(zz)电信学院 图像所引智计划系列讲座(第二十三讲)  

Title: Image estimation and component separation with numerical Bayesian techniques

Ercan E Kuruoglu Institute of Information Science and Tecnologies (ISTI) - National Research Council of Italy (CNR)


时间: 10月21日下午2:00

In this talk, we investigate the problem of source separation in two dimensions which has been motivated by our work on separation of independent components in astrophysical images. Albeit the formulation is principled, the analytical solution of Maximum-a-Posteriori (MAP) estimate may be intractable. In order to obtain the joint maximization of the a posteriori distribution of the unknown parameters, in the past approximate solution methods have been proposed. In this seminar, we present a general, flexible and unifying method for this avoiding unrealistic approximations. We adopt a fully Bayesian framework which allows us to introduce our prior information to the problem. Due to the complexity of the problem which does not lend itself to analytical solutions, we use Gibbs sampling which utilizes the conditional densities of the variables. These densities are determined by modelling images as a Markov Random Fields (MRFs). This is a key novelty of our work in that most of the studies in source separation literature have not made use of the time (or space) structure in signals and have instead considered them only as data sets. We present our results on sythetic texture images and astrophysical images which demonstrate the success and the flexibility of the technique.


评论: 0 | 引用: 0 | 查看次数: 6692 | 返回顶部
昵 称: 
验证码: =5+4(答案错误内容将被重置)
内 容:
选 项:
字数限制 1000 字 | UBB代码 开启 | [img]标签 开启