Face recognition of remote monitoring under the Ipv6 protocol technology of Internet of Things architecture

3.1 IPv6 protocol technology of IoT architecture

The IoT uses equipment such as positioning systems, sensors, and radio frequency identification technology to connect people and things to the network to exchange communication and information, so as to realize identification, positioning, monitoring, and intelligent management. The essence of the IoT is an extension of the Internet. On the basis of the Internet, it has special functions such as remote control, intelligent identification, and real-time monitoring [16]. Under the current network structure standards, the IoT is divided into application layer, transport layer, and perception layer. The face information is collected by the IP camera of the perception layer, then transmitted through the transport layer under the network protocol, and finally processed by the application layer to realize remote monitoring of face recognition. The system architecture is shown in Figure 1.

Figure 1

System architecture diagram based on IoT.

As the key protocol in transmission control protocol/internet protocol (TCP/IP), IP protocol is suitable for network communication of wide area network and local area network. The IPv4 protocol is widely used at present. With the increase in Internet machine access, the 32-bit network space it possesses has made it difficult to meet realistic requirements. The IPv6 protocol is a new generation of Internet protocol. Compared with IPv4, there are three main advantages [17]. First, the address space of IPv6 is 28 bits, which can support more levels of address space, i.e., the address space is wider. Second, there are only six deterministic domains for IPv6, with a total size of 64 bits and no additional options [18]. Finally, IPv6 has added automatic configuration support, which has not only a more intelligent and higher configuration space but also stronger confidentiality and higher security capabilities. The powerful advantages of IPv6 make it necessary to convert IPv4 to IPv6. There are translation mechanism, tunnel mechanism, and dual stack mechanism to realize this process [19]. The tunnel mechanism can avoid the dependencies when upgrading areas, routers, or nodes. The dual stack conversion mechanism not only enables free communication between IPv4 and IPv6, but also enables mapping between IPv4 and IPv6. In the combined IPv4/IPv6 network, communication between these two nodes and between IPv6 nodes is a difficult point to overcome. Therefore, the research adopts the cross-platform, object-oriented Java language to develop [20]. IPv6 addresses are divided into anycast, multicast, and unicast. Arbitrary wave address is routing policy and network addressing, calculates the distance through routing protocol, and transfers data packets to the address interface with the shortest distance. Multicast network addresses deliver packets to all address interfaces by establishing a point-to-multipoint connection, i.e., between a server and a browser [21]. A unicast network address establishes a peer-to-peer connection and only sends packets to a single address interface. The flow of server–client socket communication is shown in Figure 2.

Figure 2

Flow chart of server–client socket communication.

In the dual-stack mechanism system, the Java network stack establishes IPv6 sockets first. In a mixed programming system, Java needs to establish two sockets, IPv6 and IPv4. On the server side, the two sockets established will remain permanent [22]. On the client side, it is resolved by the IP address, and after an IP protocol is adopted, another socket is closed [23]. So, the most important thing in Java socket programming is to obtain and verify the IP address. First, the server creates a ServerSocket instance, then calls Accept and waits for the client to establish a connection. After the connection is established, the client’s Internet protocol, namely IPv6 or IPv4, is determined to process the corresponding data information. In the perception layer, the IP network camera controls the camera through the gateway interface technology, and the general gateway interface can execute the server-side program after compilation [24]. The Pan, Tilt, and Zoom (PTZ) camera used in the study can be controlled with multiple zooms and moved up, down, left, and right. In the face recognition system, the hypertext transfer protocol is obtained through the web camera and then handed over to the Common Gateway Interface program for processing, thereby realizing communication interaction. In remote video surveillance, compression coding is an extremely important part. After the IP network camera collects video data, it will be compressed into M-JPEG format and then transmitted to the face recognition system.

3.2 Face recognition system based on PCA

After the video data are collected by the IP network camera and compressed into a picture, a face recognition system is required for identification. Before face recognition, geometric normalization of face image is carried out. After the video data are collected by the IP webcam and compressed into pictures, the face recognition system is required for recognition. The research uses the whole method to extract the features of the face image because the facial expression is generated by causing facial muscle movement, resulting in facial texture and body changes. From an overall point of view, this kind of motion makes the face deform significantly and has a great impact on the global information of the static image. Therefore, the facial expression feature can be extracted from the overall point of view. Based on the overall feature extraction algorithm, the facial features are expressed by extracting the macroscopic information of the face. PCA is an extraction algorithm based on linear features, which can effectively extract the main features of high-dimensional data samples, and is one of the main features and dimensionality reduction methods. PCA algorithm, as a fast and efficient linear algorithm, has been successfully applied to extracting facial features and data dimensionality reduction. However, when the PCA algorithm is used for face recognition, it mainly faces three deficiencies [25]. First, the PCA algorithm does not have a high accuracy rate for face recognition. Second, due to the nonlinear characteristics of human faces, it is difficult for PCA algorithms to process them. Furthermore, it has no classification ability [26]. Therefore, the existing PCA algorithm needs to be optimized. The flow of PCA algorithm is shown in Figure 3.

Figure 3

Flow chart of PCA algorithm.

