“he came to my concert, and the tickets changed the handcuffs”. Recently, the zhang dynasty turned into an officer, who often caught criminals at concerts and put face recognition in the public eye again. When the human face recognition technology broke out last september, after the release of apple iphone x, fingerprints were no longer needed and the phone was easily unlocked by scanning the face. As soon as any technology enters the consumer market of smartphones, especially when it is introduced by the brand of apples, it means that it will become a trend, a branding of smart devices。
In these years of rapid growth of smartphones, their password locks have been upgraded from digital passwords, hand gestures to fingerprint recognition, to today's iris recognition and face recognition. In the near future, as a result of full screen coverage and the emergence of more secure and accessible faceid technology, fingerprint recognition will also be abandoned by smartphone manufacturers to fulfil their historic mission。
So what kind of technology is human face recognition, and what kind of technology is this。
What's face recognition?
The human face recognition technique is a biometric identification technique based on information on the characteristics of the human face. Images or video streams containing human faces are collected using cameras or cameras and automatically detect and track human faces in images, leading to a series of techniques related to the face of human faces that are detected, commonly known as human image recognition and facial recognition. Traditional human face recognition techniques, based mainly on visible photo images, are also familiar ways of identifying. Simply speaking, it's a process of making computers recognize you。
Human face recognition techniques are mainly carried out through the extraction and comparison of image characteristics. The human face recognition system matches the feature data of the human face images extracted with the feature templates stored in the database, and by setting a threshold, when similarity exceeds that threshold, the results matched are exported. The profile of the face of a person to be identified is compared with the template of the face of a person that has been received, and the identity of the face is judged on the basis of similarity. This process is divided into two categories: one-to-one image comparison and one-to-one image matching。
Human face recognition in a broad sense includes a range of relevant techniques for building a human face identification system, including face collection, face profiling, face recognition pre-processing, identification and identification; in a narrow sense, human face recognition refers specifically to a technique or system of identification or identification through a person's face。
Development of human face recognition technology
As early as the 1950s, cognitive scientists began researching the identification of human faces. In the 1960s, applied research on the engineering of human face recognition was officially launched. The methodology at that time was based on the geometric structure of the human face, which was identified by analysing the characterization of the human face organ and the relationship between it and its expansion. This method is simple and intuitive, but once the human face is changed, the accuracy is significantly reduced。
In 1991, for the first time ever, the well-known “identified face” approach introduced major component analysis and statistical profiling techniques into human face recognition, making great strides in practical effects. This idea has also been further developed in follow-up studies, for example, belhumer has succeeded in applying the fisher standard to the classification of human faces, proposing a fisherface approach based on linear analysis。
The first four characteristic vectors derived from the characterization of the cambridge face dataset
In the first decade of the twenty-first century, as the theory of machine learning developed, scholars explored human face recognition based on genetic algorithms, support to vector machines (support victor machine, svm), boosting, fluid learning and nuclear methods. Between 2009 and 2012, the thin expression became the hotspot of the study because of its excellent theory and its greatness vis-à-vis the resistance。
At the same time, there is a general consensus in the industry that the best effects of identification can be achieved by character extraction and sub-spatial methods based on locally designed and artificially well-designed descriptions. Gabor and lbp characterizations are two of the most successful manual local descriptions in the field of human face recognition to date. During this period, the targeted treatment of impact factors for the identification of human faces was also the hot spot for that phase of research, such as the integration of face light, the correction of human face postures, the over-resolution of human faces and the blocking of treatment. Also at this stage, the focus of researchers began to shift from face recognition in restricted settings to face recognition in non-restricted settings. The lfw face recognition open competition began to be popular in this context, when the best identification system, despite obtaining more than 99 per cent recognition precision on the restricted frgc test set, was only 80 per cent of the maximum precision on the lfw, which seemed to be far from practical。
For the first time in 2013, msra researchers tried large-scale training data of 100,000 and obtained 95. 17 per cent precision on the lfw based on high-dimensional lbp features and the joan bayesian methodology. This finding indicates that a large training data set is important for effective face recognition in non-restricted environments. However, all of the above classic approaches make it difficult to handle training scenarios for large-scale data sets。
