Large Scale Image Retrieval Based On Deep Hashing Method

In the era of Artificial Intelligence,it has become a trend for users to search and obtain information through smart terminals.As an important information carrier,image data is increasing rapidly.How to retrieve the images required or interested by the user in the massive library is a high-profile issue.Large scale image retrieval based on deep hashing method is a powerful algorithm to solve this problem.It has great advantages in high storage efficiency and fast retrieval speed.In recent years,the research about the deep hashing method has mainly focused on two aspects.(1)Add hash layer in the deep neural network to extract compact visual features with high retrieval precision.(2)Create efficient index structure for visual features to improve retrieval efficiency.This thesis first analyzes some famous deep hashing networks,such as CNNH,DNNH,DSRH and DLBHC networks.CNNH is not an end-to-end method and does not make full use of learning ability of the deep neural network.DNNH improves the CNNH and uses image triples for training,which improves the image expression capabilities but requires a lot of image preprocessing work.DSRH directly learns the final evaluation indicators.Due to the non-convexity of the objective function,the final retrieval accuracy is a bit poor.DLBHC network has the highest search accuracy compared with other networks,but it cannot meet the requirements of massive image retrieval tasks by traversing search queries in Hamming space.In order to make up for the shortcomings of the existing deep hashing methods,this thesis designs an efficient searching strategy to improve the efficiency of large-scale image retrieval.The main contributions are as follows:Designing an End-to-End deep hashing network.Adding hash hidden layer in the deep neural network to learn highly ordered hash features,and semantic information is embedded into the features to preserve information in the high dimensional feature space.Each neuron in the hash layer is treated as a hash function,and the expression of the function is designed using the idea of local sensitive hashing.After training the network,the final function parameter values are obtained,which effectively alleviates the instability of the local sensitive hash function.The number of neurons in the hash layer is determined experimentally in this paper,which makes the generated hash features more compact.Experimental results prove that the hash feature has the similar prefix characteristic,which improves the accuracy of image retrieval.This thesis designs a prefix index structure by using the prefix similarity characteristic and aranges similar data objects in adjacent locations to obtain a large number of similar images with a small amount of calculation.First,the original binary hash is divided and sorted,and similar eigenvectors are arranged on the adjacent disk pages.Each disk page is indexed by an index vector.Comparing the query vector with the index vector,the CBIR system finds the index vector with the smallest distance,then loads the page represented by this index vector from the disk to get a large number of similar data objects.This indexing strategy greatly reduces the distance calculation and the query time.Finally,experiments are performed on the large-scale image datasets.The algorithm is compared with other image retrieval algorithms to verify the performance.Experimental results show that the retrieval accuracy of the proposed algorithm is higher than that of other algorithms.At the same time,the query time is reduced by an order of magnitude,which greatly improves the efficiency of image retrieval.