Visual perception and selective image analysis

This paper describes organization principles, architecture, and basic algorithms of a vision system with goal-directed perception. A model of such a system which performs selective analysis of the scene depending on the current vision task is described. The tasks include: extraction of predetermined features and recognition of their location; selective extraction and identification of the target object; segmentation of the visual field into “areas of interest” to provide the visual information basis for selective image analysis. A gradient field, the result of processing the gray-scale image using a local operator, serves as the source of initial information for the visual analysis. At the upper level of the visual analysis the goals of the analysis are being determined, then the anticipated model of the object is extracted from a database, and this model is split over the levels of the analysis into the following parts: integral characteristics of the pattern (for fast detection of similar objects in the current image); models of specific spatial features of the pattern-a vector of features which indicate a set of possible local specificities of the patterns. In the process of the goal-directed analysis the system functions in response to demands from the upper level, realizing the top-down strategy of visual perception. During recognition process a goal-directed image analysis is performed through hierarchical processing and extraction of features in local receptive fields. Further on the current structural description is being formed which is compared with the description of the target object, and the degree of similarity between these descriptions is calculated. An object is regarded as identified if the degree of similarity exceeds a predetermined threshold value. An experimental study has been performed of segmentation of a series of images obtained with the help of videosensor mounted on the wrist flange of an anthropomorphic arm