用改进的Paik型Boltzmann机实现图像复原  

Realization of image restoration by improved Paik's Boltzmann machine

作　　者：张煜东[1] 吴乐南[1] 

机构地区：[1]东南大学信息科学与工程学院,江苏南京210096

出　　处：《光学精密工程》2009年第7期1737-1744,共8页Optics and Precision Engineering

基　　金：国家863高技术研究发展计划资助项目(No.2008AA01Z227);国家自然科学基金资助项目(No.60872075);高等学校科技创新工程重大项目培育基金资助项目(No.706028);江苏省自然科学基金资助项目(No.BK2007103)

摘　　要：为解决传统的Boltzmann机方法不仅容易陷入局部最小点,而且收敛速度慢问题,对传统的Boltzmann机进行了改进。将Paik′s算法与Boltzmann机结合,使串行模式推广到并行模式以加快收敛速度;使用亚单位步长增进技术增加计算精度;最后,为折中收敛速度与收敛精度这一对矛盾,采用了自适应步长策略。对算法的改进进行了理论验证、收敛性分析并对残差变化进行了讨论。实验表明,该方法能够收敛到全局最优,复原结果的峰值信噪比比改进的Boltzmann机法获得的峰值信噪比高0.5～0.8dB,且收敛速度仅为该方法的1/3,证明了本文提出的改进的Paik型Boltzmann机对图像复原是有效的。An improved Paik/s Boltzmann machine was presented to optimize the traditional Boltzmann method that suffers from not only being trapped into a local extreme but also a slow convergence. By proposed the method,the Paik＇s algorithm is integrated into the Boltzmann machine to restored images. Then, serial models are extended to parallel models to fasten the convergence speed and a subunit step is adopted to increase calculation precision. Finally a step adaptive method is used to trade off the contradiction of convergence precision and convergence velocity. Moreover, each improvement for the origonal algorithm is expatiated by theoritical validation, convergence analysis, and discussions of residual variations. Experiments demonstrate that the proposed method can be converged to the global minimum, and the Peak Signal Noise Ratio （PSNR） of the restored image is 0.5-0.8 dB higher than that of the improved Boltzmann method proposed by Bellouquid, while the consumed time is only 1/3 that of the latter. These results show that the proposed method is effective and valid for image restorations.

关 键 词：图像复原 BOLTZMANN机 神经网络 全并行算法 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]