Research On Robust Hash Of Image Retrieval Based On Deep Learning

With the rapid development of Internet technology,the image on the Internet has also shown explosive growth.How to quickly and accurately retrieve the target image in massive images has been a research hotspot in the field of information retrieval.The traditional hashing methods map the image feature to the Hamming space by the hashing method to reduce the feature dimension and retrieval cost.However,manual features are extracted by the unsupervised method,which focuses on the visual similarity of the image rather than the semantic similarity,so it can't achieve a satisfactory retrieval performance.Recently,due to the great success of deep learning in computer vision,deep hashing has become the main research goal of image retrieval.Deep hashing usually uses the end-to-end network to generate binary codes in the network directly,it will make the optimization of CNNs difficult in discrete hamming space(NP-hard in general).In addition,because the network is only trained in the training set,and the difference distribution of training set and real data,which results in poor generalization and robustness of the network.To address it,the specific works of this paper are as follows:(1)We propose an improved network structure based on CNN-F,which applies the "relaxed" hash function.By relaxing the binary code into a continuous space,the network can be learned through back-propagation algorithm.(2)We propose a novel loss function,which combines contrastive loss and classification loss.Contrastive loss is more suitable for retrieval tasks.Classification and weight parameter can reduce the impact of data imbalance.In addition,we do not constrain the output to be binary codes in classification part to avoid introducing any auxiliary variable.(3)Combining the above innovations,we design a pseudo-siamese network with a STN module in front of the input layer of the main network,which can learn more meaningful and various transformation on the basis of image,and increase the generalization and robustness of the network.Then the mutual learning strategy is adopted to make training stage stable and accelerate convergence.To the end,the experimental results show that our method has achieved the state-ofthe-art performance on benchmark datasets(CIFAR-10 and NUS-WIDE).For example,our method achieves the mAP of 79.8% on CIFAR-10 dataset with 12-bits code,which outperforms many other methods.