Wisdom (public number: zhidxcom)
Edit | Zhidong West Interior
Since the second half of the 20th century, computer vision technologies have gradually developed. At the same time, with the wide use of digital images, hardware and software technology related to digital images, digital images have become an important constituent of contemporary social information sources, and the demand and application of various image processing and analysis continue to promote innovation. The application of computer vision technology is very wide. Digital image retrieval management, medical imaging analysis, intelligent security, human-machine interaction, etc. have a computer vision technology. This technology is an important part of artificial intelligence technology and is also the frontier area of today’s computer scientific research. After continuous development in recent years, a set of digital signal processing technology has been gradually formed. Computer graphics images, information theory, and semantic mutual combination of comprehensive technologies, and has strong marginal and discipline. Among them, face detection and identification of current image processing, pattern identification, and computer visual inside, is also a branch that is most concerned by people in biometric identification.
Face recognition is a biometric identification technology identified based on human face feature information. Usually collected images or video streams with a camera or camera, and automatically detect and track people in the image. According to the information, the 2017 biological identification technology has risen to $ 17.2 billion, in 2020, it is expected that the scale of biometric market in the world may reach $ 24 billion. Since 2015 to 2020, the size of the face recognition increased by 166.6%, which increased first in many biological identification technology, and it is expected that the market size of the face recognition technology will rise to $ 2.4 billion by the 2020 people’s face recognition technology.
In this issue, we recommend the research report of Aminro’s large data mining project from Associate Professor Tangjie Leaders, explaining people’s face recognition technology and its application field, introducing people in domestic playing talents in the field of face recognition and predicting the development of this technology. trend.
If you want to collect the full text of this article (Aminer: Face Identification Research Report), you can reply to the Zhidong West header number to get the keyword “NC296” acquisition.
1. Human face recognition technology overview
1. Basic concept
The unique charm of the human visual system drives researchers trying to simulate the acquisition, processing, analysis and learning capabilities of human beings through visual sensors and computer hardware and software to make computer and robot systems intelligently visual function. In the past 30 years, scientists in many different fields have continually tried to understand the mystery of biological vision and nervous system from multiple perspectives to benefit humanity with their research results. Since the second half of the 20th century, computer vision technology gradually develops in this context. At the same time, with the wide use of digital images, hardware and software technology related to digital images, digital images have become an important constituent of contemporary social information sources, and the demand and application of various image processing and analysis continue to promote innovation.
The application of computer vision technology is very wide. Digital image retrieval management, medical imaging analysis, intelligent security, human-machine interaction, etc. have a computer vision technology. This technology is an important part of artificial intelligence technology and is also the frontier area of today’s computer scientific research. After the continuous development of in recent years, a comprehensive technology that has been gradually combined with digital signal processing technology, computer graphic image, information theory and semantics, and has strong marginality and discipline. Among them, face detection and identification of current image processing, pattern identification, and computer visual inside, is also a branch that is most concerned by people in biometric identification.
Face recognition is a biometric identification technology identified based on human face feature information. Usually collected images or video streams with a camera or camera, and automatically detect and track people in the image. According to China Report Network, “2018 China Biological Recognition Market Analysis Report – Industry Depth Analysis and Development Prospect Prediction”, 2017 Biological Recognition Technology Global Market has risen to $ 17.2 billion, and by 2020, it is expected that biometric identity around the world is expected. Market size may reach $ 24 billion. Since 2015 to 2020, the size of the face recognition increased by 166.6%, which increased first in many biological identification technology, and it is expected that the market size of the face recognition technology will rise to $ 2.4 billion by the 2020 people’s face recognition technology.
In different biometric identification methods, face recognition has its own special advantages, and there is an important position in biometric identification. Five advantages of face recognition:
Face recognition does not need to interfere with people’s normal behavior to achieve recognition effect, and it is not necessary to worry whether the recognizers are willing to put their hands on fingerprint acquisition devices, whether their eyes can be aligned with iris scanning devices. As long as you naturally stay in a moment before the camera, the user’s identity will be correctly identified.
The collection device is simple, and it is quick to use. In general, common cameras can be used to capture the collection of face images, no particularly complex dedicated equipment. Image acquisition can be done in seconds.
Through face recognition identity, people’s habits, people and machines can be identified using face pictures. There is no characteristic of fingerprint, iris, and a person who has not been specially trained and cannot be identified by other people using fingerprints and iris images.
. The collection of face image information is different from the acquisition of fingerprint information. It needs to contact the collecting device with fingerprint acquisition. It is neither hygienic and easy to cause the user’s dissent, and the face image collection, users do not need to contact the equipment. .
After face recognition, the processing and application of the next step determines the practical application of face recognition equipment, such as applications in access control, face pictures, upper and lower hands swipe, terrorist identification, etc., scalability powerful.
It is because face recognition has these good features, which makes it very widely used in application prospects, which is also attracting more and more attention to the academic community and business. Face recognition has been widely used in identity identification, living body detection, lip language identification, creative camera, human face beauty, social platform and other scenes.
2, development history
As early as the 1950s, cognitive scientists have started research on face recognition. In the 1960s, the application research of face recognition engineering was officially opened. The method at the time mainly used the geometric structure of the face, and identified by analyzing people’s face organ feature points and the topology relationship therefrom. This method is simple and intuitive, but once the face posture, the expression changes, the accuracy is seriously reduced.
In 1991, the famous “Eigenface” method first introduced primary component analysis and statistical characteristics into facial recognition, and made a long progress in practical effects. This idea also has further carried forward in subsequent studies, for example, Belhumer successfully applies Fisher discriminant criteria to face classification, and proposes a FisherFace method based on linear discrimination analysis.
