| In this thesis we propose an automatic face recognition system which can recognize faces appearing in images acquired in relatively uncontrolled environments. The system can be visualized as a computer controlled TV channel selector, which can restrict the channel selection depending upon the viewers watching the TV. A camera placed on the TV set grabs images of the audience, which are processed by an algorithm controlling the channels. We use gray scale images of 640 x 480 resolution for audience identification. The first stage searches for faces at all possible locations with all possible sizes in the test image and passes on the size and face location information to the recognition stage where each face is matched with a database of known faces created at the time of training. On recognition of a subject, the algorithm may initiate actions like restricting access of some TV channels, etc. We used 300 images for training and evaluation of the system with two to four subjects appearing in each image. System was trained to recognize the images of 10 subjects and reject all others. The results show the effectiveness of our method to recognize faces with large variations in scale, orientation, illumination, expressions, and background. Still, better methods for face location and recognition need to be developed to improve the recognition accuracy. Depending upon background and number of faces appearing in the image, processing of each image takes between 45 to 90 seconds on a Sun Ultra-1 workstation due to the computational complexity of our algorithm. |
