Visual perception based image and video abstraction
视觉感知的图像和视频抽象

Reliable estimation of visual perception reflects the understanding of the meaningful structure in an image. In this paper, we describe a complete abstraction framework for images and videos that explicitly responds to this goal. The method stylizes salient images or video parts and non-salient images or video parts differently by an edge preserving filter, and keeps a harmonious transition between the two parts. First, we introduce an automatic salient object segmentation algorithm to distinguish salient regions, and it is a saliency computation based local spatial neighbors. Taking into account the actual needs, we provide an interactive technology, which can be convenient for specifying image salient structure information on purpose. Based on the generated salient structure information mask, we use a single-scale anisotropic filter to process the salient parts, and use a multi-scale anisotropic filter to process the non-salient parts so that we can implement a strong abstraction effect.The proposed method generates a kind of image and video abstraction that does not only represent a preferable visual effect, but also can be applied to the implementation of another non-photorealistic rendering (NPR)result after soft image quantization. Experiments show that our algorithm could get the desired result for processing a certain number of images and videos.