The first decade of the 21st century, with the development of machine learning theory, scholars successively explored the genetic algorithm, support vector machine (SVM), Boosting, flow learning, and nuclear methods, etc. From 2009 to 2012, sparse expresentation (sparse representation) is a research hotspot at the time because of its robustness of the shutter. At the same time, the industry also has basically reached a consensus: feature extraction and sub-spatial methods based on the local description of artificial design, feature the best identification effect.
The Gabor and LBP feature descriptors are the most successful two artificial design parties in the field of face recognition. During this period, the targeted treatment of various face recognition factors is also the research hotspot at that phase, such as human face lighting, face posture correction, face super residual, and occlusion processing.
At this stage, the researchers’ attention will begin to transfer from face recognition from the restricted scenario to face recognition in the non-restricted environment. LFW face recognition public competition (LFW is a public face number set by the University of Massachusedu University, the test data size is 10) started to popularize in this context, the best recognition system, despite the limited FRGC test set Can get 99% identification accuracy, but the highest precision on the LFW is only about 80%, and the distance from the practical look is much farther away.
Microsoft Asian Research Institute researchers have tried 100,000 US-scale training data, based on high-dimensional LBP characteristics and Joint Bayesian methods to obtain 95.17% accuracy on LFW. This result indicates that the big training data set is important to effectively improve face recognition in non-restricted environments. However, all of these classic methods are difficult to handle the training scenes of large-scale data sets.
Before and after 2014, with the development of big data and depth learning, the neural network recovered attention, and obtained the results of the distant classic method in applications such as image classification, handwritten identification, speech recognition. Sun Yi et al., Hong Kong University, proposes to apply the convolivable neural network to face recognition, using 200,000 training data, the first time exceeding the human level in the history of human level, this is the history of face recognition A milestone. Since then, researchers continue to improve the network structure while expanding the training sample size, pushing the identification accuracy on the LFW to 99.5%. Some classic methods in the development process of face recognition and their accuracy on the LFW have a basic trend: the scale of training data is getting bigger and bigger, and the recognition accuracy is getting higher and higher.
▲ Human face recognition technology development course
3, China Policy Support
Since 2015, the national intensive has introduced “Guidance Opinions on Banking Financial Institutions (Draft for Comment),” System Technical Requirements At the same time, in 2017, the labor intelligence first writes the national government report, as an important segment of artificial intelligence, the state’s policy support related to face recognition is constantly increasing. The “Promoting the Three-Year Action Plan (2018-2020) of the New Generation of Artificial Intelligence Industry Development in December 2017 (2018-2020)” to 2020, the effective detection rate of face recognition in complex dynamic fields exceeds 97%, the correct recognition rate exceeds 90%.
▲ Face identification related policies
4, development hotspot
Through the excavation of past recognition fields, the research keywords in the field of face recognition are summarized, and the keywords of face recognition, feature extraction, sparse representation, image classification, neural network, target detection, face image, face Detection, image representation, computer vision, posture estimate, face confirmation and other fields.
The figure below is an analysis of the trend of face recognition research, which is designed to conduct research on technical sources, heat or even development trends based on historical research results data. In Fig. 2, each color branch represents a keyword area, where the width represents the research heat of the keyword, and the position of the keywords in each year is to sort the height of all the key words on this time. At first, Computer Vision (computer vision) was the hotspot of research. In the end of the 20th century, Feature Extraction exceeded CV, became a new hotspot of research, and later was more than Face Recognition at the beginning of the 21st century, so far in the second Position.
▲ Face recognition related hotspots
In addition, the research is discovered in accordance with the keywords extracted in the paper published in FG (International Conference On Automatic Face and Gesture Recognition), Face Recognition has the highest frequency, 118 times, Object Detection is second, 41 Occasion, Image Classification and Object RecoGnition also have 36 words more than ten times, there are image segmentation (32), Action Recognition (32), Sparse RePresentation (28), Image Retrieval (27), Visual Tracking. 24), SINGLEIMAGE (23). The word cloud is shown below:
▲ Face recognition word cloud analysis
5, face recognition related conference
Three top international conferences in computer vision (CV)
ICCV: IEEE International Conference On Computer Vision
The meeting was hosted by the IEE, IEE, Institute of Electrical & Electronic Engineers, mainly in Europe, Asia, and the United States. As the top academic conference, the first international computer vision conference was unveiled in London in 1987, and the first session was held two years. ICCV is the highest level of conference in computer visual fields, and the conference’s paper set represents the latest development direction and level in the field of computer visual. The thesis acceptance rate is around 20%. The direction is computer vision, pattern identification, multimedia calculation, and so on.
In recent years, the global academic community is more interesting to pay attention to the research achievements of the Chinese people in the field of computer visual fields, because the relevant research led by Chinese has achieved a long-term progress – a total of more than 1200 papers received a total of more than 1200 papers. The selection papers are only 244, of which more than 30 papers from mainland China, Hong Kong and Taiwan have more than 12% of the total number of papers. As the first Chinese team invested in deep learning technology, on the basis of the key technology of many years of layout, the team led by Tang Xiaou led by the Professor of Hong Kong’s Chinese University quickly achieved technical breakthroughs. Only two deep learning articles on the 2012 International Calculation Vision and Mode Identification Conference (CVPR) were from the Tang Xiaou Lab, while 8 deep learning fields of global scholars on the 2013 International Computer Visual Conference (ICCV) In the article, there were 6 out of the Tang Xiaolan Lab.
CVPR: IEEE Conference On Computer Vision and Pattern Recognition
The meeting is a top meeting in the field of computer vision and pattern identification. Held once a year, the admission rate is about 25%. The direction is computer vision, pattern identification, multimedia calculation, and so on.
