基于显著性分割的图像抽象化

摘要

One of the primary task of the image abstraction is to enhance the visual information of specific objects while removing some unimportant information. In order to further improve the effect of the cartoon stylization, this paper proposes a non-uniform image abstraction algorithm based on the saliency segmentation. The method effectively enhances the visual characteristics of salient foreground object regions,and discards the details in the background. It employs the frequency-tuned algorithm to obtain the saliency maps and the Mean-Shift segmentation approach to detect the whole foreground areas. Using this map and bilateral filter can obtain the abstracted map based on the edge detection of flow-guided. It uses soft quantization to improve the further sense of hierarchy,and abstracts images by taking advantage of the segmentation results with a non-uniform way. Experimental results show that the algorithm can effectively produce foreground-enhancing non-photorealistic images,and outstanding the foreground of images.%图像抽象化的目的是增加具有特定目标的信息，并过滤掉一些不相关或不重要的信息。为进一步提高图像的卡通风格化效果，提出一种基于显著性分割的图像抽象化算法，增强前景物体区域的可视特征，同时去除背景区域的无关细节。采用频率调谐算法获得显著度图，利用Mean-Shift分割算法检测出整个前景区域，与双边滤波和基于流场的边缘检测结合获得抽象化图形，通过软量化方法使产生的风格化图像更具有层次感，并以此为基础对图像进行非均匀的抽象化处理。实验结果表明，该算法能够有效地产生前景特征增强的非真实感图像，使结果图像的前景主题更加突出。