CN1629875A - Distributed Face Detection and Recognition Method in Mobile Computing Environment - Google Patents

Abstract

The invention relates to a distributed face recognition method in a mobile computing environment in the technical field of pattern recognition, comprising the steps of: obtaining an image from a camera on a mobile device, performing simple and effective light correction on the image; Detection and calibration, embedding the calibrated face area with a watermark; transmitting the watermarked face area image to the face recognition server through a wireless network; finally, verifying the authenticity of the area image through the watermark on the server, and then Perform face recognition and transmit the recognition results to the mobile terminal.

Description

Distributed Face Detection and Recognition Method in Mobile Computing Environment

technical field

The invention relates to the technical field of pattern recognition, in particular to a face detection and recognition method in a distributed mobile computing environment combining face detection and recognition technology, light processing technology and wireless communication technology.

Background technique

In recent years, wireless network technology has shown an explosive development momentum. With the development of network technology, wireless network has more and more influence on people's daily life. The expansion of wireless network bandwidth and the improvement of network reliability have made people's demand for intelligent multimedia applications in wireless networks increasingly strong. These applications include:

·Face recognition

·Expression Analysis

· Content-based video playback

·Wireless interactive games

· Wanted criminals and criminal suspects identification, etc.

Among them, face detection and recognition technology is an essential supporting technology for these applications.

The so-called face detection is to detect the position, size, number and direction of faces in multimedia data (such as digital signals such as images and videos); face recognition is to recognize the results obtained by face detection to judge Which person does the detected face belong to.

The existing face detection and recognition technologies are all based on the operating environment of desktop computers or even workstations, which cannot meet the needs of people in handheld mobile computing environments. Compared with desktop systems, handheld mobile computing environments have some special operating conditions, including:

·Weak computing power

·Limited power and data storage space

·The light in the environment changes greatly

· Different users have different usage habits, etc.

Due to the changing lighting conditions in the mobile computing environment, in order to improve the robustness of face detection and recognition, light processing is an essential link. Light processing technology can process pictures captured under various lighting conditions, minimizing the impact of lighting changes on face detection and recognition.

Since light processing and face detection and recognition are tasks with a large amount of calculation, it is necessary to optimize the algorithm of light processing and face detection and recognition to reduce the amount of calculation, and put the face recognition part on the server to reduce the workload. The computing and storage burden of handheld devices, which is very critical for small mobile devices.

Contents of the invention

The purpose of the present invention is to provide a method of face detection and recognition based on a distributed system in a wireless network environment. This system performs real-time and robust face detection on the handheld device, and then encrypts the detected face image with digital watermark and sends it to the server through the wireless network, and performs face recognition on the server and returns the recognition result.

To achieve the above purpose, the face detection and recognition steps in this distributed mobile environment include:

1. Detection and calibration: perform light preprocessing on images captured by handheld devices (handheld computers, mobile phones or PDAs) and perform face detection and face range calibration;

2. Encrypted transmission: Encrypt the calibrated face range with digital watermark and send it to the server through wireless network. The server verifies the digital watermark embedded in the image to judge the integrity and correctness of the image;

3. Recognize and return the result: use the face recognition training algorithm based on the embedded hidden Markov model Hidden Markov Models (HMM) to perform face recognition and return the result to the handheld device.

Obtain video data from the camera, use the mean value and variance of the pixels to correct the light of the video data; detect the size and position of the face from the image; embed the watermark into the detected face area; and then pass the wireless network. The area image is transmitted to the server, and the server verifies the authenticity/fake of the image according to the watermark; finally, it recognizes and transmits the recognition result to the mobile terminal.

Description of drawings

Fig. 1 is the overall frame diagram of distributed face detection and recognition method under the mobile computing environment of the present invention;

Fig. 2 is the flowchart of the distributed face detection and recognition method under the mobile computing environment of the present invention;

Fig. 3 is a training flowchart of the distributed face detection and recognition method in the mobile computing environment of the present invention.

Detailed ways

The process in Figure 1 is as follows:

·Face detection process

(1) Acquiring a frame image: capture the image through the camera C on the handheld device. Because we have to process each frame, we need to extract the image frame by frame from the video stream.

