Image Classification via Multiple-Instance Decision-Based Neural Networks

Effective image classification becomes an important issue in content-based image retrieval since it can help to organize the massive amount of digital images and serve for many applications such as object identification, web people search, etc. In this paper, the image classification problem is considered as a Multiple-Instance Learning problem, and Multiple-Instance Decision-Based Neural Networks (MI-DBNN) with a hybrid locally unsupervised and globally supervised learning is proposed as an image classifier. For each image category, a set of related (positive) and unrelated (negative) images are selected as training examples. Then, the proposed MI-DBNN is trained according to these images. The proposed system is evaluated over the SIVAL image database and the experimental results show that our method can enhance the training accuracy from 55.1% to 68.9% and testing accuracy from 54.0% to 65.5%.