1 Introduction
The local features of the human face are the unique "identity card" of the human body, and have the characteristics of "inseparable from the human". Face recognition technology is a biometric technology that uses some unique biometrics of a human face to automatically recognize a person's identity. It has the characteristics of direct concealment of face acquisition, friendliness, easy identification, high security, and portable, so it has broad application prospects in many fields such as public security, security, finance, and information network security.
In this paper, face recognition technology is applied to attendance management, and a dynamic face recognition attendance system (hereinafter referred to as face attendance system) is designed and implemented, which overcomes the traditional attendance methods such as magnetic cards, IC cards, etc., such as entrusted enrollment, card loss, etc. The shortcomings and defects effectively eliminate the human factor in the attendance management, and fully reflect the fairness and efficiency of the attendance management. Compared with the fingerprint attendance system, the face attendance system has the advantages of non-invasiveness and user-friendliness. At the same time, the collected frontal face images can be applied to other management information systems of unit employees.
2 Overview of the face attendance systemThe face attendance system is a successful product that applies advanced face recognition technology to the attendance field. The product first enters employee text data and dynamically collects facial images and trains them to establish a facial image database. During attendance, the employee’s face photos are dynamically captured through the camera connected to the computer, and the captured face photos are compared with the pre-collected photos. If the employee is the employee, the attendance is successful, and the attendance is judged based on the attendance time. Circumstances (normal, late, leave early or overtime, etc.), and record attendance information, such as attendance date, shift, working time and other information. If it is not the employee's attendance, the system will give a friendly prompt and take corresponding processing. After all attendance data is processed and summarized by the host program, the corresponding management report or data text is generated, which can be provided to the personnel department for personnel evaluation or the financial department for salary settlement.
2. 1 Composition of face attendance systemThe face recognition attendance system includes two parts: hardware and software. The hardware consists of an attendance machine, a statistical host and a printer. The attendance machine is composed of a host, a camera, and a video capture card. It is used for facial attendance recognition and image building; statistical host Used for attendance statistics, query, management, save system data, etc. The printer is used to print attendance reports. The software consists of Windows2000 Advanced Server operating system, SQL Serv2er2000 database management system and face attendance system.
2. 2 The functional structure of the face attendance systemThe face recognition attendance system consists of the face recognition subsystem and the attendance management subsystem. Its functional structure is shown in Figure 1.
The face recognition subsystem includes a video acquisition module, a face detection module, a face processing module and a face recognition module. The video acquisition module collects the user's face video image in real time from the attendance site; the face detection module periodically samples a frame of the video in the video, and uses the feature information of the face to accurately locate the face coordinates and the face range in the image; The face processing module intercepts the image of the face area, normalizes the size, equalizes the brightness, and enhances the image, and then separates the background to obtain an image containing only the face; the face recognition module uses a variety of recognition technologies to obtain the classification results and make decisions Level fusion gets the final result. The attendance management subsystem includes information registration, information query and printing report modules. The administrator enters the employee's information in the information registration interface and collects multiple photo samples of the employees on the spot; in the information management interface, you can query employee information and attendance information; and can print out the employee's attendance report for a certain period of time at any time.
3 Realization of face attendance system3. 1 Video capture technology
Use VFW (Video for Windows) technology to collect facial video images. The VFM library function is a software toolkit provided by Microsoft for the development of video applications under the Windows platform. It provides a series of application programming interfaces (API) through which users can easily realize video capture, video editing, and video playback, etc. Functions. VFM's video capture functions mainly include capturing video streams, capturing video streams to buffer, capturing video streams to AVI files, local browsing, and capturing single-frame previews. VFW is mainly composed of 6 modules, of which the AVICAP module mainly realizes the function of processing video capture, and provides a programming interface for the input and output of AVI files and the video and audio device drivers; the MSVIDEO module mainly realizes the connection of the video capture window and the drive device, and supports ICM Video encoding service. AVICAP realizes the general process of real-time video capture:
(1) Create a video capture window;
(2) Register system callback function;
(3) Get the default settings of the video capture window;
(4) Set the relevant parameters of the capture window;
(5) Connect with video capture equipment;
(6) Set the display mode of the capture window;
(7) Capture video to cache or file and process it accordingly;
(8) Terminate video capture and disconnect from the video capture device.
3. 2 face detection technologyFace detection refers to the process of judging whether there is a face in the input image, and if so, the process of determining the position, size, and posture of all faces. In order to improve the speed of face detection, the Ad2aBoost algorithm is used in this paper to perform face detection. The AdaBoost algorithm was first proposed by Freund and Schapire in 1996. At present, Ad2aBoost is widely used in face detection, data mining, machine learning and other fields.
The AdaBoost algorithm is an iterative method that selects the most meaningful combination from a large number of weak classifiers to form a strong classifier. The purpose is to improve the accuracy of detection. Viola et al. proposed the concept of integral image and the method of training face detection classifier based on the AdaBoost method as shown in Figure 2. This method uses the rectangular feature of the image as the basis for classification. In order to effectively exclude a large number of non-face windows, the entire classification The classifier is formed by cascading multiple sub-classifiers. Each classifier corresponds to a Harr feature. The input window to be inspected will be verified level by level. If the classifier in it verifies it as a non-face, it will be immediately excluded and no longer enter the next level of classifier verification. The AdaBoost method proposed by Viola et al. not only has a good detection effect, but also has a detection speed of 15fps, which is easy to implement , Has been widely used in actual system development. Due to space limitations, see the literature for specific algorithms. The final strong classifier CFINAL (x) is
In the formula, Ci (x) is a weak classifier, ai = log (1/βi), the specific meaning is that when the extracted classifier classifies some samples correctly, then reduce the weight of these samples (βi becomes smaller); When the classification is wrong, increase the weight of the sample (βi becomes larger). In this way, the simple classifier extracted after training will further strengthen the training of these misclassified samples.
