Deep CNN Features for Visual Saliency Estimation
Aymen Azaza, Ali Douik · 2018
Visual saliency detection is an important task for researchers in the field of computer vision. Recently, object proposal algorithms have been applied in the field of saliency detection. Object proposal methods provide image segments as proposals (not bounding boxes) which can be used for saliency estimation, also much recently the use of deep convolutional networks breakthroughs computer vision with the extraction of the powerful features using deep CNN, which is more efficient compared to designing manually hand-crafted features. In this context, we tried to develop a saliency algorithm based on the combination of object proposals with deep features, and we propose several saliency features which are computed from deep features combined with object proposal. We train an SVM to predict the saliency of every object proposal. Experimental results shows that we outperform other state of the art method in two data sets in terms of F-score, PR curves and in term of computational speed.