Target search by bottom-up and top-down fuzzy information

Abstract

One of the basic tasks assigned to the attentional mechanism is to decide which location in the visual field we must pay attention first. An object containing a distinctive feature can attract attention in a bottom-up way. By comparing one object with the others present in the scene, bottom-up conspicuity features are used to guide attention to the most different object. Top-down hints are based on the previous knowledge about the objects or on which features are important to locate them and also have a large influence on the attended locations. Inspired on the mechanisms of human visual attention we developed a new methodology to integrate bottom-up and top-down information by using fuzzy net containing three fuzzy subsystems. The first bottom-up subsystem allow us to combine features and infer with great flexibility some intuitive decision rules based on the visual perception principles such as the Gestalt laws. The second top-down subsystem combines different features according to the relevance of them in different tasks. Finally, the last subsystem integrates the information of the previous systems and gives a general salience index. The new methodology was tested in geometrical objects considering the features that attracts attention to human beings.

Feature Gate Computational Top-Down Model for Target Detection

“…In [5], the author derived the TD (Top Down) map generated using the decision tree classifier on the test date. A fuzzy rule-based system was used in [6], [7] and [8], where the feature is used for the specific image property (Intentional features) introduced in fuzzy inference system on learned features. A Visual Attention System for Object Detection and Goal-Directed Search (VOCUS) [9] model derived the TD map by tuning the features of the target region in the map.…”

Visual attention-based polygon level of detail management

“…Much of the work in image processing has been basic research aimed at the detection of regions of interest to model human visual attention [Osberger 2001;De Almeida Neves 2000;Milanese 1993;Itti 1998]. However, some have gone further and utilised the regions of interest to perform optimisations on image processing algorithms, usually in the areas of image and video compression [Osberger 2001].…”