face space. The projection distance of the test image window in the sub-face space is calculated, and if the window image satisfies the threshold comparison condition, it is determined to be a human face.
A method based on feature analysis, that is, a relative ratio of a face reference point and other shape parameters or class parameters describing a facial face feature are combined to form a recognition feature vector, and the overall face-based recognition not only retains the face portion The topological relationship between the pieces, but also retains the information of each component itself, and the component-based recognition is to design a specific recognition algorithm by extracting the local contour information and the gray information. Now Eigenface (PCA) algorithm has become the benchmark algorithm for testing the performance of face recognition system together with the classic template matching algorithm. Since the birth of feature face technology in 1991, researchers have carried out various experimental and theoretical analysis. The FERET'96 test results also show that the improved eigenface algorithm is the mainstream face recognition technology and one of the best performance recognition methods.
The method first determines the size, position, distance and other attributes of the facial iris, nose, mouth angle and the like, and then calculates their geometric feature quantities, and these feature quantities form a feature vector describing the image. The core of the technology is actually "local body feature analysis" and "graphic/neural recognition algorithm." This algorithm is a method that utilizes various organs and features of the human face. For example, the corresponding geometric relationship multi-data formation identification parameter is compared, judged and confirmed with all the original parameters in the database. Turk and Pentland propose a feature face method, which constructs a principal subspace according to a set of face training images. Since the principal has a shape of a face, also called a feature face, the test image is projected onto the principal subspace. A set of projection coefficients is obtained and compared with the face images of each known person. Pentland et al reported a fairly good result, with 95% correct recognition rate in 3,000 images of 200 people and only one misrecognition of 150 positive faces on the FERET database. However, the system needs a lot of pre-processing work such as normalization before performing the feature face method.
On the basis of the traditional feature face, the researchers notice that the feature vector with large feature value (ie, feature face) is not necessarily the direction of good classification performance, and accordingly, various feature (subspace) selection methods are developed, such as Peng's The double subspace method, Weng's linear ambiguity analysis method, Belhumeur's FisherFace method, and so on. In fact, the feature face method is an explicit principal component analysis face modeling. Some linear self-associative and linear compression BP networks are implicit principal component analysis methods. They all represent faces as some vectors. Weighted sums, these vectors are the main eigenvectors of the training set cross product matrix, which ValenTIn discusses in detail. In summary, the eigenface method is a simple, fast and practical algorithm based on transform coefficient features, but because it essentially depends on the gray correlation of the training set and the test set image, and requires the test image to be compared with the training set. So it has a lot of limitations.
Feature face recognition method based on KL transform
KL transform is an optimal orthogonal transform in image compression. It is used for statistical feature extraction, which forms the basis of subspace method pattern recognition. If KL transform is used for face recognition, it is assumed that people The face is in a low-dimensional linear space, and different faces are separable. Since a high-dimensional image space KL transform can obtain a new set of orthogonal bases, a low-dimensional face can be generated by retaining partial orthogonal bases. Space, and the basis of low-dimensional space is obtained by analyzing the statistical characteristics of the face training sample set. The generating matrix of the KL transform can be the overall scatter matrix of the training sample set, or the inter-class scatter matrix of the training sample set. The average of several images of the same person can be used for training, so that the interference of light and the like can be eliminated to some extent, and the calculation amount is also reduced, and the recognition rate is not lowered.
Face Recognition Algorithm Analysis 4. Elastic Model Based Method
Lades et al. proposed a dynamic link model (DLA) for object recognition of distortion invariance, describing objects with sparse graphics (see the following figure), their vertices are marked with multi-scale descriptions of local energy spectra, and edges represent topological connections. The relationship is marked with geometric distance and then plastic pattern matching techniques are applied to find the nearest known pattern. Wiscott et al. improved on this basis, using the FERET image library to do experiments, using 300 face images and another 300 images for comparison, the accuracy rate reached 97.3%. The disadvantage of this method is that the amount of calculation is very large.
Nastar models the face image (I ) (x, y) as a deformable 3D mesh surface (x, y, I(x, y) ) (as shown below), thereby transforming the face matching problem into Elastic matching problem for deformable surfaces. The surface deformation is performed by the method of finite element analysis, and it is judged whether the two pictures are the same person according to the deformation condition. This method is characterized by placing the space (x, y) and the gray scale I (x, y) in a 3D space and considering it. Experiments show that the recognition result is significantly better than the feature face method.
LaniTIs et al. proposed a flexible representation model method, which encodes the face into 83 model parameters by automatically locating the salient features of the face, and uses the method of discriminant analysis to perform face recognition based on shape information. Elastic image matching technology is a recognition algorithm based on geometric features and wavelet texture analysis for gray distribution information. Because the algorithm makes good use of face structure and gray distribution information, it also has automatic and precise positioning. The function of facial feature points has a good recognition effect and a high adaptive recognition rate. This technique is among the best in the FERET test. Its disadvantages are high time complexity, slow speed and complex implementation.
Face Recognition Algorithm Analysis 5. Neural Network Method (Neural Networks)
Artificial neural network is a nonlinear dynamic system with good self-organization and self-adaptation ability. At present, the research of neural network methods in face recognition is in the ascendant. Valentin proposed a method that first extracts 50 principals of a human face, then maps it to a 5-dimensional space using an autocorrelation neural network, and then uses a common multi-layer perceptron to discriminate, for some simple test image effects. Preferably, Intrator et al. proposed a hybrid neural network for face recognition, in which unsupervised neural networks are used for feature extraction and supervised neural networks are used for classification. Lee et al. described the characteristics of the face with six rules, and then based on these six rules to locate the five senses, the geometric distance between the five senses was input into the fuzzy neural network for identification, and the effect was better than the general method based on Euclidean distance. Laurence et al. used convolutional neural network to perform face recognition. Because the convolutional neural network integrates the correlation knowledge between adjacent pixels, the invariance of image translation, rotation and local deformation is obtained to some extent. Therefore, a very ideal recognition result is obtained. Lin et al. proposed a neural network method based on probability decision (PDBNN). The main idea is to use virtual (positive and negative) samples for reinforcement and anti-reinforcement learning, so as to obtain a better probability estimate. As a result, a modular network structure (OCON) is used to accelerate network learning. This method has been well applied in various steps of face detection, face localization and face recognition. Other studies include: Dai et al. proposed low-resolution face association and recognition using Hopfield network, Gutta et al. A hybrid classifier model combining face recognition with RBF and tree classifier is proposed. Phillips et al. use MatchingPursuit filter for face recognition. In China, the support vector machine in statistical learning theory is used for face classification.
The application of neural network method in face recognition has certain advantages over the above-mentioned methods, because it is quite difficult to describe the many rules or rules of face recognition, and the neural network method can be learned. The process obtains implicit expressions of these laws and rules, which are more adaptable and generally easier to implement. Therefore, artificial neural network recognition is fast, but the recognition rate is low. The neural network method usually needs to input the face as a one-dimensional vector, so the input node is huge, and one of the important targets for recognition is dimension reduction processing.
Algorithm Description of PCA: Identification using Principal Component Analysis (PCA) is proposed by Anderson and Kohonen. Since the PCA transforms the high-dimensional vector to the low-dimensional vector, the variance of each component of the low-dimensional vector is maximized, and the components are uncorrelated, so that optimal feature extraction can be achieved.
Kindwin Technology (H.K.) Limited , https://www.szktlled.com