Graph-based Salient Object Detection –Algorithms And Applications

In order to analyze real world information,the human visual system uses its visual attention mechanisms to identify meaningful information from a large amount of input information.Inspired by the selective attention mechanism,saliency detection uses computer vision algorithms to simulate the human visual characteristics.As a branch of visual saliency,salient object detection aims to find the most attractive region in an image.It can be applied to reduce computation time and improve the performance of algorithms in many visual applications,such as image compression,image segmentation and medical image processing.In recent years,graph-based salient object detection methods have attracted increasing attention because of their simplicity and efficiency.This thesis focuses on the algorithms and applications of salient object detection based on graph models,proposes two graph-based models of salient object detection,and applies the saliency detection technology to vessel segmentation of the optical coherence tomography angiography(OCTA)images.The main works and contributions are summarized as follows:A deformed-smoothness-constraint-based manifold ranking model is proposed and applied in the salient object detection.The graph-based standard smoothness constraints may ignore some salient object regions of low contrast against the background,while the deformed smoothness constraint is able to prevent erroneous label propagation by regularizing nodes and their neighbors locally.Thus,the object regions of low contrast against the background can be merged.Besides,the deformed smoothness constraint is further utilized in a map refinement model.It integrates the objectness feature of salient object,which can suppress the background noises in label propagation result.The results of experiments on public datasets ASD,PASCAL-S,DUT-OMRON,ECSSD and THUR15 K show the effectiveness of the proposed model.A saliency propagation algorithm with perceptual cues and background-excluded seeds is proposed.In salient object detection,the effectiveness of the label propagation of graph model relies on the accuracy of seed selection.The seed acquirement ways of the existing methods are easy to mix with background nodes.This problem is solved by a proposed graph-based framework,which incorporates perceptual cues into the framework and uses the backgroundexcluded seeds to propagate saliency.Firstly,a graph is constructed by two perceptual cues,i.e.proximity and similarity.Secondly,probable background nodes are generated by a novel background probability measuring method.These nodes are excluded from the selected seeds in order to pick out more reliable seeds.Then a label propagation model is developed to diffuse saliency based on these reliable seeds.Lastly,another perceptual cue called rareness is integrated into a cost function to optimize the propagation result.The results of experiments on public datasets ASD,PASCAL-S,DUT-OMRON,ECSSD and THUR15 K demonstrate the effectiveness of the proposed algorithm.A saliency-map-based level set image segmentation model is proposed and used for the OCTA vessel segmentation.Most target objects of image segmentation are the area that human eyes pay attention to in the scene.From this perspective,saliency map based fitting terms are added in the level set energy function of the infinite perimeter active contour model.This new model can use the information of saliency map to assist the image segmentation.Furthermore,a saliency driven retinal vascular segmentation algorithm is proposed based on this model.Firstly,compactness and rareness is utilized to detect salient regions of fovea and obtain background seeds,and boundary connectivity is used to generate foreground seeds.Secondly,these two kinds of seeds are put into the regularized random walk model to obtain the OCTA vessel saliency map.Finally,the OCTA vessel saliency map is integrated into the proposed saliencymap-based level set image segmentation model,and this model is then used to segment the vessels in OCTA images.The results of experiments on the real OCTA dataset collected from St Paul's Eye Unit of Royal Liverpool University Hospital show the effectiveness of the vessel segmentation model.Two low-rank cooperative OCTA destriping models are proposed.Stripe noise occurs during the OCTA acquisition process due to the involuntary movement of the eye.In order to remove the stripe noise effectively,two novel low-rank based models,i.e.,cooperative uniformity destriping model and cooperative similarity destriping model,are proposed to remove stripe noise at different layers of the OCTA images from the same eye cooperatively.Both models are based on the low-rank optimization model,but the assumptions of the two models for stripe noise are different.The cooperative uniformity destriping model assumes that stripe noise is identical across all the layers,while the cooperative similarity destriping model assumes that the stripe noise at different layers is not identical but similar.The two proposed models are tested and compared on the synthetic OCTA dataset and the real OCTA dataset collected from St Paul's Eye Unit of Royal Liverpool University Hospital.The results of experiments demonstrate not only the effectiveness of the two models,but also their beneficiary effect on the vessel segmentation of OCTA images.