News
Industry news
The principles and challenges of face recognition technology
In recent years, with the continuous development of the security industry, intelligence is not a major development trends in the security industry. The so-called face recognition technology, which is based on characteristics of the human face, the human face of the input image or video stream to determine, first determine whether there is a face. If there is a human face, it is further given the position of each face, size, and location information of each of the major facial organs. And based on this information, the identity of each person to extract further implied in the face, and with a known face compared to identify the identity of each face. 
 
Principle of face recognition technology 
 
Face because of their easy capture features, the concern by many industrial customers, especially public security, customs, such as shopping malls. Human face recognition are doing every day, so most can accept this kind of authentication methods. Face Recognition began in the mid-century, after several decades of effort, you can now have been used in our real life, offers convenient for us. 
 
Face recognition is divided into face detection (face detection), feature extraction (feature extraction) and face recognition (face recognition) three processes. 
 
Face Detection: Face detection means to detect and extract a face image from the input image, usually haar features and Adaboost algorithm training cascade classifier each piece of the image classification. If a rectangular area by a cascade classifier, were discriminating human face images. 
 
Feature Extraction: Feature extraction is through a number of digital information to characterize the human face of these figures is that we want to extract features. Common facial feature divided into two categories, one is the geometric characteristics, and the other is the characterization of features. Refers to the geometric relationship between the geometric characteristics of the facial features of the eyes, nose and mouth, such as between distance, area and angle. Since the algorithm makes use of a number of intuitive features, small amount of calculation. However, due to its desired characteristics can not be accurately selected point, limiting its scope of application. In addition, when the light changes, face obscured foreign objects, when the facial expression changes, characterized by large changes. So, this kind of algorithm is only suitable for rough facial image recognition, can not be applied in practice. 
 
Characterized by the use of information characterizing the gray face images, global or local features extracted by some algorithm. The more commonly used feature extraction algorithm is LBP algorithm. LBP Firstly, the image is divided into several regions, each region in the 3x3 neighborhood of pixels with the center value of the field for threshold values, the result as a binary number. Figure 3 shows an LBP operator. LBP operator is characterized by a monotonous gray change remains unchanged. Get a set of histograms for each region through this operation, then all connected together to form one large histogram histogram and histogram matching calculation for classification. 
 
Face: Face mentioned here are narrow face recognition, face recognition will soon be extracted features and characteristics of the faces in the database comparison, according to the similarity in identification and classification. The face recognition can be divided into two categories: one is confirmed, this is an image of the man than the face image in the database already exists on the process, the answer that you are not your problem; the other is to identify this is the human face image matching process all the images in the database already exists, answering the question of who you are. Obviously, the difficulty of face recognition than the human face, because identifying the need to match mass data. Commonly used classification has nearest neighbor classifiers, support vector machines. 
 
A similar way with the fingerprint application, face recognition technology is relatively mature attendance. Because in the time and attendance system, the user is actively cooperate, you can get people to meet the requirements of the face in a particular environment. This gives recognition to provide a good input source can often be satisfied with the results. But in some public places to install video surveillance cameras, face image due to the light, angle problem is difficult to obtain than for success. This is one of the future challenges face recognition technology development must be addressed. 
 
Now there are some institutions, universities conducting new field of face recognition, research new technologies. Such as long-distance face recognition technology, 3D face recognition technology. Long-distance face recognition system faces two major difficulties. First, how to get a face image from a distance. Second, the data obtained are not in the ideal case to identify the identity. In a sense, the long-distance face recognition technology is not a specific key or basic research questions. It can be regarded as an application and system design problems. There are usually two types of solutions for acquiring face images. One is the definition of fixed cameras, and the other is the use of multi-camera systems PTZ control system. The latter is more suitable for general, but its structure is more complex, the cost is more expensive. The latter need to consider how to coordinate the simultaneous operation of multiple cameras. In general, the system consists of a low-resolution wide-angle camera and high resolution telephoto camera components. The former is used to detect and track the target, which is used for image acquisition and recognition of human faces. Currently long-distance face recognition technology is still in the laboratory stage, the future if we can solve the problem of personnel dispatched to such applications is of great significance. 
 
3D face recognition can be well overcome pose, illumination, facial expressions and other problems encountered in 2D face recognition. The main reason is a 2D image can not be a good representation of depth information. Typically, 3D face recognition method uses 3D scanning technology to obtain 3D face, and then build 3D face models and used to identify. However, the disadvantage of 3D face recognition technology is also very obvious. First, it needs to calculate the amount of additional 3D acquisition devices or binocular stereo vision technology, and secondly, the modeling process requires large. I believe that with the future development of chip technology, when computing power is no longer constrained by a substantial decline in the cost of acquisition devices, 3D face recognition will become one of the hottest technology. 
 
Face the challenge in the application 
 
From a practical test run, the gap between user expectations and the current level of technology among the still relatively large. Face recognition technology in the dynamic monitoring of the pressure faced by the application actually is relatively large. 
 
1. The user wants to correct alarm rates demanding. The reality is that in theory, have to accept a high false alarm rate. On the technical side, to achieve high correct alarm rates can be achieved by lowering the threshold, but the cost of lowering the threshold are: high false alarm rate. In order to achieve the correct alarm rate of 95%, many algorithms may produce 300% or higher rate of false positives. 
 
2. The user wants to monitor the database is large enough, often require tens or hundreds of thousands, or even millions of watch lists, hoping to catch the "big fish." The reality is that you have to accept large capacity high false alarm rate. 
 
3. The user wants to form a network of large-scale construction activities can sketch out the trajectory monitoring personnel. The reality is that high investment necessary to re-build private networks and associated hardware. 
 
4. The user wishes to make use of existing monitoring equipment (cameras and network). The reality is that the existing cameras lack of clarity, poor image quality for video surveillance scenes human face is too small, causing the network bandwidth is not enough, and so can not use the existing equipment. 
 
5. The user does not want to produce fewer false positives or false positives. The reality is so bound to the loss of the right to monitor alarm rate and reduce storage capacity, contrary to the idea of ​​the user. 
 
6. Light Problems 
 
Faced with a variety of test ambient light may appear side light, top light, back light and high-light and other phenomena, but there may be different light each time, even in the light of each position within the monitored area is different. 
 
7. Face Pose and ornaments issues 
 
Because monitoring is non-fit, and monitoring personnel to monitor the area through a natural posture through, so it may appear profile, bow, rise and other various non-positive attitude and wearing a hat face, black-rimmed glasses, masks and other accessories phenomenon. 
 
8. The camera's image problems 
 
Many technical parameter affecting the quality of the camera of the video image, these factors are photoreceptors (CCD, CMOS), the size of the photoreceptor, the processing speed of DSP, the built-in image processing chip and the lens, etc., while the built-in camera setting parameters also affect the number of video quality, such as exposure time, aperture, the dynamic balance and other parameters. 
 
9. dropped frames and humiliating problem 
 
Calculate the required network identification and recognition system may cause the video frame loss and disgrace phenomenon, especially to monitor high traffic areas, because of bandwidth issues and computing power network transmission problems, often caused by frame loss and disgrace.