可进化可重构图像滤波器的设计  

An Evolvable and Reconfigurable Image Filter

作　　者：刘少腾[1] 来金梅[1] 陈利光[1] 童家榕[1] 鲍丽春 

机构地区：[1]复旦大学专用集成电路与系统国家重点实验室,上海201203 [2]北京航天飞控中心,北京100094

出　　处：《复旦学报（自然科学版）》2010年第6期709-715,共7页Journal of Fudan University：Natural Science

基　　金：国家高技术研究发展计划(863计划)(2007AA01Z285)资助项目

摘　　要：设计了一种新型的可重构、可进化图像滤波器.可重构电路采用了一种全新的可配置电路网络结构.该结构将功能运算单元以Benes拓扑结构互联,取代现有的MUX加功能运算单元的结构,使其具有更多条输出路径,更多运算单元参与到了进化中以提高进化效率.在运算单元设计方面,利用FPGA中特有的丰富LUT逻辑资源优势,设计了一种仅通过LUT配置码来控制其功能的P运算单元,节约资源且扩充了运算单元可执行的功能选项.借助于模拟退火遗传算法,通过对可重构电路配置码(基因)的不断改进和优化,该滤波器可以很好地滤除图像中的高斯噪声和椒盐噪声.实验结果表明:经过100万代进化,对于高斯噪声,噪声图像的平均每像素误差Mdpp=32.11,滤波后Mdpp=16.74,滤波质量优于一般高斯滤波器和现有的MUX加功能运算单元的可进化滤波器;对于椒盐噪声,噪声图像Mdpp=6.22,滤波后Mdpp=3.44,有效滤除了噪声.A new reconfigurable and evolvable image filter is described.A configurable circuit network（CCN） is used as reconfigurable part of this evolvable filter.Instead of traditional MUX and function unit structure,this CCN structure utilizes Benes network topology to connect function units,in order to enable more ways to output and more function units to be involved in evaluation,resulting in the efficiency enhancement of evolution.Meanwhile,for function unit design,because of the abundant LUT resources in FPGA,class P unit is created,whose function is decided directly by LUTs＇ content.Therefore,resources are saved;the function of function units is extended.By using the analog annealing genetic algorithm to improve the configuration codes repeatedly,a filter that can significantly decrease Gaussian noise or salt-pepper noise in an image is finally realized.Experimental results show that,after 1 million generation＇s evolution,for Gaussian noise,the noisy image＇s Mdpp=32.11 and filtered image＇s Mdpp=16.74.It is better than the result of ordinary Gaussian filter;while for salt-pepper noise,the noisy image＇s Mdpp is 6.22 and filtered image＇s Mdpp=3.44.It proves that the noise has been significantly reduced.

关 键 词：可进化硬件 图像滤波器 可重构电路 Benes拓扑结构 

分 类 号：TN402[电子电信—微电子学与固体电子学]