Face recognition (Face Recognition) is a biometric technology that automatically recognizes people based on facial features (such as statistical or geometric features, etc.) , facial recognition, etc. Usually what we call face recognition is the abbreviation of identity recognition and verification based on optical face images.
The machine vision
solution service provider Longruizhike (www.loongv.com) believes that face recognition uses cameras or cameras to collect images or video streams containing faces, and automatically detects and tracks faces in the images. A series of related application operations are performed on the face image. Technically, it includes image acquisition, feature location, identity confirmation and search, and so on. Simply put, it is to extract the features of the face from the photo, such as the height of the eyebrows, the corners of the mouth, etc., and then output the results through the comparison of the features.
The general process of face recognition: 1) Face collection: (1) Introduction: Different face images are collected through the camera lens, such as static images, dynamic images, different positions, different expressions, etc. When within the shooting range, the acquisition device will automatically search for and shoot face images. (2) The main influencing factors of face collection: image size: too small a face image will affect the recognition effect, and too large a face image will affect the recognition speed. Non-professional face recognition cameras commonly stipulate that the minimum face recognition pixels are 60*60 or 100*100 or more. Within the specified image size, the algorithm is more likely to improve the precision and recall. The image size reflected in the actual application scenario is the distance between the face and the camera. Image Resolution: The lower the image resolution, the harder it is to identify. The image size, combined with the image resolution, directly affects the recognition distance of the camera. At present, the farthest distance for a 4K camera to see a face is 10 meters, and a 7K camera is 20 meters. Lighting environment: Overexposure or too dark lighting environment will affect the face recognition effect. You can use the built-in function of the camera to fill in the light or filter the light to balance the lighting effects, or you can use the algorithm model to optimize the image light. Blur level: The actual scene mainly focuses on motion blur, and the movement of the face relative to the camera often produces motion blur. Some cameras have anti-blur function, and in the case of limited cost, consider optimizing this problem through an algorithm model. Degree of occlusion: The best image is an image with no occlusion of facial features and clear face edges. In the actual scene, many faces are blocked by hats, glasses, masks and other occlusions. This part of the data needs to be reserved for training according to the requirements of the algorithm.
Collection angle: The angle of the face relative to the camera is the best. However, it is often difficult to capture the front face in actual scenes. Therefore, the algorithm model needs to be trained on data including left and right faces and upper and lower faces. The angle of camera placement in industrial construction must meet the requirements that the angle formed by the face and the camera is within the algorithm recognition range. 2) Face detection: (1) Introduction: Accurately calibrate the position and size of the face in the image, and pick out the useful information (such as histogram features, color features, template features, structural features and Haar features, etc. ), and then use the information to achieve the purpose of face detection. (2) Face key point detection (face alignment): Automatically estimate the coordinates of facial feature points on a face image. (3) Mainstream method: Based on the detected features, the Adaboost learning algorithm (a method for classification, which combines some weaker classification methods to combine a new strong classification method) selects some of the most powerful classification methods. The rectangular feature (weak classifier) that can represent the face, the weak classifier is constructed into a strong classifier according to the weighted voting method, and then several strong classifiers obtained by training are connected in series to form a cascaded classifier with a cascade structure, which is effective to improve the detection speed of the classifier. The recent genres of face detection algorithm models include three categories and their combinations: viola-jones framework (generally good performance, suitable for mobile and embedded use), dpm (slower), cnn (good performance) ). 3) Face image preprocessing: (1) Introduction: Based on the results of face detection, the image is processed and finally serves the process of feature extraction. (2) Reason: The original image obtained by the system cannot be used directly due to various limitations and random interference, and it must be pre-processed such as grayscale correction and noise filtering in the early stage of image processing. (3) The main preprocessing process:
Face alignment (to obtain an image with a correct face position), light compensation of the face image, grayscale transformation, histogram equalization, and normalization (to obtain a standardized face image with the same size and the same grayscale value range) , geometric correction, median filtering (smoothing the image to remove noise), and sharpening. 4) Face feature extraction: (1) Introduction: The features that can be used in face recognition systems are usually divided into visual features, pixel statistical features, face image transformation coefficient features, and face image algebraic features. Face feature extraction is based on some features of the face, also known as face representation, which is the process of modeling features of the face (2) Methods of face feature extraction: 1. Knowledge-based representation methods ( Mainly including based on geometric feature method and template matching method): According to the shape description of face organs and the distance characteristics between them, the feature data that is helpful for face classification is obtained, and its feature components usually include the Euclidean distance between feature points. , curvature, and angle, etc. The human face is composed of parts such as eyes, nose, mouth, chin, etc. The geometric description of these parts and the structural relationship between them can be used as an important feature to identify the face, and these features are called geometric features. 2. Representation method based on algebraic features or statistical learning: The basic idea of algebraic feature-based method is to convert the high-dimensional description of the face in the air domain into a low-dimensional description in the frequency domain or other spaces, and its representation method is linear projection representation. methods and nonlinear projection characterization methods. The methods based on linear projection mainly include principal component analysis or KL variation, independent component analysis and Fisher linear discriminant analysis. There are two important branches of nonlinear feature extraction methods: kernel-based feature extraction techniques and manifold learning-dominated feature extraction techniques. 5) Matching and recognition: The extracted face feature value data is searched and matched with the feature template stored in the database, and the identity information of the face is judged by setting a threshold and comparing the similarity with this threshold. .
