Localized Generalization Error Model Based Semi-supervised Learning For Image Retrieval

Localized Generalization Error Model (L-GEM) is a new approach to estimate generalization error of a classifier. Prior researches have indicated that L-GEM out-performs current methods in neural network architecture selection and input feature selection. In this thesis, we apply the L-GEM to sample selection and proposed a semi-supervised learning method based on L-GEM. These proposed methods are adopted in a Content-Based Image Retrieval (CBIR).In CBIR, the number of unlabeled images is large while there are only a few labeled samples. Making a good use of these unlabeled samples is important to improve the retrieval precision. This has become a hot research topic in recent years. In contrast to fully labeled training dataset in supervised learning, semi-supervised learning method learns from both a small portion of labeled samples and a large portion of unlabeled samples. Co-training constructs two basic learners using two independent and sufficient sub-feature sets. Each trained basic learner classifies unlabeled samples and labels the most confident one which will be then added to the other learner's training set. This process repeats until satisfactory result is achieved. In this way, the training set is expanded and the problem of limited training samples is releaved.In the co-training process, sample selection is a vital step. Traditional confidence criterion base on posterior probabilities or approximately directly. In this thesis, we proposed the L-GEM based sample selection criterion. In co-training process, we use sample's L-GEM value as the confidence criterion and the confidence is inversely proportional to their L-GEM values. In our experiments, we use Radial Basis Function Neural Networks as basic learner and the results show that our proposed measurement is better than the traditional methods. The proposed method greatly improves image retrieval performance.Moreover, in order to reduce the number of relevance feedback iterations, we adopt active learning based on L-GEM. In detail, a few images with low confidence will be feeded-back to user for labeling. The experiment shows that it is more effective than random method.