Salient Object Detection In Complex Scenes

Salient object detection aims to enable the computer to imitate the human visually attentional mechanism via designing some intelligent algorithms,leading to automatically grabbing the most attractive object or region in an image.Being a preprocessing step to greatly reduce the data processing cost and computation time,salient object detection has been widely applied in other computer vision and machine learning fields,e.g.,image segmentation,image fusion,image recovery,etc.As a result,salient object detection has important research significance and social application value.Up to now,existing salient object detection methods have achieved close-human-level performance in simple scenes(e.g.,a single salient object locating around the image center,simple background,and high contrast between foreground and background),but it is still an urgent challenge for salient object detection in complex scenes(e.g.,salient object nearing the image boundary,complicated background,low contrast between foreground and background,and rich semantic information).To address the problems encountered by the existing salient object detection methods,in this thesis,we propose several algorithms to completely and uniformly detect the salient object in complex scenes based on sparse representation,graph theory,and deep learning.The details of this thesis are as follows:First,we propose a salient object detection algorithm based on robust sparse representation and local consistency.The existing salient object detection methods based on the traditional sparse representation represent the salient object by the squared error,which is sensitive to the non-Gaussian noise.Differently,we consider the salient object as a sparse “outlier”,and represent it by the sparse error.As a result,the problem of salient object detection is reformulated as the sparsity pursuit problem.This is the first attempt to apply robust sparse representation model for salient object detection.Moreover,two Laplacian regularizations are enforced on the consistency of representation coefficients and reconstruction errors of locally adjacent superpixels.Therefore,a local consistency is embedded into the robust sparse representation model,improving the smoothness of the saliency map.Compared with those salient object detection methods based on the traditional sparse representation,the proposed algorithm promotes the representation ability of the salient object and the uniformity of the saliency map.Secondly,we propose a salient object detection algorithm based on two-stage graphs.The above proposed method based on robust sparse representation and local consistency considers only the local correlations between the current superpixel and its neighbors,leading to some limitations for the consistency of the saliency object.To this end,we employ the graph theory to solve the problem of salient object detection.Existing graph based methods consider only a local consistency to construct a graph.Differently,we design two different graphs in two stages in a coarse to fine way.At the first stage,a graph is constructed based on the adjacently spatial consistency,and a weighted jointly robust sparse representation model is designed to calculate the saliency value of each node.At the second stage,based on the coarse detection results of the first stage,we determine the potential foreground nodes and the potential background nodes.A regionally spatial consistency is designed by connecting any pairs of potential foreground/background nodes.A refined graph is constructed to achieve a more accurate saliency map by combining the adjacently spatial consistency and the regionally spatial consistency.Compared with existing graph based salient object detection methods,the proposed algorithm enhances the contrast between foreground and background,foreground uniformity,and background uniformity,resulting in more uniform saliency maps.Thirdly,we propose a deep salient object detection algorithm based on contextual information guidance.In recent years,due to the powerful feature representation ability,deep learning has been successfully applied in salient object detection,and has broken through the performance bottleneck of the traditional methods that are based on hand-crafted features.Existing deep learning based salient object detectors integrate multi-level deep information by a concatenation of multi-level feature maps or an element-wise addition of multi-level side outputs,which have some limitations.Differently,we propose a contextual guidance strategy to apply deep side output to guide shallow feature maps.This enables that shallow layers can learn more accurate salient features with the help of deep side outputs,and deep side outputs can be propagated to large-resolution versions with the aid of shallow-level feature maps.Moreover,we propose a group convolution module to learn more discriminate feature maps.Finally,the group convolutional module is embedded into the guidance strategy to further promote the guidance roles of deep side outputs.Compared with existing salient object detectors based on multi-level deep information integration,the proposed algorithm more efficiently incorporates feature maps and side outputs,improving the accuracy of salient object location,and the completeness and uniformity of salient object segmentation.Fourthly,we propose a deep salient object detection algorithm based on a two-stream partobject assignment network.Existing deep learning based salient object detection methods mostly compute the saliency of each image part separately.This easily leads to incomplete/inconsistent salient object segmentation.To solve this problem,we involve the property of part-object relationships for salient object detection based on the fact that a salient object is usually composed of several associate parts and several associate parts can form a complete object.We adopt the capsule network(Caps Net)to explore the part-object relationships within an image to detect the salient object.This is the first attempt to apply Caps Net for salient object detection.Moreover,we propose a two-stream strategy to feed the groups of input capsules into two identical streams.Caps Net is adopted to explore the part-object relationships in each stream.The proposed two-stream strategy not only reduces the network parameters and computation complexity,but also alleviates the redundancy of capsule matchings.Compared with existing deep learning based salient object detection methods,the proposed algorithm can effectively solve the problem of incomplete/inconsistent salient object detection.Objective and visual evaluations on several public benchmark datasets efficiently validate the effectiveness and superiority of the above four proposed salient object detection algorithms.