Font Size: a A A

Studies On Image Segmentation Methods Based On Visual Saliency

Posted on:2015-03-26Degree:DoctorType:Dissertation
Country:ChinaCandidate:X F BaiFull Text:PDF
GTID:1108330461485140Subject:Systems Engineering
Abstract/Summary:Request the full-text of this thesis
Image segmentation refers to the processing that divides the whole im-age region into several non-overlap sub-regions according to some similarity criteria, so that pixels in the same region have the same characteristics, but pixels between different regions have the lower similarity. The segmentation results are always some compact and meaningful representations about the im-age, such as object edges, contours, superpixels, sub-regions, etc; and these representations can be used for some higher vision tasks including subsequent object recognition and scene understanding. Therefore, image segmentation is a key and fundamental step from the low-level image processing to high-level image analysis and understanding, and it is a research focus in computer vision, image processing and other related areas. In addition, with the continuously developments of image segmentation applications and image acquisition meth-ods, more and more image segmentation techniques are applied to the internet and mobile devices, such as image classification and marking, image thumb-nail generation. These applications require the image segmentation methods to be fast, accurate, automatic and adaptive.Although a lot of image segmentation methods have been proposed cur-rently, there are still some problems unresolved, such as poor general, user interaction dependency, discrepancies between the segmentation results and human perception. To deal with these problems, in this thesis, key issues in automatic image segmentation approaches are intensively studied focusing on the theories and applications of visual saliency, in order to effectively solve the automatic segmentation of active contour model, adaptive extraction of initial curve, construction and level set implementation of saliency based active con-tour model, automatic generation and selection of S VM training samples, and other related problems. The main research work of this thesis are as follows:(1) Due to the traditional active contour models are sensitive to initial curve, an adaptive initial curve extraction method is proposed based on visual saliency detection mechanism in visual cognition theory. The potential posi-tions of salient objects are detected by saliency map, and some morphological operations are used to specify the prediction results more accurately, so that to obtain the contour information of salient object. Experiment results show that this presented method can accurately detect and extract the prior shape information in the image adaptive and automatically.(2) Saliency information is an important image feature. A modified CV model with adaptive initial curve is proposed, in which the prior shape in-formation of object is explicitly embedded into the geometric active contour model. And a novel region saliency based active contour model is presented too. In the partial differential equation solving process of active contour model, the level set function is constructed according to the prior shape information of the object, making the initial curve to be the maximum approximation of true object boundary, so that to greatly reduce the number of iterations to con-verge to the object boundary. Experiment results on synthetic images, natural images and medical images show that the proposed method can segment the image adaptive and automatically, and effectively reduce the impact of the ini-tial curve to segmentation performance of the active contour model as well.(3) SVM method based on statistic learning theory has been widely ap-plied in image segmentation due to its excellent classification performance. However, SVM is essentially a supervised learning method, and cannot ob-tain training samples when used for image segmentation. To this end, a novel method based on SVM and saliency is presented to automatically select the training samples of SVM classifier and segment the image. A "trimap" derived from saliency map can provide the spatial feature of foreground and back-ground, which are then combined with color features achieved from histogram analysis to generate the training datasets of foreground and background, and then, training pixels are chosen from these two datasets according to a local homogeneity threshold. From the viewpoint of segmenting as a whole, the proposed SVM image segmentation method is driven entirely by the visual characteristics of the image itself. Experiment results show that the segmenta-tion performance of the proposed method is much better than other SVM based methods, and the automatic selection of SVM training sample also ensures the stability of the segmentation algorithm.(4) SVM for image segmentation method based on saliency and super-pixel is proposed in order to further improve the segmentation performance of SVM classifier. In this method, the basic processing unit is superpixel instead of pixels. Foreground and background superpixel sets are generated automati-cally by means of saliency information of superpixels stemmed from saliency map. And then, some superpixels are selected from these two sets and feature vectors of each superpixel are extracted to train a SVM classifier. Compared with the traditional pixel-wise processing, the visual granularity of superpixel is more consistent with the general processing of visual information received by human visual system. Experiment results demonstrates that using super-pixel comprised of neighboring pixels with similar visual features, making the feature distribution information of image local regions are well exploited, and the algorithm efficiency is boosted through decreasing the data amount on the other hand.(5) In order to provide a platform for experiment analysis and comparison of image segmentation researchers, this paper designs and implements an in-tegrated image segmentation system, including the management of image seg-mentation datasets, image pre-processing, image segmentation methods, im-age segmentation performance evaluation and other functions. The integrated system has good practicability, interactivity and scalability.Studies of image segmentation method relate to many research fields such as computer vision, image processing, pattern recognition and machine learn-ing. In-depth study and effective solutions of image segmentation can greatly promote and facilitate the development and maturation of these disciplines, and to provide an important inspiration to solve other complex pattern recognition problems.
Keywords/Search Tags:Image Segmentation, Visual Saliency, Active Contour Model, Support Vector Machine, Superpixel
Request the full-text of this thesis
Related items