Improving scene classification by fusion of training data and web resources
Dongzhe Wang, Kezhi Mao, Gee-Wah Ng · International Conference on Information Fusion · 2015
Scene classification is often solved as a machine learning problem, where a classifier is first learned from training data, and class labels are then assigned to unlabelled testing data based on the outputs of the classifier. In this paper, we propose a novel scene classification framework that uses both training data and open resources on world wide web. This framework is inspired by human's capability to use external knowledge such as reference books or Internet when classifying something ambiguous or unknown. Specifically, we bring in the web resources in the form of text to aid visual recognition tasks. Both the classifier learned from training data and knowledge extracted from web resources are conclusive factors in the scene classification. Experimental results show that the new framework can improve scene classification accuracy by 9%.