Content Based Image Retrieval And Image Semantics Analysis

This thesis researches some methods surrounding the semantic gap in content-based image retrieval(CBIR),such as relevance feedback,semantic segmentation and semantic analysis.For the on-line CBIR,we propose a retrieval method based on a new active learning Support Vector Machines(SVM).For the off-line CBIR,we propose a classification method based on object semantic template,a segmentation method based on graph cut for multi-label images,and a retrieval method based on region semantic bag of words.For the problems of relevance feedback,such as small size of samples,the asymmetric of positive and negative samples,imbalance between labeled samples and un-labeled samples,we proposed a new relevance feedback strategy for CBIR.Firstly,it optimizes the traditional active learning SVM.It selects the negative samples to active learn based on feature similarity and the hyperplane of SVM,which avoids the randomness of the sample selecting strategy based on uncertainty.In the process of active learning,we design an adaptive regularization rules,which proposes the generalization ability of retrieval system.The experimental results show the proposed method can improve the accuracy,the efficiency and the generalization of retrieval system.The accuracy of the global features to express the image semantics is low in CBIR.For this problem,we research the local features,and propose a method to acquire the region of interest(ROI).It segments ROI based on Gaussian Mixture Model(GMM)and Conditional random field(CRF),and use Multi-layer perceptron(MLP)to learn the semantic template for object semantic.Then use the object semantic template to classify the images.The experimental results show the proposed method has high classification accuracy for the images with explicit object and background.For the computational complexity of multi-label semantic segmentation based on graph cut,we propose an auto segmentation method.It uses the main color as label seeds without prior knowledge,the pixel seeds or region seeds.It designs two parameter to get the main colors,and uses ?-expansion move algorithm to reduce the computational complexity of graph cut.The experimental results show the proposed method has fast speed and high segmentation accuracy.In the image classification field,visual bag of words(BoW)has two drawbacks.One is the lower classification accuracy because visual BoW is commonly extracted from local low-level visual feature vectors via key points,without considering the high-level semantics of the image.The other is that time consuming is long,because the size of the vocabulary is very large.To solve these two problems,we propose a new image classification model based on visual bag of semantic words(BoSW),which is based on the obtained main semantic regions,and uses a semantic annotation algorithm based on support vector machine to label the corresponding region with a visual semantic vocabulary.The proposed BoSW model refines the image semantic via introducing the user's conception for extracting semantic vocabularies,and reduces the size of the vocabulary.The experimental results show the superiority of the proposed algorithm by comparison with the state-of-the-art methods on benchmark datasets.