Extracting Salient Objects from Operator-Framed Images

In images framed by human operators, as opposed to those taken under computer control, the position of objects can be an important clue to saliency. This paper uses the Berkeley image data set to show how locational and photometric information can be combined to extract a probability of saliency for all image pixels. This probability can then be thresholded and segmented to extract compact image regions with high probability of saliency. A self assessment procedure allows the algorithm to evaluate the accuracy of its results. The method can extract salient regions of non uniform color, brightness or texture against highly variable background.