Facial feature extraction for face characterization
Jeff McDermott · 2006
Research in the area of image processing and facial recognition has presented many interesting challenges. This research focuses on feature extraction but touch upon other topics such as Face detection. Simply put face detection attempts to accurately pick out a human face in an arbitrary scene, while feature extraction is the attempt to accurately isolate the desired portion of a digital image.Given a spatial database of facial images, the purpose of this research is to decrease the requisite training time for fuzzy queries. For example, one such fuzzy query might ask the database to "find all of the people with long noses." The queries are considered "fuzzy" if they involve the use of words with overloaded or ambiguous meanings. This database takes into account a community of users' perceptions and returns a set of images matching their criteria. Currently each image in the database must be evaluated by the community to ensure it will be retrieved at an appropriate time. If a facial feature could be accurately extracted from a digital image the requisite training time for this database could be significantly reduced. The community could evaluate some number of images and then the data from their feature extractions could be used to determine how similar images would have been classified.This leaves some major questions to be answered. Do extracted features correlate in some way with the actual measurements of the face? Will extracted features be reliable enough to predict how an image would be classified? Do extracted features correlate with the community's perception of a facial feature?To answer these questions we first need to employ some image processing techniques. Since the pictures could most closely be related to ID pictures, the problem of face detection is considerably simplified. For example there is only one person in the picture, they are facing forward, their face is unobstructed, and the lighting can be controlled as well as the background color. This simplified problem will allow the facial detection to be more accurate than under alternate circumstances. Facial detection is an important first step, because feature extraction limits the search space for facial features as demonstrated by figure 1.Active Shape Models (ASM)[1] will be used in the next step of extracting the features from the located face. ASM are deformable statistical models that find the best fit of a model to a digital image. Once placed on the image the model will iteratively attempt to find the best placement for all the points by moving them one at a time. Figure 2 shows an example of ASM in action.There are 22 points for ASM to identify that have been chosen by James Mastros as a part of the Virginia Commonwealth University Database Research Group headed by Dr. Lorraine Parker[4]. These points extract all the information needed for facial characterization within this research but could most surely be expanded if necessary. These points capture information like width and height of the eye, length of the nose, and others such as area of the mouth. See figure 3 for an example.Presented in this work will be information pertaining to face detection and feature extraction. Several examples of feature extraction will be available and possibly live demos should anyone wish to participate.