A Randomized Framework for Estimating Image Saliency Through Sparse Signal Reconstruction

This article proposes a randomized framework that estimates image saliency through sparse signal reconstruction. The authors simulate the measuring process of ground-truth saliency and assume that an image is free-viewed by several subjects. In the free-viewing process, each subject attends to a limited number of regions randomly selected, and a mental map of the image is reconstructed by using the subject-specific prior knowledge. By assuming that a region is difficult to be reconstructed will become conspicuous, the authors represent the prior knowledge of a subject by a dictionary of sparse bases pre-trained on random images and estimate the conspicuity score of a region according to the activation costs of sparse bases as well as the sparse reconstruction error. Finally, the saliency map of an image is generated by summing up all conspicuity maps obtained. Experimental results show proposed approach achieves impressive performance in comparisons with 16 state-of-the-art approaches.