Multiple deep CNN for image annotation

Abstract

Achieving better performance has always been an important research target in the field of automatic image annotation. This paper draws on the current popular deep learning model for the field of automatic image annotation. We propose a multiple convolutional neural networks (CNN) combination model for image annotation, which achieves satisfactory performance. First of all, we use three classical convolutional neural networks, and subsequently we examine the annotation accuracy for each CNN model. Then we take full advantage of the powerful feature representation capabilities of deep CNN, thus the last two layers of the deep CNN are extracted for each model and merged to form a new combined feature. Finally, we form our combination models by concatenating these features from each CNN model, and utilize these concatenated features to linear SVM classifier for image annotation. Experimental results on ImageCLEF2012 image annotation dataset illustrate that our combination method outperforms the traditional classifiers and the individual CNN models.