Research On Key Technology Of Image Marking Based On Multi - Tag Learning

With the rapid development of computer technologies and mobile photo technologies, the uploaded images in the network space grow explosively. In order to use the image information well, a large number of image retrieval algorithms have been proposed. All these algorithms can be generally categorized as three types: Text-based Image Retrieval, Content-based Image Retrieval, and Semantic-based Image Retrieval. The core technology of semantic-based image retrieval is automatic image annotation and it is the key research contents of this thesis.A lot of algorithms of automatic image annotation have been proposed, but the following three problems have not been resolved fundamentally: the semantic gap problem, the curse of dimensionality, and the data imbalance problem.In order to solve the above issues, this thesis updates four classical algorithms of machine learning based on multi-label learning:1. The algorithm ML-KNN builds the learning model by using the correlation of the major samples and the number of neighbors in different classes. But the correlation of the distances between the major samples and the neighbors is not considered. In this thesis, a new algorithm ML-WKNN based on lazy learning is proposed which builds a model by combining the above two correlations. The experimental results of ML-WKNN on Image and Yeast data sets show that it outperforms the other four classical multi-label algorithms.2. The algorithm MLNB utilizes PCA to preprocess the attribute characteristics of samples. The processed attribute characteristics are uncorrelated, but they still can’t meet the conditions of Naive Bayes classification algorithm. In this thesis, a new algorithm IMLNB is proposed. It preprocesses the attribute characteristics of samples by ICA. The processed attribute characteristics can meet the precondition of Naive Bayes algorithm. The experimental results of IMLNB on Image and Yeast data sets show that it outperforms the other four classical multi-label algorithms.3. A new algorithm LTFML based on the idea of constructing label-specific features is proposed. The algorithm LTFML only uses the positive samples of each label to build label-specific features. At the same time, a new weighting function is proposed and the different clusters are weighted differently. The experimental results of LTFML on Image and Yeast data sets show that the LTFML can give more comprehensive predictions than the other four classical multi-label algorithms.4. To address the imbalance data problem of multi-label learning, a new algorithm 3BM based on the idea of Bagging is proposed. It utilizes Bagging approach to select the same number of positive samples and negative ones of each class with replacement from the training set. The chosen samples consist of completely balanced training subsets. In this thesis, the strategy min-max modular is used to combine the predicted results of each base model. The experimental results of 3BM on Image and Yeast data sets show that it can give more comprehensive predictions than the other four classical multi-label algorithms.