A novel approach to perceptual knowledge extraction from images based on the concept of salient image objects is proposed. Salient image object - a concise description of a image fragment within a circular region - is a vector of salient image features, which describes the fragment invariantly to geometrical transformations and some intensity changes. Bayesian network of salient image objects - a kind of generative image modeling - is used as a model for the knowledge representation, which includes semantic entities (e.g., real-world objects) and provides probabilistic relations between image features and semantic entities. The proposed technique of multi-scale image relevance function permits a fast and ordered extraction of salient image objects.