Saliency Detection Via Double Random Walks And Deep Multi-level Networks

Saliency detection,which aims to find the most important and interesting regions in an image,has attracted increasingly attention in recent years.As an effective pre-processing step,it has been extensively applied to numerous computer vision tasks,such as video summarization,person re-identification,image compression and object tracking etc.Although much significant progress has been made,it remains a challenging problem.First,this paper proposes a novel saliency model based on double random walks with dual restarts.Two agents(also known as walkers)respectively representing the foreground and background properties simultaneously walk on a graph to explore saliency distribution.First,we propose the propagation distance measure and use it to calculate the initial distributions of the two agents instead of using geodesic distance.Second,the two agents traverse the graph starting from their own initial distribution,and then interact with each other to correct their travel routes by the restart mechanism,which enforces the agents to return to some specific nodes with a certain probability after every movement.We define the dual restarts to take into account the interacting and weighting of two agents.To enhance the discriminative capability of features,we extract deep features from the fully convolutional network to represent each superpixel node.Extensive evaluations demonstrate that the proposed algorithm performs favorably against other state-of-the-art methods on five benchmark datasets.Second,this paper demonstrates segment-level saliency prediction can provide the proposal-level method with complementary information to improve detection results.In addition,classification loss(i.e.,softmax)can distinguish positive samples from negative ones and similarity loss(i.e.,triplet)can enlarge the contrast difference between samples with different class labels.We propose a joint optimization of the two losses to further promote the performance.Finally,a multilayer cellular automata model is incorporated to generate the final saliency map with fine shape boundary and object-level highlighting.Extensive evaluations demonstrate that the proposed algorithm performs favorably against other state-of-the-art methods on five benchmark datasets.