Feature Fusion Based Image Retrieval Using Deep Learning

In the last decades, Content Based Image Retrieval and image classification have become popular, and among them Region-based Image Retrieval is quite active. More and more descriptors and retrieval methods have been proposed and investigated in order to improve the retrieval performance. This paper proposed a feature fusion deep learning method. The features including colors, texture and shape, which are extracted from both the entire image and regions. The features are then trained using diverse deep learning methods. The conducted deep learning methods include Sparse Auto-encoders, Denoising Autoencoding, Deep Belief Nets, Drop Out Neural Networks, and Deep Boltzmann Machine. The method is evaluated through extensive experiments on Corel 10K datasets. Experimental results demonstrate that the introduced methods are comparable with the state-of-arts in this image retrieval application.