PROJECT TITLE :

Reversion Correction and Regularized Random Walk Ranking for Saliency Detection - 2018

ABSTRACT:

In recent saliency detection research, several graph-based mostly algorithms have applied boundary priors as background queries, that may generate utterly “reversed” saliency maps if the salient objects are on the image boundaries. Moreover, these algorithms typically depend heavily on pre-processed superpixel segmentation, that could result in notable degradation in image detail options. In this Project, a unique saliency detection technique is proposed to beat the on top of problems. Initial, we have a tendency to propose a saliency reversion correction process, which locates and removes the boundary-adjacent foreground superpixels, and thereby increases the accuracy and robustness of the boundary previous-based saliency estimations. Second, we propose a regularized random walk ranking model, that introduces previous saliency estimation to every pixel in the image by taking each region and pixel image features into account, therefore leading to pixel-detailed and superpixel-independent saliency maps. Experiments are conducted on four well-recognized knowledge sets; the results indicate the superiority of our proposed methodology against 14 state-of-the-art methods, and demonstrate its general extensibility as a saliency optimization algorithm. We tend to any evaluate our technique on a brand new knowledge set comprised of images that we have a tendency to outline as boundary adjacent object saliency, on that our method performs higher than the comparison strategies.