Around 2014, with the development of big data and deep learning, the nervous network gained prominence and obtained results from far more than classic methods in image classification, handwritten recognition and speech recognition applications. Sun yi of the chinese university of hong kong, among others, proposed the application of a congested neural network to human face recognition, using 200,000 training data and obtaining for the first time in the lfw recognition accuracy above human level, a milestone in the history of human face recognition development。
Since then, researchers have continuously improved the network structure, while scaling up the training sample to over 99. 5 per cent accuracy of identification on the lfw. As shown in table 1, we give some of the classic methods of human face recognition development and their precision on the lfw, a basic trend being the increasing scale of training data and the increasing accuracy of identification。
Ten key techniques for face recognition
1. Face detection
The role of face detection is to detect the location of the human face in the image。
The human face detection algorithm was entered as an image, with the output being a sequence of human face frames, with the result being zero or one or more faces. The resulting human face frame can be square, rectangle, etc。
The rationale for human face testing algorithms is simply a "scan" plus "judgment" process. That is, the process of scanning the entire image and then deciding whether the candidate area is human face or not. So the calculation speed of the face detection algorithm is associated with the size of the image and its content. In practical terms, we can accelerate algorithms by setting the "input image size" or "minimum face size limit" and "maxim human face size limit"。
Example: the green rectangular frame indicates the human face detection algorithm
2. Face alignment
The aim of the "face alignment" was to locate the coordinates of the five key points on the human face。
The human face is matched by the algorithmed input of a “one face image” and a “person face coordinates frame”, and the output is a sequence of coordinates of the five key points. The number of five key points is a fixed number set in advance, with five points, 68 points, 90 points, etc. Common。
Some of the better face matching techniques of the current effect are largely achieved through an in-depth learning framework. These methods are based on a frame of coordinates for the human face detection, which removes the area of the human face from the surface, downsizes it to a fixed size and then calculates the critical point position. In addition, the calculation of human face matching algorithms is much less time-consuming than the human face test or the process of extracting the human face characteristics to be mentioned later。
Example: input images and output results are as follows, and green dots are marked with five official positions。
3 face attribute
"face attribute" is a technique for identifying properties such as sex, age, attitude, and expression. This has been used in some cameras ap to automatically identify and mark the gender, age, etc. Of the person in the camera view。
The normal human face properties recognition algorithm is entered as a " one face map " and " five key points of the human face " and the output is the attribute value corresponding to the human face. Face recognition algorithms usually align a person's face according to the coordinates of the five key points of the person's face. The process involves rotation, scaling, extraction, etc., and the person's face is adjusted to the intended size and form to allow subsequent attribute analysis。
The properties of the face include gender recognition, age estimates, face recognition, attitude recognition, hair recognition, etc. Generally, the process of identifying algorithms for each attribute is independent, but there are a number of new algorithms based on in-depth learning to achieve simultaneous output of attributes such as age, sex, attitude, expression, etc。
Example: the results of the human face properties recognition output are as follows:
4. Facefeature analysis
"face feature extration" is the transformation of an image of a person's face into a characteristic that can characterize a person's face, expressed in a set of values of fixed length。
The entry for the human face feature process is the “a human face map” and the “five key points of the human face”, the output being a numerical string corresponding to the human face. The human face-to-face feature algorithm is achieved by first rotating the coordinates of the five key points, scaling them up, etc., to align the human face, then extracting the features and calculating a numerical string。
Example: human face profiling process
5 face company
The “face company” algorithm is designed to measure similarities between the faces of two people。
The human face comparison input is the two-face feature human face characteristics obtained from the front-face characterization algorithm and the output is the similarity between the two characteristics。
Example: human face comparison process with a similar 96% output
Face verification
Face verification is an algorithm for determining whether the faces of two individuals are the same person。
Its input is a face-to-face feature, which is comparable to the acquisition of the face-to-face characteristics of two persons by comparing them with a predetermined threshold to verify whether the face-to-face characteristics belong to the same person。
Example: the face verification process was as follows, with a similarity of 96 per cent above the threshold of 75 per cent and was judged to be the same person
