Learning Semantic Concepts from Visual Data Using Neural Networks

For content-based image retrieval techniques, query image is used to pick up and rank some relevant images from a database using some certain similarity metric. If semantic features are not involved in the modeling of visual data, the resulting system may demonstrate a disability of retrieving images likely associated with interesting semantic concepts of objects in the images. Therefore, issues on semantics representation, automatic extraction of semantic concepts from visual data, and effects of window size on the concepts recogni- tion are needed to study. This paper describes an approach towards these prob- lems. We first define a set of semantic concepts characterizing the outdoor im- ages. Then, a neural network is employed to memory the semantic concepts through pattern learning techniques. Lastly, the well-trained neural networks will perform as a classifier to identify the predefined semantics within an image. Empirical studies and comparison with decision tree techniques are carried out.