Deep Learning To Hash For Image Retrieval

Due to the advantages of storage overhead and retrieval efficiency,the approximate nearest neighbor search algorithm based on learning to hash has been widely used in large-scale image retrieval.In recent years,due to the improvement of computer computing power,better optimization algorithms,and the emergence of large-scale image data sets,deep convolutional neural networks have achieved unprecedented development,showing a strong ability to represent image data.Compared with the traditional hash learning method which uses handcrafted features as model input,the end-to-end deep hashing combines feature representation learning with hash-code learning,which further improves the retrieval quality of hash code.In order to leverage the potential of deep convolution neural networks,an end-to-end deep hash learning model is proposed for two different application scenarios respectively: unsupervised and supervised.1)Deep unsupervised hashing with pseudo pairwise labels.The current mainstream deep hashing methods are supervised learning.In contrast,it is difficult to obtain high-quality hash codes for unsupervised deep hashing methods due to the lack of similarity information.In practical,labeling data is an extremely time-consuming and laborious work,and for some specific areas,the participation of experts in the field is often needed to complete the labeling work.To this end,an end-to-end deep unsupervised hashing based on pseudo-pairwise labels is proposed.Its learning process includes two stages: obtaining pseudo-pairwise labels and learning to hash.In the first stage,the image features with rich semantic information extracted by the pre-trained deep convolutional neural network is used to construct the semantic similarity labels of the data.Based on this,the supervised deep hashing is performed in the second stage.Experiments on benchmark image datasets show that this method performs better on image retrieval than current state-of-the-art unsupervised deep hashing methods.2)Deep hamming embedding hashing for image retrieval.The image features learned by deep convolutional neural networks have an obvious hierarchical structure.As the number of layer deepens,the features it learns become more abstract,and the discriminability of classes also increases.Based on this,a novel hash coding method which can directly rely on the existing deep image classification network is proposed,that is,a latent layer is inserted at the end of the deep convolution neural network,and then the hash-code of the image is obtained according to the activation of each unit of the layer.At the same time,according to the characteristics of the hash-code,a hamming embedding loss is proposed to explicitly control the reservation of similarity information between data.Experiments on benchmark image datasets show that the model has a significant performance improvement compared with other deep hash learning methods,and in particular,improves the retrieval performance under short encoding length.