Multi-Deep Hashing For Large-scale Image Retrieval

Due to the powerful representation capabilities of deep networks,deep hashing has proven to be effective for large-scale image retrieval.However there are still problems that limit the improvement of retrieval performance.On the one hand,the existing deep hashing method only uses a single deep hash table.In order to achieve higher retrieval recall and accuracy,a longer hash code must be used,which results in the cost of more storage space.On the other hand,although deep features from different convolutional layers have different levels of characteristics,most existing deep hashing methods only extract feature vectors from the output of the fully connected layer forward from the last fully connected classification layer.The main focus is on semantic information,while ignoring detailed structural information.This requires research on multi-level hashing,using multi-level functions to take advantage of different fine-grained CNN characteristics.In order to solve the above problem,for the first problem,this paper proposes Weighted Multi-Deep Ranking Supervised Hashing(WMDRH)to make use of the effectiveness of multi-table deep hashing.The hash layer is embedded in the deep network.It uses multiple weighted deep hash tables to improve precision/recall without increasing space usage.Besides,it presents an order-aware ranking pairwise loss to ensure that more penalties are applied to(dissimilar)similar image pairs with(small)large Hamming distances to generate discriminative hash codes.For the second problem,in order to make full use of the characteristics of different layers of CNN,this paper proposes multi-level supervised hashing with deep features(MLSH),which uses multi-layer features to use traditional features.The COSDISH hash method trains multiple hash tables.It uses a multi-hash table mechanism to integrate the hash tables obtained from the training of multiple levels of features extracted by a corresponding single deep convolutional neural network.This method verifies the use of the complementarity between the multi-layer functions from each layer of a single deep network,which can make the hash code better retain the semantic content of the advanced feature display image while providing the structural information missing from the advanced features.However,the above method is based on the field of feature extraction.The traditional hash method is trained using deep features,so that the hash training cannot provide feedback information for feature extraction.This paper further proposes a new method on the premise that the multi-layer CNN features are valid.The Bit-wise Attention Deep Complementary Supervised Hashing(BADCSH)proposes an end-to-end system that sequentially trains multiple hash tables in an boosting manner,each hash table being trained by correcting errors caused by previous hash tables.Features from different levels are used to train the corresponding hash tables.The hash table with high-level features aims to reveal the semantic content of the image,while the hash table with low-level features obtains the structural information of the data.In addition,adding dense attention blocks to the hash layer to treat various hash bits differently,which reduces bit redundancy and maximizes overall similarity preservation.