Research Of Natural Image Segmentation

As a fundamental and important issue in computer vision,image segmentation has drawn much research attention in recent years.A reliable segmentation result is highly desirable for many higher-level computer vision applications,including image caption,autonomous driving,advanced driver assistance system,robotics,etc.How-ever,most of the object appearances in natural images involve significant substantial variations that are from pose and scale changes,occlusion,background clutter,etc.All these uncontrolled effects make the natural image segmentation become a challenging problem.In this thesis,some new methods are studied aiming at the image segmentation for natural images with complex variability.The main contributions are summarized as follows:(1)An inhomogeneity-embedded active contour(InH_ACM)is proposed for natu-ral image segmentation.The ACMs developed so far have not shown powerful perfor-mance on natural images.The reason is that natural images are rich in color,intensity or texture.The object pixels are often not artifact inhomogeneous,but inherently inhomogeneous.InH-ACM describes the inhomogeneity in natural images by a pixel inhomogeneity factor and utilizes it for segmentation,unlike most of existing methods that use some averaging convolution to reduce or remove the inhomogeneity in images.Moreover,this method builds a saliency-inspired framework that can automatically locate the initial contour for InH_ACM to start the evolution.Experimental results on several datasets indicate that the proposed InH_ACM can produce reliably satisfactory segmentation in many situations,outperforming most of current popular ACMs.(2)This paper addresses the problem of segmenting objects for natural images by leveraging multiple segmentation methods.Existing image segmentation algorithms mostly partition the image into some coherent segments instead of extracting the object entirely.It can be observed that basic elements(e.g.,superpixels)in the common segment produced by many methods are highly-correlated-they belong to the same object and generally lead to a low-rank structure.To make use of this information,this paper presents a novel approach to learn a structured low-rank affinity by leveraging various algorithms for segmenting the objects.Comprehensive experiments validate that the proposed approach achieves superior results as compared to each individual method.Moreover,our method is shown to be competitive in comparison to the state-of-the-art methods.(3)Extracting objects from natural images has long been an active problem in image processing.Despite various attempts,it has not been completely solved up to date.Current state-of-the-art object.proposal methods tend to extract a set of object segments from an image,and often there are consequential differences among these results for each image.Another type of methods strive to detect one object.with a bounding box where some background parts are often covered.For these two methodologies,this method observes:1)there are generally some regions overlapped among different proposals,which are usually from one object,and thus they could be as object segment hypotheses:2)pixels outside the detected bounding box could be as background hypotheses as they are with high probability from the background.With them,this method formulates the object.extraction as a double sparse reconstruction problem in terms of the bounding box results.The idea is that object regions should be with small reconstruction errors to segment hypotheses bases.Simultaneously,they should have large reconstruction errors to background hypotheses bases.Comprehen-sive experiments and evaluations on PASCAL VOC object segmentation dataset and GrabCut-50 database demonstrate the superiority of this method.In particular,this method achieves the state-of-the-art performance for the object segmentation with bounding box prior on these two benchmark datasets.(4)Category-specific object,segmentation has been a long-standing research top-ic in pattern recognition.This paper presents an unsupervised discriminant shape(UDS)to address category-specific object segmentation by incorporating the proposed shape prior into an intuitive energy minimization framework.Recently,based on the region proposal methods,deep Convolutional Neural Networks(CNNs)provide access to candidate segments in categories of interest from images.However,the segments obtained from bottom-up proposals tend to undershoot or overshoot objects and are easily classified into one specific class.To address this problem,this paper presents an unsupervised discriminant projection based clustering algorithm(UDC)to obtain more precise shape prior to guide the segmentation,and the class-specific proposals are clustered based on their projections onto the discriminant projection direction.Based on the set of proposals,this method then obtains the prior information of foreground UDS with an easy voting scheme.The derived UDS prior is finally utilized in the subsequent energy minimizing formulation based figure-ground segmentation.Exten-sive and comprehensive evaluations on four datasets demonstrate the effectiveness and robustness of the UDS based segmentation.