PROJECT TITLE :

Dynamic Double Classifiers Approximation for Cross-Domain Recognition

ABSTRACT:

In general, the existing cross-domain recognition methods primarily concentrate on altering the feature representation of the data or modifying the classifier parameter, and the effectiveness of these methods is indicated by the better performance. Nevertheless, the majority of the currently available methods do not simultaneously integrate them into a unified optimization objective in order to further improve the learning effectiveness. In this article, we propose a new cross-domain recognition algorithm framework by integrating both of them. Specifically, this article focuses on the cross-domain recognition aspect. To be more specific, in order to learn a new feature representation that brings the data from different domains closer together as a whole, we reduce the discrepancies that exist between the different domains in terms of the conditional distribution as well as the marginal distribution. However, given that the data coming from various domains but all belonging to the same class are unable to interweave sufficiently, it would be unreasonable to combine them in order to train a single classifier. To this end, we propose learning double classifiers on each respective domain and requiring that they dynamically approximate to each other while they are being taught. This will allow us to achieve the aforementioned goal. Using this method ensures that we will eventually learn an appropriate classifier from the double classifiers by employing the tactic of classifier fusion. The results of the experiments demonstrate that the proposed method is superior to the methods that are currently considered to be state of the art.