Research Of Graphics And Image Segmentation Methods Based On The Human Visual Characteristics

With the rapid development of computer science and mathematical theory, graphics and image processing has become a very active research fields. In this field, graphics and image segmentation have been important research subjects. After many years of development, segmentation technology has been widely used in many fields, such as visualization, computer animation, medical image processing, and virtual reality.Many image processing tasks need to extract specific regions in the image, we can accomplish this operation by means of image segmentation technology. Since the subsequent visual tasks will be affected by the quality of the segmentation result, image segmentation method become a critical step in the process of the image underlying to image recognition and understanding. Human beings can separate the target region from the image accurately, but it is not an easy task for the computer. In many cases, due to the influence of image quality and content diversity, it is difficult to make the computer segment images according to the human understanding. Most segmentation methods divide the image into disjoint regions based on the criterion of regional similarity, or attribute differences, but rarely using the characteristic of human visual in the segmentation process, and the result of image segmentation is far from human visual perception. Therefore, how combining human visual characteristics with image segmentation technology to produce more reasonable segmentation results need to be further research.With the rapid development of data acquisition device and modeling technology, 3D graphic model has become a new digital media representation, and the associated processing of 3D model also become a research focus in the field of computer graphics. Similar to the case of image segmentation, we still hope that 3D models can be segmented into meaningful sub-components by the aid of human visual characteristics. However because of human subjectivity, different people hold different interpretations of the "meaningful" parts, and the definition of "meaningful" segmentation in different application background is also different. Furthermore, besides the geometry attributes, 3D model contains more complex spatial information and topological information than image data, which makes the segmentation problem of 3D model more challenging. Therefore, how to use human visual theory and cognitive theory to segment 3D models, and produce more accurate and reasonable segmentation results become an important research problem in graphics processing.In view of the segmentation results conform to human visual perception is a common problem in image segmentation and 3D model segmentation. In this thesis, we concentrate on the study of image and mesh segmentation from the perspective of human visual characteristics. The main results in this work are as follows:(1) We present a visual consistent adaptive image thresholding method (VCA thresholding method), which combines the thresholding process with the characteristics of human visual system. Firstly, the image pixels are roughly classified into two categories by utilizing the local gradient information of each pixel. After that two sub-images are constructed to retain the essential information of the original image. For each sub-image, a corresponding global optimal threshold is calculated by optimizing an objective function. The definition of the objective function is inspired from human visual characteristics. Finally, a visual consistent binary image is produced by combining the results from the previous steps with the local information of pixels. Our method has been tested on different images, the experimental results demonstrated that the overall visual qualities of our method are more suitable for human visual perception.(2) We present an algorithm which is guided by mesh seed points for segmenting a mesh into meaningful sub-meshes. Firstly, a candidate set composed of feature points is selected to highlight the most significant features of the model. The collection of seed points we defined is a subset of the candidate set. By maximizing the diversity measure between the seed set and candidate set, the appropriate seed point is selected. Because humans generally perceive desirable segmentations at concave regions, in order to partition the target into meaningful parts, we define a distance function between each pair of mesh vertices. This function is formed by arc length, angular distance and curvature-related correction term. Finally, by clustering mesh vertices to generate the visual meaningful segmentation results.(3) We present a new mesh segmentation method based on mesh saliency and spectral clustering. Our method transforms the spatial segmentation into the spectral clustering, and combines the mesh saliency with spectral clustering to achieve meaningful segmentation. Firstly, the mesh concave region is determined by using the "minimum rule" in the visual theory, and then according to the mesh saliency and curvature information to define a laplacian matrix. Through the eigen-decomposition of the matrix, the first k eigenvector of the laplacian matrix can be calculated. After that, we embed the original mesh into a k-dimensional spectral space. Finally, by utilizing the gaussian mixture model method to achieve the visually meaningful segmentation. The initial cluster centers is decided by mesh saliency. The experimental results have demonstrated the effectiveness of the proposed segmentation method. Especially for the model with convex regions and branch components, the method can achieve better visual quality.