Fast Image Retrieval Based On Deep Convolutional Neural Networks

In recent years,the popularity of smart photographic equipment has brought about an explosion in the number of images.How to retrieve the relevant contents efficiently from massive images has become a major research in the field of multimedia information retrieval.Content-based image retrieval,which allows users to retrieve the semantically similar images by inputting sample image,has been widely used in e-commerce,media design,public safety and other fields.For large-scale image retrieval,retrieval accuracy,resource consumption and retrieval efficiency are the main factors to be considered.Convolution Neural Network(CNN)can automatically learn the semantic information from images,CNN-based image retrieval has also been greatly improved in terms of retrieval accuracy.Hashing,which uses mapping functions to transform a high-dimensional feature vector into a compact and expressive binary codes,can reduce resource consumption and improve retrieval efficiency.CNN-based hashing has been an indispensable component for image retrieval.Considering the difficulty in obtaining labeled datasets for image retrieval task in large scale,we first propose a novel CNN-based unsupervised hashing method.Moreover,the existing labeled images usually have multiple labels,how to make full use of multi-label information to improve the retrieval accuracy is also a major research topic.In this paper,an unsupervised triplet hashing(UTH)method based on data augmentation is proposed.We first introduce two popular metric learning networks,i.e.,Siamese network and Triplet network.Experiment results on CIFAR-10 dataset have shown that Triplet network outperforms Siamese network in terms of retrieval accuracy.In order to learn the discriminative information from unlabeled image,we rotate each image in the unlabeled dataset by different degrees to generate multiple rotations.For each image in the unlabeled dataset,a rotation of the image,a randomly selected image from the dataset,and itself form a triplet.The discriminative loss is proposed to ensure the discriminability of the learned hash codes.The quantization loss is to minimize the discrepancy between the real-valued feature vectors and binary codes.The evenly distribution loss is to maximize the information entropy for the learned hash codes to improve their representation ability.Extensive experiment results on CIFAR-10,MNIST and In-shop datasets have shown that UTH outperforms several state-of-the-art unsupervised hashing methods in terms of retrieval accuracy.In order to learn more discriminative information from multi-label images,a multi-label deep hashing method based on mLogTriplet loss is proposed.First,we propose a mLogTriplet loss to replace the non-smooth triplet loss.Then,the WContrastive loss function is proposed to learn the multi-level similarity between images.By optimizing WContrastive loss,the similarity ranking order can be preserved.In order to make further use of the multi-label information of the images,the cross-entropy loss is used to classify multiple categories and directly learn the classification hyperplanes,thus improving the discrimination of the learned network.Extensive experiments on VOC2007 and MIRFLICKR-25 datasets have shown that the proposed multilabel hashing method outperforms several state-of-the-art multi-label hashing method.