Research On Theories And Methods Of Multi-label Classification Base On Hierarchy And Exclusion Graphs

In our daily life, the problem of classification is a hot topic. It refers to the identification of the labels and their categories. In recent years, a new popular supervised classification problem is multi-label classification problems.A number of remedies have been proposed. They are based on either on simple dimension reduction techniques or involve expensive optimization problems. These methods in present have some nice results, beside the calculation, the main issue is that they do not capture the complexity of semantic labels in the real life. When the number of labels grows to the hundreds or thousands, these labels will be hierarchical. Thus,there are all kinds of relationships between each labels, how to use these relations effectively in multi-label classification problems, is still a challenge. In recent years, a new article introduce a HEX(Hierarchy- Exclusive) graphs, which develops a new classification model that allows flexible encoding of relations based on prior knowledge. Aiming at this problem, we make some research, the main content are as follows:1.Base on the characteristics of labels in the clothes searching problem, we design and implement a methods for the multi-label classification problem, which invlove a hierarchical and mutually exclusive model and the convolutional neural network. We collect the clothing images from some online shopping websites and make an experiment to analysis the model. The experimental results show that, compare with other common multi-latbel classification models, HEX model benefits from the prior knowledge and has a better accuracy on the same training set.2.Base on a hierarchical exclusion model and the convolutional neural network, we design a clothing application for multi-label classification.