An efficient strategy for estimation of visually salient regions in images

Abstract

The information incident on the human visual system is bound by a selection mechanism, known as visual attention. This mechanism is responsible for restricting incomingvisual information to a smaller and potentially important subset for further processing,enabling the visual system to respond rapidly, despite the enormous amount of information to which it is subject. Computer vision systems often employ reproductionsof this mechanism in order to reduce visual search space, since this strategy can leadto substantial improvement in efficiency. This thesis addresses the problem of efficient computation of visual attention, particularly the case of salient region detection. A strategy based on joint upsampling of coarse-scale saliency estimates is presented for that purpose. This approach allows leveraging both the advantages of coarsescale estimation (reduction of computational cost, abstraction of unnecessary details) and fine-scale edge information (high accuracy). Based on the highly redundant data and spatially-varying importance of content in images of real-world scenes, two visual saliency formulations are presented for coarse-scale estimation in the proposed strategy. The first approach operates on a pixel level, based on random color distances. The second approach operates on a patch level, based on the reconstruction error computed from the Principal Component Analysis of the image boundaries. The efficacy and efficiency of the proposed strategy are demonstrated through assessment on the ASD, MSRA10K, ECSSD, and DUT-OMRON datasets. Comparison with other seven state-of-the-art methods in terms of precision, recall, F-measure, and execution time demonstrate that the proposed strategy is highly competitive, achieving one of the best trade-offs between accuracy and execution time.