Multi-negative samples with Generative Adversarial Networks for image retrieval

摘要

The task of image retrieval has received considerable attention from the visual Al community. However, collecting large-scale labeled images for training is rather laborious and expensive. Moreover, negative samples are treated equally without considering their differences when compared with the query. To overcome those problems, in this paper, we propose to utilize generated virtual images and multiple negative samples to simultaneously learn image representations for the task of image retrieval. In our method, we first utilize the Generative Adversarial Networks in a semi-supervised fashion to produce virtual images with an adversarial loss. Second, considering the neighborhood structure within negative samples, a random sampling algorithm is proposed to effectively mining the potentially hard samples. Third, we propose a multi-negative loss function with the Kullback-Leibler divergence. Finally, by optimizing the total loss the deep neural networks are trained. Then the learned networks are further used to obtain image representations. Extensive experiments are conducted on publicly available datasets. Our model demonstrates better performances in the task of image retrieval. (C) 2019 Elsevier B.V. All rights reserved.