The system collects 10 different frontal face images of each person and saves them to the training library. Figure 3 is an example of dynamically collecting 10 photos of a certain person. In the process of collecting 10 samples of the same face, if the face collected later is very similar to the face collected earlier, it is considered that the face collected this time already exists in the training library. The result is not saved and the collection continues. Until the images with higher similarity are collected, they are then saved in the training library. The calculation of similarity is shown in equation (2), where fi and fj represent the i-th and j-th faces, respectively, where fi=(ui1, ui2, ?, uip). Finally, in the training library, different samples of the same person are classified into the same category, and samples of different attendees belong to different categories.
The face image is high-dimensional data, and the distribution in the high-dimensional space is not compact, and the calculation complexity is high. Therefore, it is necessary to reduce the dimensionality of the high-dimensional face data. Common dimensionality reduction methods include principal component analysis (PCA), independent component analysis (ICA), linear discriminant analysis (LDA/FDA), non-negative matrix factor (NMF), multidimensional scaling analysis (MDS), and nuclear principal component analysis (KPCA), Nuclear Fisher Decision Analysis (KFDA), Manifold Learning (ML) and other methods. Considering that the face image of this system is basically a frontal face during attendance, the PCA method is used to reduce the dimensionality and feature extraction of the face image and finally recognize the face. Kirby and Sirovich [6] first introduced the PCA idea to face recognition; Turk and Pentland [7] further developed the Eigenface method for frontal face recognition and achieved great success. PCA is a face recognition method based on DKLT (Discrete K-L Transform). DKLT transform is an optimal orthogonal transform for image compression, and its generating matrix is ​​generally the overall dispersion matrix of training samples. After the high-dimensional image space is transformed by DKLT, a new set of orthogonal bases is obtained, and the important orthogonal bases are retained. These bases can form a low-dimensional linear space. PCA is a simple, fast, and practical multivariate analysis technique based on algebraic features. It retains the topological relationship between face components and also retains the information of various organ components. It is currently widely used.
The purpose of PCA is to find a set of optimal unit orthogonal vector bases (ie principal components) through linear transformation, use their linear combination to reconstruct the original sample, and make the mean square error (MSE) between the reconstructed sample and the original sample The smallest. Suppose there are N samples xk (k = 1, 2, ?,
N), the mean of all samples is E(x), then the covariance matrix of x
Let u1, u2,? , Un and λ1, λ2,? , Λn are the eigenvectors and eigenvalues ​​of S respectively, and λ1 ≥λ2 ≥? λn. Then decompose the covariance matrix S into:
In the formula, U = [u1, u2,? , Un], Ω = diag (λ1, λ2, ?, λn) Choose the optimal m (m << n) non-zero eigenvalues ​​as the principal components, and minimize the mean square error (MSE). Let Wopt = [u1, u2,? , Um ], then the sample of the original space can be used to describe the projection coefficients on the low-dimensional principal component subspace:
After the feature extraction is performed on the face image, the similarity measurement is then performed. Common similarity measures include distance-based methods, probability-based methods, and entropy function-based methods. Among the distance-based methods, there are Euclidean distance, Manhattan distance, Hamming distance, Chebychev, s-order Minkowski and other methods. In recognition methods based on geometric features, the distance function between feature vectors is usually used for measurement. Aiming at the feature vector information obtained by this system, the Euclidean distance method is used to calculate the distance between the face to be recognized and the training sample. According to the result of similarity calculation, if the similarity is greater than the given threshold δ, the system determines that the face collected on site and the face image collected in advance are the same person, and records the date of attendance, working hours and other information in the database. The basis for determining the face similarity threshold δ during attendance comes from the sample similarity value during face collection, to overcome the arbitrariness of the threshold set by human factors and ensure the rationality of the threshold value.
4 Experimental results and analysisThe dynamic face attendance system collected 10 images of each of 50 PhD students from the Tiansi Intelligent System Research Institute for recognition experiments. The results show that the one-time recognition rate of the system reached 96%. The main reasons are: (1) Face training library Relatively small and fixed personnel; (2) When attendance is required to face the front or close to the camera; (3) Normally, the posture and expression of the user during attendance are normal, and there is not much change, and there is no complicated background and Exaggerated makeup, etc., so the recognition rate is high.
5 concluding remarksIn this paper, a dynamic face recognition attendance system based on Adaboost and principal component analysis (PCA) is designed and implemented. The AdaBoost algorithm selects multiple weak classifiers to form a strong classifier to improve the speed of face detection; the PCA algorithm implements dimensionality reduction and compression from high-dimensional feature space to low-dimensional feature space, and extracts high-dimensional face images in low-dimensional space Characteristic description, realize face recognition, and finally realize attendance management. System operation results show that the dynamic face recognition attendance system has the characteristics of face recognition dynamics, high recognition rate, good practicability, and strong reliability. The next step is to further optimize the selection and extraction algorithms of facial features to improve the speed of recognition and the performance of the system.
SHENZHEN CHONDEKUAI TECHNOLOGY CO.LTD , https://www.szfourinone.com