Multi-Negative Deep Metric Learning For Image Retrieval

Image retrieval,the task of finding images similar to user-supplied query image,has been an active topic for decades.At present,with the massive growth of image and video data,image retrieval has more and more extensive and im-portant application value.Recently,deep metric learning has achieved promising results in image retrieval by pulling similar pairs close and pushing dissimilar pairs away using siamese or triplet networks.However,penalizing individual pairs or triplets of examples may lead to poor local optimization because of insufficient neigh-borhood structure.Moreover,mining effective negative examples is still a challenging problem.In this thesis,we propose a novel deep metric learning method which penalizes multiple negative samples with different degrees of difficulties simultaneously to formulate a more complete neighborhood struc-ture.Meanwhile,in order to effectively exploit negative examples,an effective sampling strategy that combines hard negative mining and random sampling was proposed.We evaluate our method on four public datasets.Experimental results demonstrate significant improvement over state-of-the-art deep embed-ding methods.Based on the above deep metric learning model,an image retrieval system is designed and implemented.