The team of Tang Xiaolan, professor Tang Xiao, led a large number of deep learning original technologies in the world: 2012 International Calculation Vision and Mode Identification Conference (CVPR) is from its laboratory; 2011 – 14 deep learning papers were published on the two top conferences in the computer visual field, and the total number of deep learning papers in these two conferences (29) were occupied by the world. He won the Best Paper Award in the two top international academic conferences in the computer visual field in 2009, which was the first prize from Asia’s papers in CVPR.
ECCV: EUROPEAN Conference On Computer Vision
ECCV is a European conference, and each meeting is around 300 papers worldwide. The main employment papers are from top laboratories and research institutes such as the United States, Europe, and the number of papers in mainland China is generally between 10-20 articles. The ECCV2010 papers an admission rate is 27%. One year in two years, the reciprocation of the paper is about 20%. The direction is computer vision, pattern identification, multimedia calculation, and so on. 2018 ECCV was held in Munich, September 8, 2018.
Asian computer visual conference:
Accv: Asian Conference On Computer Vision
ACCV is the Asian Computer Visual Conference, which is the two-year meeting of the AFCV (Asian Federation of Computer Vision, Asian Computer Vision) since 1993, aiming to provide a good platform for researchers, developers and participants. To display and discuss new issues, new programs and new technologies in computer visual fields and related fields. The 14th Asian Computer Vision Conference on 2018 will be held in Australia from December 4, 2018.
Face and gesture identify specialized meeting:
FG: IEEE International Conference On Automatic Face and Gesture Recognition
“International Conference On Automatic Face and Gesture Recognition” is an authoritative academic conference in the field of face and gesture identification worldwide. There are face detection, face recognition, expression identification, posture analysis, psychological behavior analysis, etc.
Second, face recognition technology detailed
1, face recognition process
The principle of face recognition technology is mainly three major steps: First, establish a database containing a large number of face images, and the other is to obtain the target face image of the current identification through a variety of ways, and the third is to target people. The face image is compared and screening existing face images in the database. According to the technical process implemented according to the principles of face recognition, the following four parts, namely, the collection and pre-processing, face detection, face characteristics extraction, face recognition and living body identification.
▲ Face recognition technology process
Collection and pre-treatment of face images
The acquisition and detection of a face image can be specifically divided into the acquisition of face images and the detection of face images.
Collection of face images: There are two ways to collect face images, which are both bulk import and real-time acquisition of face images. Some advanced face recognition systems can even support the conditional filtering that does not conform to face recognition quality requirements or low-quality face images, and do clearly and accurate acquisition as much as possible. Batch import of existing face images: It is about to collect a face image batch through a variety of ways to face recognition system, and the system will automatically complete the acquisition of personal face images. Real-time acquisition of face images: Tune camera or camera automatically capture face images in the device’s shooting range and complete the collection work.
Pre-processing of face images: The purpose of pretreatment of face images is to make further processing of face images on the system of face images to facilitate the feature extraction of face images. The pretreatment of face images is specifically a series of complex processing processes such as light, rotation, cutting, filtration, noise reduction, and reduction of the human face image collected by the system to make the face image from light. , Angle, distance, and size, etc., the standard requirements capable of conforming to the characteristics of the face image. The image is collected in a realistic environment. Since the image is different from light, the interference of the face expression, the shadow occlusion, etc., resulting in the quality of the collection image, then you need to preprocess the collected image, if image Pre-treatment is not good, will seriously affect subsequent face detection and identification. Study introduces three image pretreatment means, namely grayscale adjustment, image filtering, and image size normalization.
Gray adjustment: Because the final image of face image processing is generally a binarized image, and due to differences in locations, equipment, lighting, etc., resulting in uniform grayscale processing of the image to the image. To smoothing these differences. The common method of grayscale adjustment has average value method, a histogram transform method, a power transformation method, and a logarithmic transformation method.
Image filtering: During the actual face image acquisition process, the quality of the face image will be affected by various noise, these noise derived from a plurality of aspects, such as a large number of electromagnetic signals in the surrounding environment, digital image transmission is electromagnetic signals Interference and the like affect the channel, which in turn affects the quality of the face image. In order to ensure the quality of the image, reduce the impact of the noise on the subsequent processing, the image must be noise reduction processing. The principle and method of removing noise processing, common mean filter, median filtering, etc. Currently used median filtering algorithms for pretreatment of face images.
Image Size normalization: When a simple face training, when the image pixel size of a human face is different, we need to do size to the image before the endorsement is more than the identification. It is necessary to compare a common dimensional normalized algorithm has a double linear interpolation algorithm, nearest neighborographic algorithm and cubic volume algorithm.
A picture containing a face image usually may also contain other content, at this time, it is necessary to perform the necessary face detection. That is, in a face image, the system will accurately position the position and size of the face, and automatically remove other excess image information while selecting useful image information to further guarantee the precision collection of face images. .
Face detection is an important part of face recognition. Face detection refers to the application of a certain policy to be retrieved, and it is determined whether there is a face, if there is a process of positioning each face, size and posture. Face detection is a challenging target testing problem, mainly in two aspects: the inherent changes in human face targets: 1 ), Different faces have different appearances, such as face shape, skin color, etc. 2, the face of face, such as glasses, hair and head ornaments, etc. External conditions Changes: 1. Due to the difference in the imaging angle, the multi-posture of the face, such as rotation, depth rotation, and upper rotation, where the depth rotation has a large influence; 2. The influence of the light, such as the brightness in the image , Contrast change and shadow, etc .; 3, image conditions, such as the focal length of the image pickup apparatus, imaging distance, and the like.
