Task-based visual saliency for intelligent compression

Abstract

In this paper we develop a new method for highlighting visually salient regions of an image based upon a known visual search task. The proposed method uses a robust model of instantaneous visual attention (i.e. “bottom-up”) combined with a pixel probability map derived from the automatic detection of a previously-seen object (task-dependent i.e. “top-down”). The objects to be recognised are parameterised quickly in advance by a viewpoint-invariant spatial distribution of SURF interest-points. The bottom-up and top-down object probability images are fused to produce a task-dependent saliency map. We validate our method using observer eye-tracker data collected under object search-and-count tasking. Our method shows 10% higher overlap with true attention areas under task compared to bottom-up saliency alone. The new combined saliency map is further used to develop a new intelligent compression technique which is an extension of DCT encoding. We demonstrate our technique on surveillance-style footage throughout.