Abstract
By restricting `what-to-be-observed' in naturally complex scenes to `easy-to-compute' patterns, this paper presents a stochastic systems approach to the stabilization of early visual perception. In this approach, attentional landmarks are identified with saliency images latently specified in terms of a set of generic parameters. To stabilize the EM algorithm for the estimation of the latent parameters,a truncated version of innovation process is visualized via the multiplexing of residual distribution spanning entire the scene images; resulted visualization, simultaneously, yields an effective cue to confine the images of not-yet-identified landmarks. Based on experimental studies, it has been demonstrated that the focus control algorithm is available in the analysis of the first visit scenes; in this algorithm, it is sufficient to a priori define the attentional landmarks in terms of generic rules governing the perspective of the scene images; in contrast with conventional algorithms, no adhoc pre-processing is required to extract the saliency images under not-yet-identified photographing conditions.
