Atenção visual bottom-up guiada por otimização via algoritmos genéticos.

Abstract

Visual attention is a biologically inspired mechanism, which corresponds to the ability of selecting and processing only the most relevant regions of a visual scene. For didactic purposes, visual attention can be divided into two main categories: bottom-up and top down. Bottom-up visual attention guides the attention focus by using primitive visual features (such as discontinuities in intensity across different scales and orientations) computed directly from the input image,without the need of any context information. Top-down visual attention, on the other side, performs a search for interest regions from higher-level features, specified in the form of previous knowledge or models about what is being sought in the scene. Themain research question that we intended to answer in this dissertation was the following: how it would be possible to incorporate some higher-level be haviour into a typical bottom-up visual attention mechanism (thus guiding the attention focus to pre-established classes of objects)? The most known bottom-up visual attention model uses several primitive feature maps to form a saliency map, which indicates the importance of the different scene regions. In this work, we assigned weights to the feature maps and developed an optimization process based on genetic algorithms running on a computational grid. Experiments involving four object classes (cars, human faces, generic objects and pistols)have been performed. The results of the optimized bottom-up mechanism have been compared with the results of a mechanism not using optimized weights and with an existing system that implemented the well known visual attention mechanism proposed by Itti et al. [Itti et al., 1998]. The results have shown an improvement of up to 30% when using the optimized mechanism. Thus,this work shows that visual attention can indeed be guided towards pre-defined regions and can be used as part of object detection systems.