一种结合深度学习特征和社团划分的图像分割方法  

Image Segmentation Based on Deep Learning Features and Community Detection

作　　者：胥杏培 宋余庆[1] 陆虎[1] XU Xing-pei;SONG Yu-qing;LU Hu(School of Computer Science and Communication Engineering,Jiangsu University,Zhenjiang 212013,China)

机构地区：[1]江苏大学计算机科学与通信工程学院,江苏镇江212013

出　　处：《小型微型计算机系统》2018年第11期2533-2537,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61375122;61572239)资助;中国博士后科学基金项目(2014M551324)资助;江苏大学高级人才科研基金项目(14JDG040)资助

摘　　要：图像分割方法是一种非常重要的图像分析技术,现有常用图像分割方法都需要依靠人工提取特征来抽取图像的特征.本文提出了一种新的基于深度学习特征和社团划分方法结合的图像分割方法.基于深度学习特征抽取方法,采用卷积神经网络(CNN)模型抽取了图像的深度学习特征.首先,SLIC超像素算法将图像由像素级转化为区域级,划分成超像素区域,针对每个超像素区域,我们提取了深度学习特征.另外在结合超像素区域颜色特征的基础上,构建成了新的超像素区域相似度矩阵.然后我们基于社团划分的思想对相似度矩阵进行了划分.为了能自动识别图像分割的个数,我们使用了模块度Q自动确定最佳的社团个数,实现了图像的自适应分割.为了说明本文提出方法的有效性,我们在BSDS500数据集上进行了实验测试,并与现有的几种著名图像分割方法进行比较.在不同图像上的分割实验结果表明,我们提出的图像分割算法优于其它几种方法.Image segmentation is an important technique in image analysis.Existing methods of image segmentation rely on artificial neural networks to extract the actual features of the image.In this paper,we present a new image segmentation method based on deep learning features and community detection.We propose the use of a pre-trained convolution neural network （CNN） to extract deep learning features of the image.The deep CNN is trained on ImageNet dataset and transferred to image segmentation for constructing potentials of super-pixels.We first convert the original image from the pixel level to the regional level using the simple linear iterative clustering （SLIC） algorithm which divides the original image into superpixel regions.For each superpixel region,we extract deep learning features.In addition,a new super pixel similarity matrix is constructed based on the color feature of the superpixel regions.We finally perform community detection based on the similarity matrix obtained.Modularity Q is used in order to automatically identify the number of image segments which corresponds to the best number of communities,and realize the adaptive segmentation of the input image.In order to illustrate the effectiveness of the proposed method,the later was tested on the BSDS500 dataset and compared with several well-known image segmentation methods.Segmentation experiments conducted on different images showed that our proposed image segmentation algorithm outperforms other methods.

关 键 词：卷积神经网络(CNN) SLIC超像素 谱聚类 模块度Q 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]