Image Retrieval Based on Fused CNN Features

Abstract

Recently, various convolutional neural network (CNN) models have demonstrated their powerful ability as a universal representation for different image recognition tasks. In this paper, image retrieval with different kinds of CNN is considered. Firstly, we extract three kinds of CNN features with the current popular pre-trained CNN models to process image retrieval, respectively. Then, we compute weighted average of the similarity scores of these CNN features and propose an image retrieval algorithm based on the fused CNN features. Extensive experiments on two publicly available datasets well demonstrate that the proposed algorithm is clearly better than the retrieval algorithms based on individual CNN features and other current image retrieval algorithms.