The role of face detection is to be in a face image, and the system will accurately position the position and size of the face, and automatically remove other excess image information while picking out useful image information. Empress the precise collection of face images. Face detection focuses on the following indicators:
Detection rate: Identify the correct face / figure in the figure. The higher the detection rate, the better the detection model effect; the error rate: Identify the wrong face / recognition of the face. The lower the error rate, the better the detection model; the leak detection rate: all the faces of the face / figure in the figure. The lower the leak detection rate, the better the detection model; speed: time from the acquisition image to the time of face detection. The shorter the time, the better the detection model.
The current face detection method can be divided into three categories, which are based on the detection of the skin color model, based on the detection of edge characteristics, based on the statistical theory method, the following will be briefly introduced:
1. Detection of skin color model: Different modeling methods can be used when skin color is used for face detection, mainly with Gaussian model, Gaussian mixed model, and non-parameter estimate. The use of the Gaussian model and the Gaussian mixed model can establish a skin color model in different color spaces to perform face detection. By extracting the facial area in the color image to achieve a variety of illuminated methods, the algorithm needs to be effective in the premise of fixed camera parameters. Scholars such as Comaniciu use non-parametric nuclear function probability density estimation method to establish a skin color model, and use the Mean-Shift method to achieve local search to achieve the detection and tracking of the face. This method improves the detection speed of the face, and there is a certain robustness for occlusion and light. The shortcomings of this method are not very high and other methods are not very high, and there is difficult to handle complex backgrounds and multiple people’s faces.
In order to solve the illumination problem in face detection, it can be compensated for different light, and then the skin color area in the image is detected. This can solve the problem of polarization, background complex and multi-faceted face in color images, but is insensitive to face color, position, scale, rotation, posture and expression.
2. Detection based on edge characteristics: When detecting a face characteristics of an image, the amount of calculation is relatively small, and real-time detection can be realized. Most algorithms that use edge features are based on the edge of the face, which matches the established template (such as an elliptical template). There are also researchers to use the elliptical ring model and the edge direction characteristics to achieve a simple background of face detection. Fröba et al. Adopts a method based on edge-orientation matching, eom, in the edge direction map. The algorithm is relatively high in the complex background, but it can achieve good results after fusion with other features.
3. Based on statistical theory: This article focuses on the Adaboost face detection algorithm based on statistical theory. The AdaBoost algorithm is the process of seeking the optimal classifier by countless cyclic iterations. In any feature of the HAAR characteristics of the weak classifier, it is placed in a face-of-foot-based value. The characteristic characteristics of the face characteristic, through more classifier, to distinguish between people’s face and non-face . The haar function consists of some simple black white vertical or rotated 45 ° rectangular. The current HAAR feature is generally widely divided into three categories: edge characteristics, line characteristics, and central characteristics.
This algorithm is proposed by PAUL VIOLA and Michael Jones, Cambridge University, the algorithm has not only calculated speed, but also meets the performance and other algorithms, so it is relatively wide in people’s face testing, but also exists Have a higher error rate. Because in the process of using AdaBoost algorithm, there are always some faces and non-face mode difficult to distinguish, and there are some windows that are not similar to face mode in the results of their tests.
Face characteristic extraction
At present, the mainstream face recognition system can support usage can usually be divided into face visual features, face image pixel statistical features, and the characteristic extraction of the face image is to extract some specific features on the face. The characteristics are simple, and the matching algorithm is simple, suitable for large-scale construction libraries; it is suitable for small scale libraries. The method of feature extraction typically includes a knowledge-based extraction method or an extraction method based on algebraic features.
Taking one of the knowledge-based face recognition extraction method, because the face is mainly composed of eyes, forehead, nose, ear, chin, mouth, etc., and the structural relationship between these parts and their structure can be The geometric shape features are described, that is, everyone’s face images can have a corresponding geometric shape feature, which can help us as an important difference character for identifying people, which is based on knowledge-based extraction methods. One.
We can set a value of a face similarity in a face recognition system, and then compare the corresponding face image with all face images in the system database. If the preset similar value is exceeded, the system will Will output more than the face image, at this time, we need to accurately screen according to the similar level of face images and the identity information of the face itself. This precise screening process can be divided into two categories: one It is a one-to-one screening, that is, a confirmation process for face identity; its second is a pair of screenings, that is, the process of matching comparison according to face similarities.
One of the common problems identified by biometric identification is to distinguish whether the signal comes from a true organism, for example, the fingerprint recognition system needs to distinguish the identified fingerprint from the human finger or fingerprint gloves, the face recognition system is collected. Face image is from a real face or a photo of a face. Therefore, the actual face recognition system generally needs to increase the living identification link, for example, requires people to turn around, blink, open the mouth, etc.
2, face recognition main method
The research of face recognition technology is a high-end technical research work across multiple disciplines, including multi-discipline expertise, such as image processing, physiology, psychology, pattern identification. In the field of face recognition technology research, there are currently several directions in research, such as: a method of identifying statistics based on face characteristics, its main method and hidden Markov model (HMM, Hidden Markov Model Method, etc .; another face recognition method is about the connection mechanism, mainly artificial neural network (Ann, Artificial NETWORK) method and support vector machine (SVM, Support Vector Machine) method, etc. It is a method of integrated a variety of identification methods.
Method based on feature face
The method of the characteristic face is a relatively classic and widely used face recognition method. Its main principle is to make the image to reduce the dimensional algorithm, making the data processing easier, while the speed is faster. The face recognition method of the characteristic face is actually transforms the image to transform the image, converts a high-dimensional vector into a low-dimensional vector, thereby eliminating the correlation between each component, so that the image corresponding to the conversion The feature value is decremented. After the image is converted through K-L, it has good displacement invariance and stability. Therefore, the face recognition method of the feature face is convenient, and can be faster, and the recognition rate of the front face image is equivalent. However, this method also has insufficient places, which is more likely to be affected by factors such as face expressions, posture, and illumination changes, resulting in low recognition rate.
