Research On Large-Scale Image Retrieval Based On Deep Hash

With the rapid development of big data,deep learning and other advanced technologies,the interaction between the Internet and user terminals has produced a huge amount of media data with various types,complex structures,and high potential value.Among them,image data has semantics,scenes,and internal data relevance.Such characteristics have gradually become the main media for people to actively share information.Image retrieval has become a representative category of current multimedia data retrieval and recognition tasks.Hash image retrieval algorithms have the advantages of compressing data,saving storage costs and fast retrieval,and can be efficiently and accurately achieved retrieval tasks.However,most of the existing researches are based on the traditional hash method.In order to facilitate the solution of the discrete optimization problem in the hash method,the traditional hash method chooses to convert the discrete constraint problem into a continuous relaxation problem,but the continuous relaxation strategy will inevitably cause quantization errors.Affect retrieval accuracy.In response to this problem,three hash algorithms are proposed in this dissertation that significantly improve retrieval accuracy in combination with deep learning.The main innovations of the dissertation are as follows:(1)To solve the problem of quadratic quantization error caused by continuous relaxation strategy,combined with a simple three-layer convolutional neural network,a deep balanced discrete hash(DSBH)method is proposed.The discrete gradient propagation of the neural network model is realized through the straight-through estimator,which solves the discrete optimization problem and reduces the quantization error caused by continuous relaxation.Under the same network framework,use pairwise label information and classification information to learn hash codes.In addition,a loss function that considers the maintenance and balance of hash code similarity is also proposed.The experimental results show that the performance of this method is better than other deep hash methods on two benchmark data sets.(2)Aiming at the problem that the regularization term is used to represent quantization error,which leads to the difficulty of neural network convergence,combined with the Alex Net,a large-scale image retrieval algorithm based on pairwise label(DPHB)is proposed,which improves the loss function to avoid the phenomenon of discrete loss of information.Based on the Alex Net network framework,a loss function considering the pairwise loss and sample label loss is proposed to better measure the error in the model training process.Based on four commonly used evaluation indicators,compared with seven advanced image retrieval methods,the experimental results show that the retrieval accuracy of the DPHB model is higher than that of the sub-optimal model,and the m AP on the two commonly used data sets are increased by 3.94% and 2.37%,respectively.From the perspective of retrieval time and tuning time,the DPHB model can respond to user instructions more quickly and return results with higher accuracy.(3)Aiming at the problem that the model based on attention mechanism pays too much attention to high-frequency features while neglecting some representative lowproportion features in the process of feature extraction,combined with the Res Net18,a large-scale image retrieval algorithm(DAHP)based on the attention mechanism is proposed.The DAHP model consists of the Res Net18 network and a dual attention module,and combines spatial anchor loss and pairwise loss to help train the model.Through a series of ablation experiments and comparative experiments,the DAHP algorithm was proved to be able to solve the problem of feature information loss in the discrete optimization process,and uses the context information of the image to improve the feature representation ability and generate a highly discriminative hash code.Based on four commonly used evaluation indicators,experiments were conducted on four benchmark data sets.The results show that the retrieval accuracy of this method is better than the baseline methods.