Feature-based method is a bottom-up face detection method. Because human eyes can treat faces differently, researchers believe that there is a potential assumption that faces or parts of faces may have features or attributes that will not change under various conditions, such as shape, skin color, texture and edge information. The goal of feature-based method is to find these invariant features and use them to locate faces. This method is very effective, with fast detection speed in specific environment, and insensitive to face posture, expression and rotation. However, because the extraction of face parts usually depends on edge operators, such methods require high image quality, illumination and background, because illumination, noise and shadows are likely to destroy the edge of face parts, thus affecting the effectiveness of the algorithm.
Template matching algorithm first needs people TN as standard template (fixed template) or template parameterization (variable template). Then, when detecting the face, the correlation value between the input image and the template is calculated, which is usually a comprehensive description obtained by independently calculating the matching degree of the face contour, eyes, nose and mouth. Finally, whether there is a face in the image is determined according to the correlation value and the preset threshold. The face detection algorithm based on variable template is much better than the fixed template algorithm, but it still cannot effectively deal with the changes of face scale, posture and shape.
The method based on appearance shape does not need complex preprocessing of the input image, nor does it need to manually analyze facial features or extract templates. Instead, we use specific methods (such as principal component analysis (PCA), support vector machine (SVM), neural network method (ANN) and so on. ) to train a large number of face and non-face samples (generally, the capacity of non-face sample set is more than twice that of face sample set to ensure the accuracy of the trained detector). Therefore, this is also a J bottom-up method. The advantage of this method is that it uses powerful machine learning algorithm to achieve good detection results quickly and stably, and it can also get effective detection results in multi-pose face images under complex background. However, this method usually needs to traverse the whole picture to get the detection results, and it needs a large number of face and non-face samples and a long training time in the training process. In recent years, the research on face detection based on this method is relatively active.
Face recognition method based on algebraic features
In face recognition based on algebraic features, each face image is regarded as a matrix with pixel gray level as the element, and the features of face are represented by data features reflecting some properties. Let's assume a face image), (y x I is a two-dimensional N M × gray image, which can also be regarded as N M × = dimensional column vector and can be regarded as a point in N M × dimensional space. But in such a space, not every part of the space contains valuable information, so generally speaking, these points in such a huge space need to be mapped to a space with lower dimensions through some transformations. Then, the similarity between images is determined by some metrics between image projections, the most common being various distance metrics. Among the face recognition methods based on algebraic features, Principal Component Analysis (PCA) and Fisher Linear Discriminant Analysis (LDA) are the most studied methods. This chapter briefly introduces PCA.
The application of complete PCA(PrincipalComponentAnalysis) face recognition includes four steps: face image preprocessing; Reading a face database and training to form a feature subspace; Projecting the training image and the test image onto the subspace obtained in the previous step; Select a distance function to identify. The detailed description is as follows:
4. 1 Read into the face database
After the face database is normalized, everyone in the database selects a certain number of images to form a training set. Let the normalized images be n×n, and connect them in columns to form an n2-dimensional vector, which can be regarded as a point in n2-dimensional space. This image can be described by a low-dimensional subspace through K-L transformation.
4.2 calculate the generating matrix of k.l transformation.
The total deviation matrix of the training sample set is the generating matrix, i.e.
Or write:
Where xi is the image vector of the ith training sample, |l is the average vector of the training samples, and m is the total number of training samples. In order to find the eigenvalues and orthogonal normalized eigenvectors of n2×n2 dimensional matrix, the calculation amount is too large, so the singular value decomposition theorem is introduced to solve the high-dimensional problem.
4.3 Use the Singular Value Decomposition (AVD) theorem to calculate the eigenvalues and eigenvectors of the image.
Let A be a row of n×r-dimensional matrices with rank r, then there are two orthogonal matrices and diagonal matrices:
Wherein the two orthogonal matrices and diagonal matrices satisfy the following formula:
Where is the nonzero eigenvalue of the matrix,
4.4 Project the training image and the test image into the feature space, and project each pair of face images into the feature face subspace to obtain a set of coordinate coefficients corresponding to a point in the subspace. Similarly, any point in the subspace also corresponds to the sub-image. This set of coefficients can be used as the basis of face recognition, that is, the characteristic face features of this face image. That is to say, any face image can be expressed as a linear combination of this group of characteristic faces, and each weighting coefficient is the expansion coefficient of K. L transform, which can be used as the recognition feature of the image and represent the position of the image in the subspace, that is, the vector.
It can be used for face detection. If it is greater than a certain threshold, F can be considered as a face image, otherwise it is not. In this way, the original problem of face image recognition is transformed into the problem of classification according to the training sample points in the subspace.
Face recognition method based on connection mechanism
Representative identification methods based on connection mechanism include neural network and elastic matching method.
Neural network is a research hotspot in the field of artificial intelligence in recent years, and facial feature extraction and recognition based on neural network technology is a positive research direction. Neural network forms a complex system by connecting a large number of simple neurons, and has achieved good results in face recognition, especially in frontal face images. Commonly used neural networks include BP network, convolution network, radial basis function network, self-organizing network and fuzzy neural network. BP network has a small amount of calculation and short time consumption, and its adaptive function enhances the robustness of the system. Compared with other methods, neural network can obtain the implicit expression of recognition rules, but its disadvantages are long training time, large amount of calculation, slow convergence speed and easy to fall into local minimum. Gutta et al. combined the hybrid classifier model of RBF and tree classifier for face recognition. Lin et al. used virtual samples for reinforcement and anti-reinforcement learning, and used modular network structure to accelerate learning, and realized the neural network method based on probability decision, and achieved ideal results. This method can be applied to all steps of face detection and recognition. Elastic matching method uses attribute topological graph to represent human face, and each vertex of topological graph contains a feature vector to record the feature information of human face around the vertex position. The vertices of the topological graph are represented by wavelet transform, which has certain adaptability to light, angle and size, and can adapt to the changes of expression and perspective. It theoretically improves some shortcomings of eigenface algorithm.
Face recognition method based on three-dimensional data
A complete face recognition system includes three parts: face data acquisition, data analysis and processing and final result output. Figure 2- 1 shows the basic steps of 3D face recognition: 1, and the 3D shape information of the face is obtained by 3D data acquisition equipment; 2. Preprocessing the obtained 3D data, such as smooth denoising and facial region extraction; 3. Extract facial features from 3D data and compare them with the data in the face database; 4. Use the classifier to make classification judgment and output the final decision result.
Representative methods based on three-dimensional data include model synthesis method and curvature method.
The basic idea of the method based on model synthesis is to input a two-dimensional face image, recover (or partially recover) the three-dimensional information of the face with some technology, and then re-synthesize the face image under specified conditions. Typical examples are 3D deformable model and 3D enhanced face recognition algorithm based on shape recovery. 3D deformable model Firstly, a deformable 3D face model is constructed from 200 high-precision 3D face models, and a set of specific parameters are obtained by fitting a given face image with this model, and then a face image with arbitrary posture and illumination is synthesized. The 3D enhanced face recognition algorithm based on shape recovery is to synthesize a new face image by using a general 3D face model, and the synthesis process changes certain posture and light source conditions.
Curvature is the most basic local feature to express surface information, so it is the curvature of face surface that is first used to deal with 3D face recognition. Lee and lJ use the average curvature and Gaussian curvature values to segment convex regions in the face depth map.
If you are a developer, you can go to Tel and colorreco to have a look and get better technical answers.