Research On Combined Method For Image Object Segmentation

Image segmentation is a key part of Image Analysis and Computer Vision, also one of the supporting techniques of object-based image or video coding. And it has many perspective applications. However, most existing methods consider only low-level image features. Thus they can only partition the image into a set of regions with homogeneous low-level features, but not into meaningful objects. This reduces the utilities of segmentation technology dramatically, and causes it to be a bottleneck restricting the development of related fields. To overcome this bottleneck, it is necessary to research on methods aiming at image objects extraction.Studies on vision mechanism have proven that prior knowledge about object features plays a crucial role in the course of object segmentation. However, top-down methods using high-level features alone always cannot achieve high segmentation precision. In view of this and under the illumination of Systems Engineering, this thesis takes an integrated way to address the problem of image object segmentation. It considers the problem of bottom-up segmentation as well as that of top-down segmentation. And more importantly, it manages to combine them together in a systematic way, which makes them benefit from each other and, hence improves the ultimate segmentation accuracy. To make the work more technologically feasible, this thesis limits its researching domain to class-specific image object segmentation. More detailedly, it makes the following contributions:1. It proposes the concept of Image Object Segmentation and makes a comprehensive investigation on related works. First, the thesis introduces the formal definition of image segmentation and presents its hierarchical conceptual framework. Based upon this, it clarifies the concept of Image Object Segmentation. After that, a survey of existing segmentation methods is performed. It is notable that this survey covers the recently reported object detection and segmentation methods incorporating high-level features which are not included by existing overviewing literatures.2. Though bottom-up segmentation methods cannot extract meaningful objects, they are able to detect image region boundaries accurately, which is helpful for improving the precision of object boundary location. Therefore, the thesis researches firstly on bottom-up segmentation, mainly of color images.(1) Color quantization is adopted to reduce the data complexity in color images. And the LBG algorithm is borrowed to implement this process. Since clustering results of LBG algorithm depend greatly on the initial palette colors, the thesis proposes two complementary initial palette selection schemes: the Merging Scheme and the Popularity Scheme. Experiments prove that they can improve quantization results and reduce iterations considerably. Moreover, to speed up the quantization process, a heuristic palette searching algorithm is presented, whose name is ENESPS (Equal-Norm Equal Sum Palette Searching). Experimental results show its power of enhancing speed.(2) Distribution evenness of colors is proposed as a quantitative measure for color texture description and analysis. Upon this, a bottom-up segmentation method ISBEC (Image Segmentation Based on distribution Evenness of Colors) is proposed. It uses the resulted color index map from color quantization as input for simultaneous color segmentation and multi-scale texture analysis. Region growing technique is exploited to implement color-texture segmentation. Experimental results display its fairly good segmentation abilities. Furthermore, replacing color quantization with gray-scale quantization, ISBEC is also capable of segmenting gray-scale images.3. A top-down segmentation method, SBIOSEG (Shape-element Based Image Object SEGmentation), is proposed. It makes use of shape feature of interested class to segment objects. This method formulates object segmentation as a binary classification problem. It contains two stages. At the learning stage, AdaBoost algorithm is adopted to build up the Shape-element Codebook of the class and to train the classifier. At the testing stage, segmentation is implemented in a coarse-to-fine way, i.e., detect object first, and then segment it. Experiments verify SBIOSEG's capability of class-specific image object segmentation. And results also show it to be superior to existing top-down method.4. Both bottom-up and top-down segmentation methods have their advantages and disadvantages. But luckily, their relative merits suggest that they can be combined together to achieve higher object segmentation accuracy. Observing this, the thesis proposes two schemes to integrate bottom-up and top-down segmentation together. And it presents a combined object segementation method, B&T-IOSEG (Bottom-up and Top-down combined Image Object SEGmentation), which adopts both integration schemes. Experimental results display that, compared with SBIOSEG, the B&T-IOSEG method can achieve remarkably higher segmentation precision. Furthermore, the thesis introduces the prototype design and implementation of a class-specific image object segmentation system, and discusses its application in constructing 3D visual models of objects from images.