A Divide and Conquer Method for Automatic Image Annotation

Abstract

Fast and accurate automatic image annotation is of great significance. Linear regression provides a fast and simple automatic image annotation method. However, it is a linear model and it is trained on the whole training data set. The computational complexity of linear regression increases with the number of training samples. In this paper, we propose a new automatic image annotation method based on data grouping. First, training samples are mapped into a new space. Next, these samples are grouped in this new space by constrained clustering. Finally, a system consisting of a softmax gate network and multiple experts is trained on the partitioned data sets. Each expert is a single-hidden-layer feedforward neural network. Experimental results on three image annotation benchmark data sets show that our method achieves better results. In addition, our experimental results show that effective grouping of training set and training an expert on each sub training set can improve the automatic image annotation performance.