Evolving Visual Routines

Traditional machine vision assumes that the vision system recovers a complete, labeled description of the world [Marr, 1982]. Recently, several researchers have criticized this model and proposed an alternative model which considers visual perception as a distributed collection of task-specific, context-driven visual routines [Aloimonos, 1993, Ullman, 1984]. Some of these researchers have argued that in natural living systems these visual routines are the product of natural selection [Ramachandran, 1985]. So far, researchers have handcoded task-specific visual routines for actual implementations (e.g. [Chapman, 1992]). In this paper we propose an alternative approach in which visual routines for simple tasks are created using an artificial evolution approach. We present results from a series of runs on actual camera images, in which simple routines were evolved using Genetic Programming (GP) techniques [Koza, 1992]. The results obtained are promising: the evolved routines are able to correctly process up to 98% of the test images, which is better than any algorithm we were able to write by hand. In addition, the solutions generalize to problems outside the training data. The fact that low-level, reuseable primitives can be composed into visual task solutions implies that GP is a good way to automatically create multiple visual routines.