FoE模型的训练方法研究

A Study on Training Methods of FoE Model

FoE模型是Roth等于2005年提出的一种新的马尔可夫随机场模型,其以优异的处理效果引起学界的瞩目。介绍了FoE模型及其训练的过程,给出了应用student-t专家和charbonnier专家分别在几种不同训练设置下训练得到的FoE模型进行图像去噪的试验结果。并针对FoE模型训练时间较长的问题,提出了一种改进的批训练方法应用于FoE模型的训练。这种方法通过逐步增大训练过程中所应用的批数量的方式,将FoE模型的训练过程分为4个层次。图像去噪仿真试验的结果表明,该方法可以在有效减少约50%FoE模型训练时间的同时,取得与原训练方法相似的训练结果。

FoE model is a new MRF model demonstrated by Roth et al in 2005.This model has attracted a lot of attention by its excellent performance.In this paper,described FoE model and its training progress,present image denoise experiment results with student-t expert and charbonnier expert under several training settings.To solve the time consuming training problem of FoE model,an improved batch training method is proposed.Through increasing batch number step by step,this mothed split the training progress of FoE model into 4 stages.Practical example of image denoise shows that the proposed method can effectively reduce approximate 50% training time of FoE model at the same acquire similar training result.