多特征融合图像检索方法及其应用研究

发布时间：2018-05-14 12:06

本文选题：图像检索 + 多特征融合　；参考：《哈尔滨理工大学》2016年博士论文

【摘要】：从图像检索方法产生至今,已经形成了三个重要的分支:基于文本的图像检索、基于内容的图像检索、基于语义的图像检索。基于文本的图像检索,用文本来描述用户的需求,如图像名、图像特征等,但因为文本表达能力的局限以及文本标注过程中的歧义性,往往造成检索结果和用户需求的不符合;基于语义的图像检索,在图像视觉特征基础上进一步提炼其高级的语义表达能力,但检索过程复杂且方法体系发展尚不完善;基于内容的图像检索,以颜色、纹理、形状等作为图像的特征表达,以此为依据开展检索并作为相似性判断的依据,是当前图像检索领域的研究热点。基于内容的图像检索方法,为提升查准率和查全率需要同时使用多种特征。这些特征的融合策略和融合规则,是制约基于内容图像的检索方法性能的关键问题。本文以基于内容的图像检索方法为研究对象,以多特征融合框架设计、融合规则设计为核心研究内容,以检测领域中的应用为研究背景,开展了如下的研究工作:第一,在单一特征的图像检索方法框架基础上,研究多特征融合的图像检索理论框架,深入分析多特征融合过程中的特征选择、特征提取、特征向量生成、相似性比较等问题。从特征向量融合和相似性测度融合两个角度,构建多特征融合图像检索理论的通用框架。依托层次分析法,构建图像检索方法的评价框架。第二,综合考虑查准率和检索时间,提出了基于颜色特征和边缘特征融合的图像检索方法。此方法在相似性测度中,将红色、绿色、蓝色三个颜色分量特征和边缘细节特征融合在一起,以此提升检索结果的准确性。同时,为了降低检索过程的执行时间,颜色特征和边缘特征的提取都放在二次小波分解的低频分量图像上。第三,综合考虑查全率和检索时间,提出了基于小波基特征的图像检索方法。此方法融合了两类抽象的小波基特征,作为图像检索过程中的相似性判断依据,充分发挥了小波基特征的自适应能力,使得整个方法对不同类别的图像都具有较强的针对性。同时,小波基特征通过泰勒级数近似替代,也具备较好的速度优势。第四,综合考虑查准率和查全率,提出了多特征ESN融合的图像检索方法。此方法融合了颜色特征、纹理特征、形状特征,并用ESN网络训练得出三种特征的融合权重,不仅增强了相似性判断的全面性也增强了检索过程的适应性。
[Abstract]:Since the emergence of image retrieval methods, three important branches have been formed: text-based image retrieval, content-based image retrieval, and semantic-based image retrieval. In text-based image retrieval, the user's needs are described by text, such as image name, image feature, etc. However, because of the limitation of text expression ability and the ambiguity in the process of text annotation, the result of retrieval is often inconsistent with the user's demand. Semantic image retrieval further refines its advanced semantic expression ability on the basis of image visual features, but the retrieval process is complex and the method system is not perfect, and content-based image retrieval is based on color and texture. Shape, as the feature expression of image, is the research hotspot in the field of image retrieval, based on which the retrieval is carried out and the similarity is judged. Content-Based Image Retrieval (CBIR) requires multiple features to improve precision and recall. The fusion strategy and fusion rules of these features are the key problems that restrict the performance of content-based image retrieval methods. In this paper, the content-based image retrieval method is taken as the research object, the multi-feature fusion framework design, the fusion rules design as the core research content, and the application in the detection field as the research background, the following research work is carried out: first, Based on the framework of a single feature image retrieval method, the theoretical framework of multi-feature fusion image retrieval is studied, and the problems of feature selection, feature extraction, feature vector generation and similarity comparison in the process of multi-feature fusion are analyzed in depth. Based on feature vector fusion and similarity measure fusion, a general framework of multi-feature fusion image retrieval theory is constructed. Based on Analytic hierarchy process (AHP), the evaluation framework of image retrieval method is constructed. Secondly, an image retrieval method based on color feature and edge feature fusion is proposed. In the similarity measure, the red, green and blue color component features and edge detail features are fused together to improve the accuracy of retrieval results. At the same time, in order to reduce the execution time of the retrieval process, the extraction of color features and edge features are placed on the low-frequency image of quadratic wavelet decomposition. Thirdly, considering recall rate and retrieval time, an image retrieval method based on wavelet basis feature is proposed. This method combines two kinds of abstract wavelet basis features, which can be used as the basis of similarity judgment in the process of image retrieval, and gives full play to the adaptive ability of wavelet basis features, which makes the whole method have strong pertinence for different kinds of images. At the same time, the wavelet basis feature is replaced by Taylor series approximation, which also has a good speed advantage. Fourthly, an image retrieval method based on multi-feature ESN fusion is proposed considering recall ratio and recall rate. This method combines color features, texture features and shape features. The fusion weights of the three features are obtained by ESN network training, which not only enhance the comprehensiveness of similarity judgment, but also enhance the adaptability of retrieval process.
【学位授予单位】：哈尔滨理工大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】