Abstraction For Static And Dynamic Images Based On Visual Saliency

Non-Photorealistic Rendering (NPR) is an important field of computer graphics, which aims at representing certain artistic characteristics, simulating artistic works. NPR in general involves abstraction and stylization of the target scene, especially abstraction rendering technology, which helps simplify the visual cues and weaken more information than necessary, and convey certain aspects of the scene more effectively. Thus, the abstraction rendering provides a means of simulating the visions of artists and designers to observe and describe the reality.First of all, the paper uses a tree structure to introduce the research status and developing trend of the technologies, take the computer graphics for the root node, go down to the existing NPR technologies in succession, and then discuss visual saliency based abstraction technologies. Meanwhile, the knowledge of cognitive psychology is combined to give some influential methods of visual saliency detection. The paper describes a complete abstraction framework for images and videos, and the results in the experiment show that our algorithms are more effective and precise compared with recent methods.The main contributions of the dissertation are as follows:1. This paper proposes an automatic salient regions detection algorithm, and it is a Local Spatial Neighbors (LSN) based saliency computation. Based on three characteristics of defining salient object, we extract full-resolution saliency maps of interesting targets. And to capture complete salient information, we compute saliency at region level. The paper first segment the input image into regions, then define the saliency for each region as the weighted sum of the region’s color contrasts to spatial neighbors associated with this region. The weight is the ratio between the area of each neighbor and total area of the neighbors. We further incorporate spatial information to define the salient part which is placed near the center of the image by introducing the Gaussian falloff function to define spatial weight.2. We extend single-scale region based saliency computation to multi-scale region based saliency computation in order to make our algorithm more robust under complicated circumstances, which is obtained by setting several groups of different parameters using a graph-based image segmentation method. Starting from the first group, we compute single scale saliency, and the process is repeated until the last group is reached, and then we compute the pixel saliency value with spatial weight ratio of the color distance based on multiple scales. Experiments indicate the proposed schemes to be superior in terms of both precision and recall, while still being simple and efficient.3. The paper improved anisotropic Kuwahara filter by redefining the weighted local averages and the squared standard deviations, to make the results satisfy the expected purpose. On this basis, we propose a saliency-driven composition that integrates multiple scales to implement rendering styles. A saliency map, either obtained from automatic saliency computation or designed by a user, is introduced both in the final composition and in the creation of the multiple styles, namely a meaningful rendering style, which is generated by integrating styled lines and retaining the salient part as realistic.