Image Retrieval via Gated Multiscale NetVLAD for Social Media Applications

Image retrieval on the social media platforms propels the tourism information sharing on the Internet. Existing image retrieval methods lack the capability of discovering global statistical distribution of feature representations at multiple scales. In this article, we propose a gated multiscale NetVLAD network, which constructs feature pyramid network based on ResNet backbone and computes NetVLAD features at each pyramid level to capture the multiscale information. In addition, we use gate mechanism for each level of the pyramid to adaptively represent the contribution of each level of features to the retrieval task. Experimental results on CIFAR-10, MNIST, and the Google street view datasets show that our image retrieval method has achieved better results than several existed methods.