【作者】 张志远；

【导师】 郭捷；

【作者基本信息】 上海交通大学 ， 计算机应用技术， 2011， 硕士

【摘要】 近年来,数字图像目标识别技术在人脸检测、车牌识别、网络图像索引和智能机器人视觉模拟等诸多领域得到了广泛应用。早期图像目标识别技术以图像的颜色、纹理和边角等局部视觉特征为主,缺乏整体性的考虑。近期的研究则注重加入图像内物体轮廓以及位置等全局特征,以提高目标识别的准确率。本文在研究图像局部特征和全局特征提取的基础上,建立了一个基于多特征融合的多目标识别分类模型。本文根据马尔可夫随机场理论,建立的多目标分类模型融合了图像的局部视觉特征和全局位置等相关特征。模型所采用的局部视觉特征包括图像的颜色、纹理、边界和边角等17维特征。模型所采用的全局位置相关特征表征了图像内物体与物体之间在空间布局上的位置相关关系。首先,我们提出了一个图像角度特征提取算法,该角度特征用于描述图像内单个物体各个部位在空间布局上的位置关系。该算法经拓展,可用来提取图像的全局位置相关特征。基于马尔可夫随机场模型的多目标识别系统的建立包括以下四个步骤:基于聚类的图像预分割处理、图像局部特征提取及局部分类模型的训练、图像全局位置相关特征提取和综合识别模型学习。经训练获得的图像多目标识别分类器可用于后续测试图像的目标识别与分类。通过在不同图像数据库上的测试分析,结果表明融合了图像局部视觉特征和全局位置相关特征所建立的综合马尔可夫随机场分类模型在处理多目标分类问题上表现出了较大的优越性,与仅基于图像局部视觉特征的分类模型相比,本文系统大大提高了目标识别的准确率。更多还原

【Abstract】 In recent years, digital image object recognition has been applied in various as-pects of life, touching areas such as camera face detection, illegal road vehicle licenseplate recognition, image indexing and intelligent robot vision. Early object recognitionresearch mainly focused on the extraction of local features like color, texture, edge andcorner, which lacks the sense of global understanding. Recent research focus has putmore emphasize on global features such as contour and positions in order to improveobject detection precision. In this paper, we focus on extracting both local and globalfeatures and establish a multi-object classi?cation model by integrating these features.Based on the Markov Random Field theory, we establish a multi-object recogni-tion model. This model encodes local features and global position features. For localfeatures, we extract information from color, texture, edge and corner to form a 17 di-mensional feature vector. For global features, this paper ?rst proposes an angle featureextraction method that uses angle feature to describe various parts of a single object inthespatialdistributionofmutualrelations. Then, byextendingthismethod, wecapturethe spatial layout information of di?erent objects within an image.The Markov Random Field based multi-class recognition system is established bythe following four steps: segmenting image into superpixels,extracting image’localvisual features and training the basic classi?er based on these local features, extractingthe global spacial layout features from images, and training the synthesized multi-classrecognition model. The learnt model can be used to classify the objects in the testimages.Weevaluateourmodelontwodifferentimagedatabases. TheresultsshowthattheMarkovrandomfieldincorporatingspaciallayoutinformationallowsustosigni?cantly increase the classification accuracy compared with the models based on local features.更多还原