Method based on geometric characteristics
The identification method based on geometric character is a face recognition method according to the characteristics of facial facial organ and its geometric shape, which is the first recognition method for people to study and use, which is mainly used by different characteristics of different facets. Matching identification, this algorithm has a faster recognition speed, while the memory occupied is relatively small, but its identification is not high. The method is mainly to detect the position and size of the human face’s mouth, nose, eyes and other face main feature organs, and then use these organs to match the geometric distribution and ratios to meet face recognition.
Based on geometric character recognition is generally as follows: First, each feature point and position of the face face, such as nose, mouth and eye, and then calculate the distance between these features, and obtain each characteristic face. Vector feature information, such as the position of the eye, the length of the eyebrow, etc., followed by calculating the corresponding relationship between each feature, comparing the corresponding feature information in the face database, and finally draws the best matching person . The method based on geometric character is in line with people’s understanding of human face characteristics. In addition, each face is only stored, so the space occupied is relatively small; at the same time, this method does not reduce its identification rate, Moreover, the matching and recognition rate of the feature template is relatively high. However, the method based on geometric characteristics also has a good robustness. Once the expression and attitude are slightly changed, the identification effect will be greatly reduced.
Depth learning method
The emergence of deep learning makes human face recognition techniques have made breakthrough progress. The latest research results of face recognition have shown that deep learning, the characteristics of human face characteristics have no important characteristics of manual characteristics, such as moderately sparse, have strong selectivity to face identity and face properties. It has good robustness to local occlusion. These features are naturally obtained through large data training, and the model is not added to explicit constraints or post-processing, which is the main reason for deep learning to successfully apply in face recognition.
Deepest learning is a typical application of seven aspects in face recognition: a face recognition method based on convolutional neural network (CNN), depth nonlinear face shape extraction method, deep learning-based face gesture robust modeling A fully automatic face recognition in the constraint environment, a face recognition under deep learning video surveillance, based on deep learning low-resolution human face identification and other recognition of deep learning-based face-based information.
Among them, convolutional neural networks, CNNs is the first learning algorithm for truly successfully training multi-layer network structure, a human face recognition method based on convolutional neural network is a machine learning model under depth supervision learning. Mining data local feature, extract global training characteristics and classification, whose weight sharing structure network makes it more similar to biological neural network, and has been successfully applied in various fields in pattern identification. CNN makes full utilization of locality such as data itself by combining partial perceived region, sharing weight, in space or time slimming samples, and optimizing the model structure to ensure a certain displacement invariance.
With the CNN model, the DEEP ID project of Hong Kong Chinese University and Facebook’s DEEP FACE project on the LFW database is 97.45% and 97.35%, respectively, is slightly lower than the correct rate of 97.5% more than human visual recognition. After getting breakthrough results, the DEEPID2 project of the Chinese University of Hong Kong increased the recognition rate to 99.15%. DEEP ID2 The variation of the class in the class is minimized by learning nonlinear feature transformation, while maintaining the distance between the face images of different identities, more than the current recognition rate of all leading depth learning and non-depth learning algorithms in the LFW database. And the recognition rate of human beings in the database. Deep learning has become a research hotspot in computer vision, and the new algorithm and new direction of deep learning are constantly emerging, and the performance of deep learning algorithms has gradually exceeded shallow learning algorithms in some international major evaluation competitions.
Method based on support vector machine
Applying the support vector (SVM) method to face recognition originated from statistical theory, the direction of its research is how to construct effective learning machines and use to solve the classification of patterns. It is characterized by categorizing image transform spaces in other spaces.
Supporting vector machines is relatively simple, and can achieve global optimal features, so support vector machine has achieved extensive applications in the field of face recognition. However, this method also has the same shortcomings as the neural network, which requires a large storage space, and the training speed is relatively slow.
Other comprehensive methods
The above several common face recognition methods, we are not difficult to see that each identification method can do the perfect recognition rate and faster recognition speed, there are their own advantages and disadvantages, so many researchers now It is preferred to use a variety of identification methods to integrate application, take advantage of various identification methods, comprehensively use to achieve higher recognition rates and identification effects.
Face recognition three classic algorithm
Feature Face Law (Eigenface)
Scrubbing technology is a recent development for face or general rigid body identification and other methods involving face treatment. Methods using feature faces are first submitted by Sirovich and Kirby (1987) (“Low-Dimensional Procedure for the Characterization of Human Faces”) and is used by Matthew Turk and Alex Pentland for face classification (“Eigenfaces for Recognition) “). First, a group of face images is converted into a feature vector set, called “Eigenfaces”, “feature face”, which is the basic component of the original training image set. The process of identifying is to project a new image to the feature face space and determine and identify by its projection point in the position of the sub-space and the length of the proof line.
After transforming the image to another, the image of the same category will be together, and the images of different categories will be far more than that of different categories of images in the original category space, which is difficult to cut, transform, transform To another space, you can separate them very well. The spatial conversion method selected by the Eigenface is PCA (main component analysis), using the PCA to get the main component of the face distribution, the specific implementation is the resolution matrix of the covariance matrix of all face images in the training set, and obtain the corresponding ethical vector These intrinsic vectors are “feature faces”. Each feature vector or feature face is equivalent to capturing or describing a variation or characteristic between face. This means that each face can represent a linear combination of these feature faces.
