The Application Of Perceptual Organization In Image Prominence And Visual Attention Prediction

Perceptual organization, including perceptual segmentation and figure-ground organization, plays a key role in the process of human visual perception. Perceptual segmentation deals with image signal and object boundary information originated from retina, and therefore forms the concept of image regions; figure-ground organization further classifies the obtained regions into foreground and background, which greatly reduce the cost and complexity of following object recognition. Several perceptual organization cues have been identified by psychologists after decades of research, including convexity, symmetry, surroundedness and orientation, etc. Each of them play an essential role in figure-ground organization. Among these cues, convexity implies that, if there’s a curved boundary between two adjacent regions, then the region on the convex side tends to be foreground; symmetry implies that, region with high symmetry has more possibility to be foreground; surroundedness means that, region fully surrounded by other regions are more likely to be foreground. We can make use of these cues in multiple computer vision tasks, like foreground segmentation, image saliency detection, and attention prediction, etc.In this article, we propose an automatic convexity detection algorithm, and validate the algorithm on real image datasets. We find our proposed algorithm not only accords with conclusions from artificial psychology experiments, but is also effective on natural image. Therefore, in the task of image saliency detection, we utilize the convexity detection algorithm, together with a hierarchical segmentation model, to form a weighted graph. After adjusting weight between nodes using convexity context window, we use a graph cut to obtain final saliency detection results. Moreover, in attention prediction, we calculate convexity, symmetry as well as surroundedness, and construct foreground maps for images. Therefore we predict attention using this foreground map together with bottom-up low-level feature saliency and top-down object detection. In our experiments, we validate the two proposed framework and evaluation result demonstrates that saliency detection using perceptual organization feature performs as well as algorithms using low-level features. Furthermore, accuracy for attention prediction is also enhanced when foreground map is involved.