Research On Image Cosegmentation

Image segmentation is the fundamental research for many computer vision and multimedia processing applications. Recent practice applications mainly focus on web image dataset, which is usually very large and contains much image information. This requires the segmentation on image group rather than traditional single image. Hence, cosegmentation is proposed to discovery foreground regions from a group of images. Although several cosegmentation methods have been proposed in the last several years, the key problems of cosegmentation are still unresolved, such as the design, and theoretical analysis of the cosegmentation model, the semantic similarity evaluation of the common objects, and the unsuccessful cosegmentation performances in practice applications. Cosegmentation thus becomes an urgent and challenging research task in computer vision and multimedia processing. In this thesis, we focus on the research of cosegmentation.Based on the number of common objects and the difficulties of cosegmentation, this thesis includes a series of cosegmentation researches, such as single class cosegmentation, multiple class cosegmentation and cluttered background cosegmentation. In each research, we focus on several cosegmentation key problems, e.g., model design and foreground consistency evaluation. Furthermore, we apply the cosegmentation to improve the practice applications, such as object detection and extraction. The details are listed as follows.1. We first research on color based low-level cosegmentation problem. In this research, a color feature based foreground similarity measurement and an active contours based cosegmentation model are designed, which can segment similar color common objects from a group of images. This model also provides a fundamental framework for the model design and theoretical analysis of more complicated cosegmentation problem.2. Since the color based cosegmentation can not handle color variation very well, we next research on middle-level feature based cosegmentation, such as saliency and shape middle-level features. Two models, i.e., saliency and shortest path based cosegmentation,and directional shape feature and median graph theory based shape model generation and matching model are proposed. The first one introduces saliency information to handle color variations. The second one focuses on the shape matching and model generation from a group of images. The experimental results show that the middle-level feature based models can achieve more robust cosegmentation performance with color variations.3. Because the foreground consistency evaluation is usually unknown in cosegmentation, we research on the adaptive foreground similarity measurement and propose image complexity based feature adaptive cosegmentation model. The model includes image complexity analysis method, foreground semantic similarity evaluation model and the learning of the model parameter. This model is able to adaptively learn the foreground similarity measurement from an unknown image group, which significantly extends the practice applications of the cosegmentation.4. Furthermore, we research on multiple class based cosegmentation, and propose two cosegmentation methods, i.e., directional graph clustering based multiple foreground cosegmentation and multiple group cosegmentation. The first model formulates the multiple foreground extraction as clustering problems and combines the segment propagation strategy to achieve multiple foreground extraction. The second model tries to extract the common objects when there are multiple image groups. The foreground information translation among the image groups is introduced, which can provide more object prior information and result in the improvement of the cosegmentation performance.5.Since it is difficult to segment the common objects from complicated background,we propose a similar scene cosegmentation model. The research is based on the observation that many images are usually taken under the same background scene. Based on the active contours cosegmentation model, we introduce the background consistency in the cosegmentation and implicitly provide the foreground information by modeling the background prior. The experimental results show the successful segmentation of the interesting objects from similar background scene.6. We finally apply the cosegmentation model in practice applications, and propose a logo based object segmentation model. Starting from the shape model generated from cosegmentation, a shape representation based on the relationship between the logo and the object boundary is proposed. Furthermore, the corresponding shape matching method that is robust to shape variations is introduced to locate the object boundary from unknown images. The experimental results show the successful performance of the proposed method in web e-business images.