A central debate in visual perception theory is the argument for indirect versus direct perception; i.e., the use of intermediate, abstract, and hierarchical representations versus direct semantic interpretation of images through interaction with the outside world. We present a content-based representation that combines both approaches. The previously developed Visual Alphabet method is extended with a hierarchy of representations, each level feeding into the next one, but based on features that are not abstract but directly relevant to the task at hand. Explorative benchmark experiments are carried out on face images to investigate and explain the impact of the key parameters such as pattern size, number of prototypes, and distance measures used. Results show that adding an additional middle layer improves results, by encoding the spatial co-occurrence of lower-level pattern prototypes.