Visual saliency refers to the distinct subjective perceptual quality that allows a particular object in a scene to stand out from the background and from other neighboring objects to grab our attention. It is a fundamental problem in several fields, ranging from neuroscience to psychology and computer vision. Simple computational frameworks to compute saliency have been studied over the past three decades.

In this paper, the authors develop a supervised-learning-based feature integration technique, which uses a random forest regressor to discriminatively integrate the saliency features for saliency computation. This is the first successful attempt using this technique, which outperforms the integration approach over saliency maps. As a single image segmentation might not be reliable enough, the final saliency map is generated using multi-level fusion over the saliency maps generated using an unsupervised graph-based image segmentation approach previously used in the literature.

Three regional saliency features were developed, namely the contrast descriptor, image-specific backgroundness descriptor, and the objectness descriptor. The latter two features were found to be more significant than the contrast descriptor feature. The authors’ proposed salient object detection technique was evaluated on several standard datasets, including MSRA-B, SED, and DUT-OMRON. Evaluation metrics included the precision-recall and receiver operating characteristic curves, as well as the mean absolute error.