针对实时视觉通信的图像序列自动提炼

Automatic Abstraction of Image Sequences for Real-Time Visual Communication

本文提出了一项新技术,它可以实现从任意的图像序列中自动提炼出简洁的表达方式,以便进行高效的视觉通信.我们认为,视觉通信的全过程可分为视频数据的传输和人眼对视觉信号的理解两个阶段.因此,本文以心理学中人对图像的认知规律的相关理论为指导,专注于研究如何同时提高图像的可压缩性和可理解性.我们借助一个缘提取算法来保留对人的视觉系统最为敏感的物体边界,再用一个非线性扩散算法来减弱无足轻重的细节信号.为了使最终生成的动画保持时间上的一致性,本文的技术方案是在整个时空域上设计的.而我们依然能够保持实时的处理速度,因为该方法可以方便地使用GPU作并行计算.为了演示新技术的实用性,我们还建立了一个以本文算法作为处理内核的完整的视觉通信系统并在该系统进行所有实验.统计数据表明,本文方法不仅可以明显地降低传输带宽,而且提高了图像序列的可理解性.

We present a novel technique for efficient visual communication, where a compact representation is abstracted from arbitrary image sequences automatically. We aim at improving the entire visual communication process, including transmission of video data and perception of visual signal by human eyes. Therefore,guided by some related theory in Psychology,our work increases both the compressibility and comprehensibility of images. This goal is achieved by two steps. One step is an edge extraction process for preserving object boundaries which are believed to be the most sensitive features to human vision,and the other step is a non-linear diffusion process for reducing extraneous details. Both are designed in a spatiotemporal manner to guarantee the temporal coherence of resulting animations, while real-time processing speed is maintained by facilitating parallel computation on a GPU. We additionally build a complete visual communication system with the proposed algorithm as its core to demonstrate the practicality of our technique. Experimental statistics collected on the system indicates that transmission bandwidth is obviously saved while perceptibility of image sequences is improved.