The general process of face recognition: 1) Face collection: (1) Introduction: Different face images are collected through the camera lens, such as static images, dynamic images, different positions, different expressions, etc. When within the shooting range, the acquisition device will automatically search for and shoot face images. (2) The main influencing factors of face collection: image size: too small a face image will affect the recognition effect, and too large a face image will affect the recognition speed. Non-professional face recognition cameras commonly stipulate that the minimum face recognition pixels are 60*60 or 100*100 or more. Within the specified image size, the algorithm is more likely to improve the precision and recall. The image size reflected in the actual application scenario is the distance between the face and the camera. Image Resolution: The lower the image resolution, the harder it is to identify. The image size, combined with the image resolution, directly affects the recognition distance of the camera. At present, the farthest distance for a 4K camera to see a face is 10 meters, and a 7K camera is 20 meters. Lighting environment: Overexposure or too dark lighting environment will affect the face recognition effect. You can use the built-in function of the camera to fill in the light or filter the light to balance the lighting effects, or you can use the algorithm model to optimize the image light. Blur level: The actual scene mainly focuses on motion blur, and the movement of the face relative to the camera often produces motion blur. Some cameras have anti-blur function, and in the case of limited cost, consider optimizing this problem through an algorithm model. Degree of occlusion: The best image is an image with no occlusion of facial features and clear face edges. In the actual scene, many faces are blocked by hats, glasses, masks and other occlusions. This part of the data needs to be reserved for training according to the requirements of the algorithm.
Collection angle: The angle of the face relative to the camera is the best. However, it is often difficult to capture the front face in actual scenes. Therefore, the algorithm model needs to be trained on data including left and right faces and upper and lower faces. The angle of camera placement in industrial construction must meet the requirements that the angle formed by the face and the camera is within the algorithm recognition range. 2) Face detection: (1) Introduction: Accurately calibrate the position and size of the face in the image, and pick out the useful information (such as histogram features, color features, template features, structural features and Haar features, etc. ), and then use the information to achieve the purpose of face detection. (2) Face key point detection (face alignment): Automatically estimate the coordinates of facial feature points on a face image. (3) Mainstream method: Based on the detected features, the Adaboost learning algorithm (a method for classification, which combines some weaker classification methods to combine a new strong classification method) selects some of the most powerful classification methods. The rectangular feature (weak classifier) that can represent the face, the weak classifier is constructed into a strong classifier according to the weighted voting method, and then several strong classifiers obtained by training are connected in series to form a cascaded classifier with a cascade structure, which is effective to improve the detection speed of the classifier. The recent genres of face detection algorithm models include three categories and their combinations: viola-jones framework (generally good performance, suitable for mobile and embedded use), dpm (slower), cnn (good performance) ). 3) Face image preprocessing: (1) Introduction: Based on the results of face detection, the image is processed and finally serves the process of feature extraction. (2) Reason: The original image obtained by the system cannot be used directly due to various limitations and random interference, and it must be pre-processed such as grayscale correction and noise filtering in the early stage of image processing. (3) The main preprocessing process:
Face alignment (to obtain an image with a correct face position), light compensation of the face image, grayscale transformation, histogram equalization, and normalization (to obtain a standardized face image with the same size and the same grayscale value range) , geometric correction, median filtering (smoothing the image to remove noise), and sharpening. 4) Face feature extraction: (1) Introduction: The features that can be used in face recognition systems are usually divided into visual features, pixel statistical features, face image transformation coefficient features, and face image algebraic features. Face feature extraction is based on some features of the face, also known as face representation, which is the process of modeling features of the face (2) Methods of face feature extraction: 1. Knowledge-based representation methods ( Mainly including based on geometric feature method and template matching method): According to the shape description of face organs and the distance characteristics between them, the feature data that is helpful for face classification is obtained, and its feature components usually include the Euclidean distance between feature points. , curvature, and angle, etc. The human face is composed of parts such as eyes, nose, mouth, chin, etc. The geometric description of these parts and the structural relationship between them can be used as an important feature to identify the face, and these features are called geometric features. 2. Representation method based on algebraic features or statistical learning: The basic idea of algebraic feature-based method is to convert the high-dimensional description of the face in the air domain into a low-dimensional description in the frequency domain or other spaces, and its representation method is linear projection representation. methods and nonlinear projection characterization methods. The methods based on linear projection mainly include principal component analysis or KL variation, independent component analysis and Fisher linear discriminant analysis. There are two important branches of nonlinear feature extraction methods: kernel-based feature extraction techniques and manifold learning-dominated feature extraction techniques. 5) Matching and recognition: The extracted face feature value data is searched and matched with the feature template stored in the database, and the identity information of the face is judged by setting a threshold and comparing the similarity with this threshold. .