|Project Description: ||Two subsystems that could be utilized in the creation of a face recognition system were investigated, a face detection subsystem, and a normalization subsystem. The detection of a face in a digital image is not a simple process and numerous methods have been proposed to accomplish this. Among these methods the face detection via color segmentation method is investigated. This method involves detecting pixels of ‘skin’ color in the image, then grouping and filtering to determine which sets are likely to contain a face. This method was found to be very susceptible to variations in lighting and background colors. Grouping and filtering to determine like face candidates was also very error prone. Overall, face detection via color segmentation was found to be insufficient to accurately detect faces in images.
The normalization process is the process of removing variations in the facial image and preparing the image for the recognition process. A method based on neural networks performing a nonlinear PCA (Principle Component Analysis) of the face images was investigated. No neural network was found that was able to perform this transformation successfully. Although on some training image sets, some network configurations were trained to produce rudimentary results similar to those expected from the desired nonlinear PCA transformation. Recommendations are made as to how to continue research in this area.