In this work, I focus in a simple parameter-free statistical model that requires few training data and can be trained fast. I show that the model is well suited for face detection, person identification, and classification of facial properties.For face detection, the well known elastic bunch graph matching algorithm is adapted to learn appearance probabilities of facial features. Furthermore, texture features are transformed to be used for the detection of faces in different sizes and in-plane rotation angles. In order to place facial landmarks more reliably and to increase face recognition accuracy, images are automatically standardized according to the found scale and angle of the face. It is shown that both extensions of the elastic bunch graph matching algorithm work well with only few hand-labeled training examples and that the face detection can be accelerated.After applying small changes to the model, it can be employed for identifying a person that is shown in an image. In opposition to other state-of-the-art identification algorithms, the model learns how two facial images can be compared most reasonably. For both the intrapersonal and the extrapersonal class, each one statistical model is approximated. The intrapersonal class consists of comparisons of images showing the same person, while the extrapersonal class contains image comparisons of different identities. Utilizing face graphs, it is shown empirically that the statistical model is able to reliably recognize faces in different sizes and with different facial expressions. Identification under illumination variation is still a tough problem, but it is illustrated that the proposed model is indeed able to outperform state-of-the-art face recognition approaches. This is also reached by exploiting parts of the texture descriptors that are ignored by most current algorithms.The very same model is employed for classification of facial expression and illumination condition by simply exchanging training image pairs in the intrapersonal and extrapersonal class. Another current classification challenge is to diagnose several genetic syndromes, which have an impact on the facial appearance, by processing facial images. The statistical model is modified slightly in order to implement this classification automatically. It is shown that the proposed parameter-free model is able to classify genetic syndromes better than highly specialized classification methods with carefully chosen parameters. To be applicable in clinical practice, a Java program was developed, which allows medical personnel to administrate a database of facial images, automatically detect the faces in the images, manually adjust landmark positions if necessary, and diagnose a syndrome according to these images.The visualization of the used texture features is solved by a combination of a solid mathematical foundation with an engineer's approach. Using this reconstruction method, the possibility of visualizing modified texture features is shown. As an example, texture features of different patients with the same genetic syndrome are combined into one texture feature, in whose visualization medical experts could positively identify the corresponding syndrome.