Local binary pattern (Local Binary Patterns, LBP)
Local Binary Patterns LBP is a visual operator for classification in the computer visual field. LBP is an operator for describing image texture characteristics, which is submitted by T.Ojala, Olu, Finland, tipped in 1996 (“a Comparrative Study ON Featured Distributions). In 2002, T.OJALA et al. Published an article on LBP (“MultiResolution Gray-Scale and Rotation Invariant Texture Classification With Local Binary Patterns”). This article is clearly described with a multi-resolution, gray scale and improved LBP characteristics of the rotation, equivalent mode. The core idea of the LBP is to be a threshold with the gray value of the central pixel, and the corresponding binary code is obtained in the field to represent the local texture characteristics.
LBP is based on the basis for extracting local features. The significant advantage of the LBP method is that the illumination is not sensitive, but there is still no problem with posture and expression. However, compared to the characteristic face method, the identification rate of LBP has improved greatly.
Linear identification analysis considers category information while despising the design, by the statistian SIR R. A. Fisher1936 (“The Use of Multiple Measurements in TaxOom”). In order to find a characteristic combination, the maximum type interval dispersion and the smallest depersion degree are reached. This idea is very simple: under low dimension, the same class should be tightly together, and the farther distance from different categories. In 1997, Belhumer successfully applied Fisher discriminant criteria to face classification, and proposed a FisherFace method based on linear discrimination analysis (“Eigenfaces vs. Fisherfaces: Recognition Using Class Specific Linear Projection”).
Sirovich, L., & Kirby, M. (1987) .LOW-DIMENSIONAL Procedure for the Characterization of Human Faces.Josa A, 4 (3), 519-524. Research proves that any special person can be called EigenPictures The coordinate system is expressed. EigenPictures is an intrinsic function for the average covariance of the facial collection.
Turk, M., & Pentland, a. (1991). Egenfaces for Recognition.journal of Cognitive Neuroscience, 3 (1), 71-86. Research has developed a close-time computer system, which can locate and track people’s heads Then, then identify the person by comparing the characteristics of the face features and the features known individuals. This method treats facial recognition as a two-dimensional identification issue. The identified process is to project a new image to the feature face space, which captures significant changes between known face images. Important features are feature faces because they are the feature vectors of the face set.
Ojala, T., Pietikäinen, M., & Harwood, D. (1996) .a Comparrative Study of Texture MeasureS with Classification Based On Featured Distributions.Pattern Recognition, 29 (1), 51-59. Research on different graphics textures Comparison, and proposed an LBP operator used to describe image texture characteristics.
Ojala, T., Pietikainen, M., & Maenpaa, T. (2002) .Multiresolution gray-scale and rotation invariant texture classification with local binary patterns.IEEE Transactions on pattern analysis and machine intelligence, 24 (7), 971-987. The study proposed a very simple and effective gradation and rotational uncarreled texture classification method, which is based on the non-parametric discrimination of localized two-value modes and prototype distributions. This method has a steady gradation and simple computation.
Fisher, RA (1936) .The Use of Multiple Measurements in TaxOomics, 7 (2), 179-188. Study finds a characteristic combination to achieve the largest inter-zoning degree and the smallest class Dispersion. The solution is: Under the low dimension, the same class should be tightly together, and the different categories are so far.
Belhumeur, Pn, Hespanha, JP, & Kriegman, DJ (1997). Egenfaces vs.fisherfaces: recognition using class specific linear project. Yale University New Haven United States. Research on Fisher’s linear discrimination for facial projections, can be in low-dimensional space A well-separated class is produced, even in large changes in light and facial expression. The extensive experimental results show that the error rate of the proposed “Fisherface” method is lower than the characteristic facial features of Harvard and Yale face database test.
Common human face database
Mainly introduces several common face databases:
ERET face database
Creremely created by the Feret project, contains a large number of face images and only one face in each picture. This focus, the same photo of the same person has different expressions, light, posture, and age. A face image containing more than 10,000 multi-posture and illumination is one of the most widely used face databases in the field of face recognition. Most of them are Western, and the changing changes in the face images included in each person are single.
CMU Multi-Pie Face Database
Established by the US Carnegie Mellon University. The so-called “PIE” is an abbreviation for posture (pose), illumination and expression. The CMU Multi-Pie Face Database is developed on the basis of the CMU-PIE face database. Contains more than 75,000 multi-postures, light and facial images of 337 volunteers. The posture and light change image is also collected under strict control, and it has gradually become an important test collection in the field of face recognition.
Yale Human Face Database (USA, Yale)
Create a visual and control center by Yale University, including 165 pictures of 15 volunteers, including changes in illumination, expressions, and postures.
Yale Human Face Database 10 samples, compared to the sample collected in the ORL face database Yale library contains more obvious illumination, expressions, and postures, and occlusion.
Yale Human Face Database B
The 5,850 images of 10 people include 9 postures, 64 lighting conditions. The image of the gesture and light change is collected under strict control, mainly used for the modeling and analysis of light and posture problems. Due to few collected people, further application of the database has been relatively large.
MIT face database
It is created by the Massachusetts Institute of Technology Media Lab, including 1692 different attitudes (27 photos per person), light and size facial images.
ORL face database
Created by the AT & T laboratory of the UK Cambridge, including 40 people with a total of 400 facial images, and some volunteers include changes in gestures, expressions, and facial ornaments. The human face is often adopted in the early stage of face recognition research, but due to few variations, the identification rate of most systems can reach more than 90%, so the value of further utilization is not large.
