A Fuzzy Logic Based Top-Down Attention Modulation Framework for Selective Observation

The work presents a framework for top-down modulating the visual process under intention which is typically represented by words or short sentences. A fuzzy logic mapping process is developed to ground the word-form intention into a suitable weight vector to combine different visual feature channels during local saliency representation. Since the knowledge about the time-frequency characteristic of each feature extractor and their potential performance contribution to the higher-level qualitative intention is embedded, such a mapping solution is then feasible. Employing this top-down fuzzy logic mapping process, a novel task-oriented visual attention model for selective observation is then proposed. Implementing the model with a computational one, related experiments on traffic scene images has been done. Their result illustrates that the goal of selective surveillance on certain vehicle speed can be validly achieved. It implies that this framework is a competent model for visual selective attention, which expands the way to implement a computational mechanism for top-down modulating the bottom-up process especially in the case of task-related attentional action in machine vision systems.