Upload
duongkiet
View
220
Download
3
Embed Size (px)
Citation preview
IntroductionAlgorithms of face recognition have long been researched and are maturely developed. However,the issue becomes intractable when these algorithms are implemented on an embedded platformwhich has comparably low computation power.In our project, a face recognition algorithm on a portable device (an embedded platform) isimplemented. A input image will first be captured by the digital camera and is processed by theFace recognizer and the Search engine. Information will be displayed to the user through the Videoeyeglasses interface.
The face recognition algorithm in this project is mainly divided into three parts, Face detection,Face extraction and Face recognition.
The algorithm is implemented on two embedded platforms and is compared with aimplementation in a paper called "Characterization of visual feature recognition“.
Aims and ObjectivesThe project objective is to implement face recognition in an optimum way in terms of run timeonto the embedded system. Various algorithms and methodologies are studied and hardwareresources planning will be done to achieve the goal.This kind of face recognition embedded system can be widely used in our daily life in differentsectors. We hope that human life can be greatly helped with this technology. Some typicalapplications are listed as follows.
Methodology
Methodology (cont’d)Faces are stored in the database for the face-recognition function. At most five faces of differentviewing angles can be stored for each person/record. This leads to the best recognitionperformance if the five faces are captured in the following orientation: -45o, -20o, 0o, +20o, +45o.
Face recognition algorithms flow chartThe Face recognition algorithms are mainly divided into three parts.
User InterfaceFace recognition mode (left): Face recognition function is performedFace record mode (right): New record for a person can be added to the database
ResultsRun time comparison 1:Compare the face recognition algorithm on two difference embedded platforms.
Run time comparison 2:Compare between the embedded configuration in the paper "Characterization of visual featurerecognition“ and the face recognition algorithm in this project on Devkit8000 platform.
Successful rate test:A test on the successful rate is done with 104 test faces, which are different form the 85 facesstored in the database. Threshold is set to identify the unrecorded face. With the threshold is set ataround 2400, a successful rate over 70% can be achieved.
ConclusionThe face recognition algorithms are successfully implemented on the embedded platform. Inaddition, a face recognition software application with a friendly user interface is designed and thenew display interface---the video eyeglasses is used to display. This provides user with a totallynew experience.
To summarize, four major goals of the project have been achieved. They are•An effective face detection and recognition algorithm that can be run on embedded platform withacceptable run time•A face recognition software application with a friendly user interface•An embedded system running the face recognition software application•The video eyeglasses display interface
Instructions
Captured face preview
Sample figure
Camera preview
Information bar
Saved face
Database
Correlation percentage
Instruction
Camera preview
Mode selection
Inputimage/video
Face detection
Face extraction
Face recognition
Identification data
A digital camera is used to capture images, new detectedfaces can be recorded. Information like name,organization, character and hobby with respect to eachindividual can be input and stored in the device. Personalinformation/remarks will be prompted out whenever thesystem recognizes any recorded faces under the display.
Original Image
Hue Transform
Decision Maker
Initial Face Region Detection
Decision MakerMasking
Extract Corresponding region from the original image
Extraction stop
Eigenface DecompositionEigenvectors
Euclidian Distance MeasurementDatabase
Decision Maker
Recorded Face Unrecorded Face
Face Non-Face
Successful
Rate (%)
Threshold
Operations YC2440 DevKit8000
Database setup (10 images stored in the database) 10.7388s 0.476s
Face detection(per frame) 9.853s 0.505s
Face recognition(per frame) 0.267s 0.011s
Embedded platform
Embedded Platform Embedded Platform Improvement Remarks
Operations
Embedded configuration in
"Characterization of visual
feature recognition"
Devkit8000
Flesh tone + Viola detection Flesh tone
2.16s 0.505s 76.6%
Face recognition( per frame) 0.992s 0.487s 50.9% Include database setup time
Face recognition( per frame) NA 0.011s NA Exclude database setup time
Face detection(per frame)
Sector Applications
Business Meeting, Gathering
Education Teaching assistant
Psychology Eye disorder helper
-45o -20o 0o +20o +45o
Sample Input:
Database: