The Automantic Image Annotation And Application Research Based On CCA Subspace

Automantic image annotaton is a very important work in the image annotation, specifically in the image retrieval field. Since the birth of AIA technology, the study has not been stopped. The question of how to obtain a valid image feature that describes image semantic information, how to build an annotation model that describes iamge semantic accurately, and how to refine the candicate label word. This is quite important to solve the "semantic gap" problem in image annotation. The construction of the valid annotation model or method, plays a great role to improve the performance and efficiency of AIA.(1) Firstly, this thesis proposes a new technology which combines image color and texture based on Canonical correlation analysis (CCA). The first thing this method should consider is feature extraction. By comparing the advantages and disadvantages of the exiting feature information, we select color and LBP texture operator as the feature parameters to integrate. By virtue of CCA in this field, we obtain excellent information that express image visual semantic. This method sloves the problem that image feature could not reflect effectively the semantic parameters of image. Have a very great role in promoting the improvement the performance of image annotation and retrieval. The below image annotation is also based on this feature, so the image feature used in the thesis in AIA is acquired according to the mean. We demonstrate the good performance of this kind of feature.(2) In order to raise the accuracy of antumantic image ananotation, we proposed an AIA algorithm based on CCA and Gaussian mixture model (GMM) using CCAtechnology and in combination with GMM model. This algorithm processes the chosen two global image feature using CCA, make up for the lack of sigle feature in the work of describe image semantic. To avoid the unclear semantic express error brought by the unperfect image segmantation. It annotates image with the new feature. Then estimate the Joint probability density between image feature and annotation words using GMM, so obtained the probability density distuibution, build annotation model. We verify the performance in the Coral5k image set, shows that the accuration has been improved using this method for image annotation. (3) According to the relationship between image feature and tags, fusion the two features using CCA, then form an annotation and refine strategy combining with the CCA. This method uses the local feature of an image, and also considers the relationship between low-level image features and annotation words, we find the kind of relation using CCA, obtain the useful information between the two features finally, that is CCA subspace feature. For the consistency of image feature between training set and testing set, projecting these local parameter information of the image to be taged to the CCA feature subspace of training set throught CCA projecting variables. We construct another annotation model using the CCA features, combining with GMM model and the bayesian classifier. In the same time, for the candicate words, we execute annotation refination according to the similar relationship between keywords. For this method it builds a very useful "brige" between low-level image feature and high-level image semantic. Get closer to the real semantic exiting in image. To test the effectiveness of the proposed method, we demonstrate it in the JMLR2003dataset. Precision, recall and F score are selected to measure and analyse the performance of every algorithm. The results show that this algorithm has significantly improved the precision and recall.(4) For the later of introduction of Graph spectral theory to the AIA, we can improve the annotation performance based on graph theory. For this, we propose an antumantic image annotation algorithm based on K-harmonic average spectral clustering algorithm (KHMSC). We join the harmony mean concept on the basis of the K-mean, then combineing the knowledge of spectral clustering, the KHMSC algorithm is completed. The chief thing is that we obtain the area/semantic blocks throught Quadratic clustering. Clustering in the keyword vector space using KHMSC firstly, n semantic concept classes are formed, then executeing the secondary clustering in the image feature space in each class, then the similar feature will be clutered into the same space, in the end, semantic blocks would be obtained. We determine the best number of semantic blocks using Davies-Bouldin index. The feature used in this method is also extracted throught CCA method. The probability between semantic blocks and keywords is the statistical variables obtained in the approach. We use the multiple Bernoulli’s model to estimate the distribution information. At last, obtain the joint probability distribution (PDF) between annotation words and testing images using Naive Bayes model finally, then build the annotation model. Select the R biggest PDF tags as the final image annotation words. In the same way, we verity the performance of this algorithm based on JMLR2003image set. The experimental results show that the method improves the precision of image tagging greatly. It also turned out that this method is very useful to construct the consistency between low-level visual feature and high-level semantic.Canonical correlation analysis is of great importance to the relationship of the different visual features. Based on this property, we extract the accurate image feature to depict the image visual parameters. On this basis, annotate image rely on different image annotation algorithms. This thesis improves the performance of image annotation form three aspects:extraction of image feature、build valid annotation model、refine the obtained image tags. And at last, this thesis achieves a good result.