基于知识域与数据域协同的图像压缩算法研究

Most recently, due to the rapid development of smart phones such as iphone and the popularity of cutting-edge Internet services such as Weibo and Youku, the multimedia data delivered over wireless networks have become increasingly huge. Therefore, the demand for image compression is becoming urgent in wireless multimedia communications. Traditional methods approach the image compression in data domain via the transform with a set of orthogonal bases that have been predefined, thereby requiring the large storage space for data and making against the fast delivery for image/video. In knowledge domain (i.e. the sets of knowledge generalized from natural objects), the image understanding methods target to extract the features and class labels of objects in an image, and they may be applied to image compression, as another way, in order to improve the efficiency of image compression. However, currently the research on this topic is still in its infancy. .For improving the compression efficiency of images, in light of human logic and perception mechanisms, we shall focus on the fundamental problems of image compression on the basis of the previous research results of this project. More specifically, (1) in knowledge domain we shall deal with the image understanding problem by combining the statistical learning and logic-rule learning together; (2) in data domain, we shall explore a learning method to obtain the over-complete texture dictionaries of different object classes and a computational approach for sparse representation with respect to the learnt dictionary; (3) towards image compression, we shall develop a unified collaborative framework to integrate the data and knowledge domains. In a word, this project aims at providing the new theoretical and technical foundation for the next generation image compression technology.

近年来，随着苹果等智能手机的发展以及无处不在的微博、在线视频等各类新型网络业务的普及，无线通信需传输的多媒体数据量变得日益庞大。因此，图像/视频压缩在无线多媒体通信中日趋重要。传统图像压缩算法大多基于数据域，通过预定义的正交基变换实现，占大量存储空间，不利于图像/视频高效传输。另一种新的研究思路为：在知识域上，即利用事物归纳后的知识集合，通过图像理解标识图像中物体的类别，进而应用于图像压缩，从而大幅提高图像压缩效率；该研究在国际上尚处于起步阶段。本项目将利用已有研究基础，以提高图像压缩率为目标，研究内容为：（1）借鉴人类逻辑思维与图像认知机理，在知识域上研究统计学习与逻辑法则学习相结合的图像理解算法；（2）在数据域上研究各类物体纹理字典的机器学习算法及其稀疏表示的计算模型；（3）融合数学模型与图像认知，构建数据域与知识域协同的图像压缩体系框架。本项目将为图像压缩提供新的理论依据与技术支持。