The PCA algorithm uses a small number of main features to represent the original main part of the data set information by projecting the data set from the high-dimensional space into the low-dimensional feature space [27]. The projection expression of a certain vector in a certain direction is shown in equation (1).

where x _j represents the vector, v is the direction, and θ represents the angle between the two vectors. When the data are standardized, the variance derivation after projection is shown in equation (2).

where σ 2 represents the variance and v T x i is the numerical value after projection. Since v T x i = 0 represents a numerical value, we obtain v T x i = ( v T x i ) T as shown in equation (3), and hence

where v T x i represents the m × m matrix and m is the number of eigenvalues. Then, to obtain the maximum variance, the corresponding vector must be found. Therefore, equation (3) is converted into the solution eigenvalue and deduced by the Lagrange multiplier method, and the final eigenvalue equation, namely the variance equation, is obtained as follows (equation (4)):

where λ represents the indefinite multiplier. Therefore, the PCA algorithm first needs to calculate the projection vector of the vector in a certain direction and then calculate the inner product of the vector and the direction. Then, the variance after projection is obtained, and the eigenvectors and eigenvalues are obtained. Finally, the eigenvalues are arranged, i.e., the largest eigenvalues are selected, and the eigenvectors are obtained, i.e., the final projection. After analyzing the PCA, a linear discriminant analysis method is introduced, which can perform learning supervision and effectively handle the sample category information [28]. It first projects the sample features in the high-dimensional space to the low-dimensional space and thus obtains a plane in the low-dimensional space to collect the data samples of the same features projected in the plane. At the same time, the difference feature samples are separated to achieve the purpose of “large variance between classes and small variance within classes,” as shown in Figure 4.

Figure 4

Schematic diagram of LDA algorithm projection.

In Figure 4, five-pointed asterisks with different colors represent different data categories, m 1 and m 2 represent the average value, u 1 and u 2 represent the data features, and W represents the plane. The linear discriminant algorithm (LDA) algorithm uses a straight line to separate the u 1 and u 2 and takes the best separability of the sample as the purpose to reduce the variance of the sample. The “inter-class” walk matrix of training samples is shown in equation (5):

where N i represents the sum of training samples, m i represents the mean vector, and S b is the “inter-class” scatter matrix. The “intra-class” scatter matrix is shown in equation (6):

where S w represents the “intra-class” scatter matrix and m i ( K ) represents the Kth face. The optimal projection matrix is shown in equation (7):

where J represents the optimal projection matrix. Therefore, the LDA algorithm can process the category information of face samples to the greatest extent. Therefore, the PCA and LDA algorithms are combined and applied to the extraction of facial features. First, PCA algorithm is used to extract the global features of the face image, and then LDA algorithm is used to identify the face categories to further compress the face data, so as to achieve face feature enhancement and face feature extraction. After the combination, we first need to obtain the sample mean value, select samples from the face database, and divide them into test sets and training sets. The overall face mean value is shown in formula (8):

where m i ( K ) represents the average of the overall face, m is the number of faces, and c is the number of pictures. After obtaining the mean difference value of the training samples, the LDA algorithm can effectively utilize the information of “intra-class” and “inter-class,” and its distance judgment equation is shown in equation (9):

where S t is the overall scatter matrix. From this, the mean difference value is obtained, and the one-dimensional sample is redefined. The final equation of all scattering matrices is shown in equation (10):

where M represents the mean difference. Then, select a suitable kernel function, and input the judgment criteria of “inter-class” and “in-class” distance into the kernel function to obtain the kernel matrix K . Then, the radial basis kernel function is selected as the input of the kernel function after the combination of the two algorithms, as shown in equation (11):

where σ 2 represents the variance. Finally, center the kernel matrix to obtain eigenvalues and eigenvectors, as shown in equation (12):

where η is the eigenvalue, α is the eigenvector, and N is the N-order matrix. Finally, a new projection matrix is obtained, and the new projection matrix is obtained by using the kernel principal component, and its expression is shown in equation (13):

where Y represents the projection matrix. Therefore, the face recognition process after the fusion of PCA and LDA algorithms is as follows: first, the system reads the image data from the face database and preprocesses it. Then, a fusion algorithm is used to extract the processed facial image features, and finally, a classifier is used to train and recognize the extracted features [9]. To further improve the accuracy of face recognition, an SVM classifier is used to identify the face after feature extraction. SVM classification is powerful and performs well in handwritten digit recognition and face recognition [29]. The SVM will segment the training sample set to obtain planes of different classes. One of the multiple planes can maximize the classification distance between the sample points, i.e., the optimal hyperplane. At this time, the closest distance to this plane is the support vector. Then, the training samples are divided by the optimal hyperplane. The farther the support vector is from this plane, the better the classification effect, as shown in Figure 5.

Figure 5

Classification diagram of SVM algorithm.

In Figure 5, different graphs represent different training samples, H₁ and H₂ represent the parallel planes closest to the optimal hyperplane, and the distance between them is the classification interval, denoted by Margin. The determination equations of H₁ and H₂ are shown in equation (14):

where x i represents the sample set and y i represents the label. Therefore, after correctly classifying all sample points into their respective categories, the satisfied expression is shown in equation (15):

where i represents the number of samples. Therefore, the algorithm used in the final face recognition system is PCA + LDA + value stream mapping, which is named as the PLV algorithm and is applied to the application layer of the remote monitoring face recognition system under the Ipv6 protocol technology of the IoT architecture.