Image retrieval research based on significant regions

Abstract

Deep Convolution neural networks (CNN) has achieved great success in the field of image recognition. But in the image retrieval task, the global CNN features ignore local detail description for paying too much attention to semantic information of images. So the MAP of image retrieval remains to be improved. Aiming at this problem, this paper proposes a local CNN feature extraction algorithm based on image understanding, which includes three steps: significant regions extraction, significant regions description and pool coding. This method overcomes the semantic gap problem in traditional local characteristic and improves the retrieval effect of global CNN features. Then, we apply this local CNN feature in the image retrieval task, including the same category retrieval task by feature fusion strategy and the instance retrieval task by re-ranking strategy. The experimental results show that this method has achieved good performance on the Caltech 101 and Caltech 256 classification datasets, and competitive results on the Oxford 5k and Paris 6k instance retrieval datasets.