Mohamed Ismail, Professor, Dept of Computer Science, Alexandria
A near-real-time computer system was developed to locate, track a subject's head, and then recognize the person by comparing characteristics of the face to those of known individuals. The system functioned by projecting face images onto a feature space that spans the significant variations among known face images.
In this work, we tackled all main aspects of the problem of face recognition. Here are the main approaches we evaluated throughout that project:
Edge detection based method
In this approach we investigated the use of edges as a tool for the location of the face.
Location from the motion of the face
We used image flow analysis to detect the motion of the moving object, which we assume to be the face. The moving region is then segmented from the image.
Color based methods
Human faces share nearly the same color properties. We used this information as a tool for locating the face, and/or facial features.
Feature-based methods. (Neural Networks)
We trained a neural network to classify the facial features such as eye, nose and mouth. This network is then used to locate the features, thus locating the face as well.
Space gray‑level dependence (SGLD) matrix.
Space gray‑level dependence matrix encodes certain features of the image. These features can be utilized to characterize the face.
An approach that uses the edge properties of the facial features is used to locate and measure the features according to a certain knowledge base.
The neural network mentioned above can be used also to extract the features.
The use of deformable templates is attempted for precise feature measures. An eye corner detection technique is used to guide the template as a good intialization step.
Features that are utilized to characterize the face do not necessarily correspond to the intuitive features of the eyes, eyebrows, nose, mouth or measurements of the head size. The eigen-representation provides an interesting alternative that proves to be very useful.
The neural network can be used instead of Euclidean distance, as similarity measure. The network learns the features of the ‘known’ faces. The use of fuzzy sets to absorb the variation of feature values has been attempted.
The eigenfaces approach provides an almost complete solution to the problem of face recognition. By projecting images into the selected face space, similarities between faces can be detected.
There were two major advantages of our work. First, no human intervention was required, as opposed to manual feature extraction. Second, less constraints on input images are imposed; the effect of the background has been reduced, the glasses do not impose any problems, facial expressions can be tolerated, no conditions on the clothing are required, and the illumination effect is nearly eliminated.