Homeostatic image perception: An artificial system

This paper describes how a visual system can automatically define features of interest from the observation of a large enough number of natural images. The principle complements the low-level feature extractors provided by PCA filters by analyzing their spatial interactions. This is achieved by modeling an internal representation in the system, composed with ternary variables obtained by thresholding the filters, using a Markov Random Field model. A stochastic gradient algorithm, based on statistics computed from an image database, is used to train this model. The result is a probability distribution on the internal state of the system which adjusts with its environment, under what is referred to as a principle of homeostasis. When new images enter the system, they are confronted to this internal distribution, and images which appear as salient in this regard are detected as visually relevant. A classification of these relevant images is provided, as an illustration of the model.