(2) Light processing: Since light has a huge impact on face detection and recognition, this includes the effects of different angles, different intensities, and shadows. At the same time, due to the weak computing power of mobile devices, complex light models cannot be adopted. Here, the mean and variance of pixel values are used to correct the influence of light on each pixel of the image, that is, the pixel value of each pixel is subtracted from the mean, divided by the variance, and then multiplied by a coefficient. After the above processing, the influence of light on the pixel value of the face is eliminated as much as possible.

(3) Perform face detection: perform face detection on the image after light processing.

(4) Calibrate the detection area: After the image passes through the detection module, we will get the coordinates and size of the face part. The face area is demarcated by these coordinates.

·Encrypted transmission process

(1) Embed digital watermark: Since the security and reliability of the wireless transmission path are worse than that of the wired network, we must ensure the integrity and correctness of the transmitted data. This will be achieved by embedding a digital watermark in the picture. At the same time, in order to improve the transmission speed, the object we process is only the face area in the image. This area was extracted by face detection in the previous process. The detected face area is quickly embedded with an integer wavelet watermark, so that the image authenticity/false verification of the area can be performed on the server side.

(2) Wireless transmission: The face area image embedded with watermark is transmitted to the server through wireless network. The wireless path here can be a wireless local area network or a GPRS network.

(3) Distributed calculation process: Since there are multiple recognition servers in the background for data processing, it is very important to coordinate the communication between the handheld terminal and the server. We add a proxy server between the handheld terminal and the identification server. The proxy server is a computer running a special scheduling software. The specific steps of the scheduling software are illustrated in Figure 2. The proxy server is like a commander-in-chief, redirecting the requests sent by many handheld terminals to different servers, which can reduce the time for data to be processed. The face databases on all servers are consistent, and all databases are updated synchronously during the face registration phase. As shown in Figure 2, the processing process is as follows:

a) The proxy server searches for an idle face recognition server after receiving the processing request. The face recognition server is a computer running a face recognition program. The specific steps of the program are illustrated in FIG. 2 . If no server is idle, wait in line, otherwise, send the received processing request to an idle server, and mark the server as busy;

b) The face recognition server verifies the digital watermark embedded in the image, and if it finds that the data has been changed or destroyed, it notifies the handheld terminal to resend the data, and at the same time notifies the proxy server that it is in an idle state;

c) After processing the undamaged data, the face recognition server returns the recognition result to the handheld terminal and sends a message to the proxy server, telling the proxy server that it is in an idle state.

·Face recognition

1) Face recognition training algorithm based on embedded hidden Markov model. Before face recognition, it is necessary to register and train the registered face images for subsequent recognition. What adopt in the present invention is to analyze the proportional relationship of each part inside the human face to recognize the human face. Therefore, a hidden Markov model is used for training in the training phase. However, the face pixel data is very large. If the hidden Markov model is directly used for training, the model will be too large to be recognized. In the present invention, an embedded hidden Markov model is adopted to divide the human face into five states of forehead, eyes, nose, mouth, and chin, and each state is a hidden Markov chain. During training, the hidden Markov parameters between the five states are trained, and the hidden Markov model inside each state is trained. This way each face gets an embedded Hidden Markov Model.

2) Face recognition based on embedded hidden Markov model. According to the images obtained by face detection, use the Viterbi algorithm to obtain various probabilities of capturing images under various embedded hidden Markov model conditions. The model with the highest probability is the most likely model of the captured face. And each model corresponds to a person. In this way, the person corresponding to the face image can be identified.

In Fig. 2, the distributed face detection and recognition method under the mobile computing environment, its processing steps are as follows:

S2-1 The proxy server looks for an idle face recognition server after receiving the processing request,

S2-2 Go to S2-3 if no server is idle, otherwise go to S2-4,

S2-3 waits in line, and returns to S2-2 to continue looking for an idle face recognition server,

S2-4 Send the received processing request to the idle server, and mark the server as busy at the same time,

S2-5 The face recognition server verifies the digital watermark embedded in the image, if it finds that the data has been changed or destroyed, go to S2-6, otherwise go to S2-7,

S2-6 Notify the handheld terminal to resend the data, then go to S2-9,

S2-7 for face recognition,

S2-8 Send the recognition result to the mobile terminal,

S2-9 Send a message to the proxy server to inform the proxy server that it is idle.

