Region‐based saliency detection

Abstract

In this study, the authors propose an unsupervised approach to detect saliency of each pixel in an image. The proposed region‐based pixel‐wise saliency detection approach produces full resolution (same as that of the original image) saliency map and precisely locates visually prominent region/object of interest in the input image. There are two parts in the authors approach. In the first phase, they partition the input image into homogeneous regions using split‐and‐merge technique. In the second phase, they rank the regions based on its proximity to the centre of the image, visual significance, size and completeness. Based on the ranking of the regions, the significance of each pixel is computed. The proposed saliency detection approaches improves the accuracy of content‐based applications such as salient object segmentation and content aware image resizing. Experimental results show that their proposed approach qualitatively better than the state‐of‐art approaches and quantitatively comparable to ground truth information which are collected from human observers.