A Progressive Attention Guided Recurrent Network For Salient Object Detection

Salient object detection,which simulates the human vision system to judge the importance of image regions,has received increasing attention in recent years.It can be used as a pre-processing step in computer vision problems,such as scene classification,object tracking,and video compression etc.During the past years,many saliency methods have been proposed.In conventional works,hand-crafted low-level features such as color,intensity,contrast have been widely explored.Hand-crafted features are designed based on limited human knowledge.It is difficult for these methods to detect salient objects in complex scenarios.Recently,Convolutional Neural Networks(CNNs),which intelligently extract high-level and multi-scale complex representations from raw images directly,have achieved superior performance in many vision tasks.Due to the semantic information obtained from high-level feature,CNN based approaches have successfully broken the bottleneck of hand-crafted features.How to design a reasonable network which is able to learn effective features and how to process these features for saliency estimation become the key issues to be addressed.Many state-of-the-art methods design saliency models by integrating multi-level convolutional features.However,not all features are of equal importance to saliency.Attention mechanisms,which add weights on image features,provide a feasible solution.On this basis,we propose a progressive attention driven framework,which selectively integrated multi-level features and effectively suppress the distraction of background to obtain powerful features.And the saliency map is generated based on the obtained features.On the other hand,it is observed that shallower layers of backbone network lack the ability to obtain global semantic information,which limits the effective feature learning.To address the problem,we introduce multi-path recurrent feedback connections to transfer global semantic information from the top convolutional layer to shallower layers to intrinsically refine the entire network.Our proposed model is evaluated with thirteen state-of-the-art methods on six benchmark saliency detection datasets.Experimental results show that our proposed method performs favorably against other methods in form of P-R curve,F-measure,MAE and visual quality.Finally,we conduct ablation analysis of each part of the algorithm.