In Fig. 3, the distributed face detection and recognition method under the mobile computing environment, its training steps are as follows:

S3-1 Establish HMM model,

S3-2 Evenly divide the face image to obtain the initialized face HMM model parameters,

S3-3 Initialize HMM parameters such as observation sequence parameters, observation vectors and state transition matrices,

S3-4 uses the embedded Viterbi segmentation method for parameter adjustment,

S3-5 Estimate the model parameters by calculating the average K in segments,

S3-6 Determine whether the iteration of the model is converged, if not, go to S3-4 to continue parameter adjustment. If converged then go to S3-7,

S3-7 After iterative convergence, the trained HMM model is obtained.

Using the detected face area to transmit in a wireless network environment solves two problems:

1. The large amount of calculation for multi-category problem classification on mobile devices;

Second, it is difficult to transmit large images to the server in real time in a wireless network environment.

Claims (7)

1, distributed people's face detects and recognition methods under a kind of mobile computing environment, comprises step:

Detect to demarcate: the image that handheld device is caught carries out the light pre-service and goes forward side by side that pedestrian's face detects and the demarcation of people's face scope;

Encrypted transmission: the people's face scope that calibrates is carried out sending to service end by wireless network after digital watermarking is encrypted, and service end is verified the digital watermarking that embeds in the image, judges the integrality and the correctness of image;

Identification and return results: adopt recognition of face training algorithm to carry out recognition of face and the result is returned to handheld device based on built-in type hidden Markov model.

2, distributed people's face detects and recognition methods under the mobile computing environment according to claim 1,

Step is as follows: obtain video data from camera, utilize the average of pixel and variance that video data is carried out the light rectification; From image, detect size, the position of people's face; Watermark is embedded in the detected human face region; By wireless network the human face region image is transferred to server then, server carries out very/pseudo-checking this image according to watermark; Discern at last, and recognition result is transferred to the mobile terminal.

3, method according to claim 2 is characterized in that, also comprises step: adopt the average of pixel and variance to carry out the light rectification, this mode is simple and effective, can satisfy the requirement of mobile computing.

4, method according to claim 2 is characterized in that also comprising step: detected human face region is embedded the integer wavelet watermark fast so that server end carry out this area image true/pseudo-checking.

5, method according to claim 2 is characterized in that also comprising step: adopt built-in type hidden Markov model in the recognition of face training, the proportionate relationship at each position of analyst's face improves discrimination and recognition speed greatly.

6, distributed people's face detects and recognition methods under the mobile computing environment according to claim 1, and its treatment step is as follows:

The S2-1 acting server receives the request back of processing and seeks the recognition of face server that is in idle condition,

If S2-2 does not have the server free time then forwards S2-3 to, otherwise forwards S2-4 to,

The S2-3 wait of ranking, and return S2-2 and continue to seek the recognition of face server that is in idle condition,

The processing request that S2-4 will receive sends to the server that is in idle condition, is busy state with this server-tag simultaneously,

S2-5 recognition of face server is verified the digital watermarking that embeds in the image, if find that data are changed or destroyed, just forwards S2-6 to, otherwise forwards S2-7 to,

S2-6 notice handheld terminal resends data, forwards S2-9 then to,

S2-7 carries out recognition of face,

S2-8 sends to mobile phone terminal with recognition result,

S2-9 sends message to acting server, and the notification agent server lays oneself open to idle condition.

7, distributed people's face detects and recognition methods under the mobile computing environment according to claim 6, and its training step is as follows:

S3-1 sets up the HMM model,

S3-2 is evenly cut apart facial image, obtains initialized people's face HMM model parameter,

S3-3 initialization observation sequence parameter, observation vector sum state-transition matrix HMM parameter,

S3-4 adopts Embedded Viterbi dividing method to carry out parameter adjustment,

S3-5 comes the estimation model parameter by segmentation calculating mean value K,

Whether the iteration of S3-6 judgment models restrains, and proceeds parameter adjustment if do not restrain then forward S3-4 to.If the convergence would forward S3-7 to,

S3-7 has just obtained the HMM model that trains through after the iteration convergence.

Images