基于自组织神经网络SOM和K-means聚类算法的图像修复  

Image Inpainting Method Based on Self-organizing Maps and K-means Clustering

作　　者：孙震[1] 王兆霞[1] 白明[1] 张俊生[1] 

机构地区：[1]天津理工大学计算机与通信工程学院,天津300384

出　　处：《科学技术与工程》2012年第8期1790-1794,共5页Science Technology and Engineering

摘　　要：近来自然图像的修复已经成了一个热门话题。提出了一种基于K-means聚类算法的自组织神经网络(SOM),称为SOM-K。它首先利用SOM来训练每一个像素的特征向量,并把一幅图像分层。这样就能把每个破损像素分到每层,同时SOM训练后的输出也通过K-means聚类算法来聚合,分别在各个层中修复破损的像素。最后把修复好的各层溶合到一起。与单独使用SOM相比,SOM-K具有更精确的分类能力。Natural image inpainting has been a hot topic in recent year. A SOM based K-means （SOM-K） method for inpaintingis presented. Feature vectors of each pixel are first trained by a SOM neural network for divid- ing an image into several layers, and assign each damaged pixel to one layer, then the output of SOM are clustered by K-means clustering method, restoring these damaged pixels by the information of their respective layer. At last,these inpainted layers are fused together. Compared to SOM, SOM-K makes a more precise segmentation in most cases by dividing an image into several layers.

关 键 词：图像修复 自组织神经网络 K-MEANS聚类算法 

分 类 号：TG391.41[金属学及工艺—金属压力加工]