Each acquisition target in the ORL face database, each acquisition object contains 10 normalized grayscale images, and the image size is 92 × 112, and the image background is black. The facial expressions and details of the acquisition objects have changed, such as laughing and not laughing, the eyes are smiling or closed, and the gestures of different face samples have changed, their depth rotation and flat rotation can reach 20 Spend.
BioID face database
Contains 1521 gray surface images in various illumination and complex backgrounds, the eye position has been manually labeled.
Umist image set
Established by the University of Manchester, England. Includes 20 people with a total of 564 images, each person has a variety of images of different angles, different attitudes.
Age identification data set IMDB-Wiki
Contains 524,230 celebrity data pictures crawling from IMDB and Wikipedia. Applied a novel age algorithm for classification. The essence is a 101 class classification between 0-100, and the resulting score and 0-100 are multiplied and the final result is summed, and the age of final identification is obtained.
Third, technical talents
1, scholar profile
Aminer calculates the Chinese version of the TOP1000 scholars in the field of face recognition, and draws the global distribution map of people’s face recognition. From a global perspective, the United States is a human face recognition study. The most aggregated countries have absolute advantages in the field of face recognition; the British is followed, and the second column is the third, and the land is located. Some talents have also gathered in Canada, Germany and Japan.
▲ Human face recognition scholar TOP1000 global distribution map
▲ Face recognition expert country number ranking
▲ Human face recognizes global scholar H-Index statistics
H-Index: Internationally recognized index that can accurately reflect the academic achievements of scholars, the calculation method is that the scholars have at least H disciplines, respectively.
Global face recognition scholar’s H-Index average number is 48, the H-Index index is between 20 and 40, accounting for 33%; H-Index Index is between 40 and 60 scholars and greater than 60 proportion Without, the former is 27%, the latter is 28%; the H-Index index is less than the least equal to 10 scholars, accounting for only 2%.
▲ People’s face recognition global talents migration
Aminer selection of expert scholars ‘scholars’ scholars in the field of human face recognition, analyzed their migration path. As can be seen from the above figure, the loss and introduction of people in people’s face recognition are slightly different, of which the US is a large-scale mobile big country in the field of face recognition, the talent input and output is greatly leading, and the income from data is slightly larger than flowing out. . Other countries such as Britain, China, Germany, Canada and Australia have followed, in which the Britain, China and Australia have a slight talent loss.
Studies are based on the previous five-year reference in the IEEE International Conference on Automatic Face and Gesture Recognition, FG. Face recognition expert and intercepting some lead scholars.
The Citation Row in the top ten columns are as follows:
▲ Ten people face recognition experts in Citation
The top twelve-related scholars in the top of the H-Index is listed below:
▲ H-Index top ten face recognition experts
2, domestic and foreign talents
The report enumerates 6 experts and 5 domestic experts in the world, see this intravenous attachment.
Fourth, application field
From an application perspective, face recognition is widely used, and can be applied to automatic access control system, identification of identification, bank ATM cash machine, and home security. Specifically, mainly:
1, public safety:
Public security criminal investigation, criminal identification, border security inspection;
2, information security:
Computer and network login, file encryption and decryption;
3. Government functions:
E-government, household registration management, social welfare and insurance;
4, commercial enterprises:
E-commerce, electronic currency and payment, attendance, marketing;
5, the place entrance:
Military machine department department, financial institution’s access control and access management.
1, access control person face recognition
With the improvement of people’s living standards, people pay more attention to the safety of home environment, and the security concept is constantly being strengthened; with this increase in this demand, the intelligent access control system has emerged, more and more companies, shops, families have installed each Various access control systems.
The access control system that is commonly used is not adjacent video access control, password access control, radio frequency access control or fingerprint access control, etc. Among them, video access control is simply transmitted to the user, there is no intelligence, essentially inherently inquiry “human defense”, users do not absolutely protect home security when they are not present; password access control is easy, password easy Forgot, and easy to crack; the shortcomings of radio access control are “recognition cards”, and RF card is easy to lose and easy to be used by others; in addition, the safety hazard of fingerprint access control is easy to copy. Therefore, there is a problem that the above access control system provided in the prior art has a problem of lower security. The face recognition system is installed, as long as the camera is revealed, it can easily go into the community and truly realize “brush card”. Biological identification access control systems do not need to carry verification media, and verify feature has uniqueness and is extremely good. Currently widely used in locations with high confidential levels, such as research institutes, banks, etc.
Facial identification technology has two main applications on marketing: First, you can identify basic personal information of a person, such as gender, approximate age, and what they have seen, how long have you been. Outdoor advertising company, such as Val Morgan Outdoor (VMO), starts with facial recognition techniques to collect consumer data. Second, this technology can be used to identify known individuals, such as thieves, or members who have already joined the system. This application has attracted some service providers and retailers.
In addition, facial recognition techniques can improve the effects of advertising, and allow advertisers to respond promptly to consumers. VMO has introduced a measuring tool Dart, which can see the direction of consumer eyes in real time as well as long, so that they can judge their level of attention to an advertisement. The next generation of DART will also include more demographic information, except for age, including consumers when watching a digital signage.
3, commercial bank
Using face recognition technology to prevent network risks: For my country’s widely used magnetic strip bank card, although technology mature, standard, but production technology is not complex, bank magnetic stripe card tracking standards are open secret, only one computer and A magnetic strip reader can successfully “clone” bank card. In addition, the sales management of the card machine is not strict. If the criminals use bank card fraud cases, the main means is to “clone” or steal bank cards through a variety of ways. At present, all commercial banks have also taken some technical means to prevent forgery and clone cards, such as using CVV (Check Value Verify) technology, generating a set of checksures while generating card magnetic strip information, the check value and each The characteristics of the card itself are associated, thus achieving the function of copying. Although a variety of measures have been taken, the defects inherent in the magnetic stripe card itself have seriously threatened the interests of our customers. For these bank network security issues, we can use face recognition technology to prevent network risks. Face recognition technology is to capture people’s face areas through image acquisition devices, and then match the captured faces and people in the database to complete identification tasks. Use face recognition technology to accurately recognize the true identity of cardholders, and ensure that the cardholder’s funding is safe. In addition, it is also possible to further lock the criminals through face recognition technology, which is advantageous to quickly solve the case.
