Deep saliency quality assessment network

In this paper, we propose to predict the quality score of saliency map by only looking over the saliency map itself. In order to achieve this goal, we propose deep saliency quality assessment network (DSQAN), to predict the saliency quality scores directly from saliency maps. First of all, we model this saliency quality assessment task as a regression problem. To train an efficient regression model, we utilize state-of-the-art deep convolutional neural networks, and reform the canonical architecture, which is used to accomplish image classification problem, to regress the saliency quality scores of saliency maps. As a direct application of the proposed DSQAN, the predicted saliency quality scores can be utilized to rank the saliency maps or choose the best-K saliency maps from a set of saliency maps. The experimental results on the MSRA10K dataset demonstrate that our proposed method has the ability to precisely predict the quality of a given saliency map. The quantitative experimental results show that the prediction error is less than 5.3% in terms of MAE score on the whole test set. We also show the ranking results using the predicted saliency quality scores to show the effectiveness of proposed method.