Many visual objects can be decomposed into a set of individual features related with some prior spatial relations. It is desirable to incorporate both data-driven (individual feature characteristics) and model-driven knowledge (spatial relations) to extract this class of objects. This is the rationale of the spring model [7], which becomes one of the cornerstones of our work. The objects of interest are not limited to rigid objects, i.e., different instances of the object may have different perturbations. Typical examples are human faces and Chinese characters. Thus, given a complete interpretation of the object in the image, it is not enough to check whether it matches with the spatial relations, but have to quantify how much it distorts from them. To achieve this, a new computation mo...[ Read more ]

Many visual objects can be decomposed into a set of individual features related with some prior spatial relations. It is desirable to incorporate both data-driven (individual feature characteristics) and model-driven knowledge (spatial relations) to extract this class of objects. This is the rationale of the spring model [7], which becomes one of the cornerstones of our work. The objects of interest are not limited to rigid objects, i.e., different instances of the object may have different perturbations. Typical examples are human faces and Chinese characters. Thus, given a complete interpretation of the object in the image, it is not enough to check whether it matches with the spatial relations, but have to quantify how much it distorts from them. To achieve this, a new computation model is formulated, pinpointing some deficiencies identified in the spring model. The model exhibits some nice properties, such as rotation and scale invariance.

In computational aspects, our work is related to the active contour model (snake) [17, 1, 29, 20]. We observe the generality of its computational framework in a wide class of problems other than boundary detection (i.e., by appropriately redefining the cost functionals in the model, it can solve different problems). The solution is obtained by energy minimization, and dynamic programming is used as the optimization algorithm.

To realize the usefulness of the general methodology, it is applied in the problem of facial feature extraction. Face analysis has developed as a separate branch of research, and several topics have emerged within it, such as face recognition [26], face coding [28], lip reading [6] and recognition of human expressions [18]. Examples of real-life applications include security system, credit-card verification, videophone system, low-bandwidth teleconferencing, criminal identification, and measurement of driver awareness. Facial feature extraction is fundamental in the aforementioned topics of face analysis. A prototype system is developed which can extract facial feature locations in complex scenes, with arbitrary tilt of heads and varying scales.