The application of face recognition technology in governance counterfeit banknotes: At present, my country’s commercial banks exist in self-service equipment: First, some self-service equipment installation does not meet the requirements. Some self-service installations of commercial banks did not work with ground reinforcement of equipment in accordance with the requirements of the public security department; some electrical environments did not meet the requirements: Some did not set up 110 connection alarms or no visual monitoring alarm, some monitoring records are not clear enough The monitoring video storage time has not reached the specified requirements, and the other equipment is seriously destroyed. The second is the design of the self-service equipment software design. In particular, some domestic equipment software is not reasonable enough, and the software changes are random, there is a loophole, causing the wrong possibility. The third is that there is no counterfeit differential equipment in the Bank’s ATM machine. Due to the problems in my country’s commercial banks in self-service equipment, the counterfeit banknotes are endless. Since there is no counterfeit identification equipment in the Bank’s ATM machine, it is only true that the clever personnel have been identified before cash, and such measures are not perfect, and it is easy to cause disputes between banks and cardholders. Even the cash deposit (CRS) has a counterfeit value, but is often utilized by the crosses due to the lag of the fake-based identification feature. The criminals are first deposited into the counterfeit banknotes, and then immediately extract the true bills on the counter or other self-service equipment.
Five, future trends
Overall, the trend of face recognition includes the following aspects.
1, machine identification combined with manual identification
At present, some face recognition companies on the market are tested at home and abroad, and the accuracy of their face recognition can generally reach more than 95%, and the speed of precise face recognition is also very fast. This also provides powerful practical prove from the side to face recognition technology into practical applications.
However, in the actual life, everyone’s face is not kept moving relative to the camera. In contrast, it is in a high-speed motion state, and the face image collected by the camera will be due to his face’s posture. Expression, light, decoration, etc. It may not be possible to achieve fast and accurate face recognition.
Therefore, after setting a certain degree of face image similarity, the face recognition company system prompts a face image higher than the similarity value, and then screens themselves one by one, using machine identification and artificial Identify the combined manner to maximize accurate recognition of face images.
2,3D Application of Face Identification Technology
Whether it is a face image that has been saved in the mainstream face image database or a face image collected by the street, the face image is collected in real time, and most of them are actually a 2D face image. The 2D face image itself has an inherent defect, which is that it cannot be deeply expressing face image information. It is particularly vulnerable to light, posture, expression and other factors during shooting. For people’s face, the face of the face includes the eyes, nose, ear, chin, etc., is not in a plane. The face has a stereo effect, and the shooting 2D face image is not very good to reflect. All key features of the face face.
In 2017, the iPhone X is equipped with a smart phone with many of the latest cutting-edge technology, which causes great attention to the industry. Among them, it is a black technology: 3D face unlocking function, namely Face ID, a new identity authentication method. When unlocked, the user only needs to look at the phone, and the Face ID can realize a face recognition unlock.
▲ Apple layout in the 3D visual field
Apple iPhone X joins the 3D facial recognition function is not a cardiogramine, because it has started layout in the 3D visual field in 2010. Especially in 2013, Apple acquired the 3D Visual Company PrimeSense of Israel with $ 345 million. This acquisition is one of the biggest handles of Apple’s history. Since then, Apple has also invested in some 3D visual technologies and face recognition technology companies.
In addition, the FACE ID can also be used in Apple Pay and third-party applications. For example, Apple uses Face ID to upgrade the EMOJI function, you can use the Face ID to create 3D expressions Animojis, which can use animation to express emotions, but current features can only be used in Apple’s own iMessage. This way to “brush face” brings to the user’s more real human-machine interaction experience.
3. Wide application of face recognition technology based on deep learning
Most of the mainstream face recognition technology is a lightweight face image database, which is not much mature for future fully presence billion-level face image database, so it is necessary to focus on the deep learning-based face. Identification technology.
In terms of common sense, there are more than 13 billion people in the country, and the strong face recognition company takes the lead in the future to establish a unified face image database that covers nationwide unified face image database, then the person The face image of the face image database stored in the face image may reach billions or even billions of levels. At this time, there will be a large number of people who are similar, key feature points, if there is no deep learning-based face Identification technology, establish a more complex and diverse face model, then it will be more difficult to achieve precise and fast face recognition.
4, the substance of the face image database
A face image database with excellent diversity and versatility is also an inevitable thing. It is mainly compared with the database referenced by the current mainstream face recognition company. It is mainly reflected in the following aspects: First, The improvement of the face image database will increase from now 100,000 yuan to the future billion or even hundreds of billions; the second is the improvement of the level, will be upgraded by the mainstream 2D face image to each The key feature point is more obvious and clear 3D face image; the third is the type of face image, will collect each person’s face image under different attitude, expression, light, ornament, etc. Enrich everyone’s face metrics and precise face recognition.
It is believed that face recognition is a wide range of applications that have developed faster AI technology and more applications. At this year’s Ambo, face recognition and dynamic capture technology, almost become “standard” of each exhibitor. With the research and development of national scientific research institutions, the company’s research, the promotion of the market, and the face recognition will usher in a better development wave. Future face recognition or becomes a valid identity to identify mainstream, when the face recognition is not a new word.