Salient Object Detection Based On Structured Low-Rank Representation And Background-Driven

Human visual system(HVS)has the ability to quickly capture the regions of interest from a scene.With the development of computer technology and artificial intelligence,computer vision has been widely concerned,which is to make use of computer to simulate the human visual system to recognize and understand the world.As a significant part of the computer vision,image salient object detection aims at detecting the regions of interest in a natural image.Due to its advantage of reducing the complexity of scene analysis,salient object detection plays an important preprocessing role in many computer vision tasks,including image segmentation,image classification,image recognition,to so on.However,how to accurately and efficiently detect the salient objects in the complex scenes is still a challenging research topic because of the technical difficulties of complicate scenes and different kinds of objects.Having studied the previous research works,the existing methods related to ours are summarized and analyzed.Then a salient object detection algorithm is proposed based on structured low-rank representation and background-driven.Several sets of experiments are conducted on three benchmark datasets to demonstrate the validity of the proposed algorithm.The main contents of this dissertation are summarized as follows.First,some existing image salient object detection methods are summarized.Especially,three existing algorithms are discussed in detail in this dissertation.These algorithms include the saliency optimization method based on robust background detection,the dense and sparse reconstruction error based saliency detection algorithm and the salient object detection model via structured matrix decomposition.Secondly,an image salient object detection based on structured low-rank representation and background-driven is proposed in this dissertation to address the following problems.The existing algorithms can hardly construct a more accurate background dictionary in the complex scenes,and the traditional salient object detection methods based on low-rank representation do not consider the low-rankness of the local background regions and the underlying structure of the image,which is difficult to completely and uniformly detect salient objects.The main steps of this work are summarized as follows:(1)Segment the input image into a set of superpixels.As well,color,edge and texture features are extracted as the features of superpixels.(2)An index tree is constructed to encode multi-scale structure information of the image via a hierarchical graph-based segmentation algorithm.(3)A primitive background dictionary is employed in the proposed model,which is constructed by using the coarse detection result from the VGG16 network.(4)A structured low-rank representation model is proposed for salient object detection,in which a tree-structured low-rank constraint and a local spatial consistency are imposed to capture the low-rankness of complicated background and uniform the foreground objects with diverse types of regions,respectively.(5)The final saliency map is obtained jointly based on reconstruction coefficients and reconstruction errors.Finally,the proposed algorithm is implemented in Matlab environment.The proposed model for salient object detection is evaluated on MSRA10 K,ECSSD and DUT-OMRON datasets,and shows competitive results with the state-of-the-art methods,including 4 deep learning based salient object detection algorithms.