Research On Fast Retrieval Method Of Image Data Based On Learning To Hash

With the continuous development of handheld network terminals in recent years,the popularity of social networks has exploded.The storage and retrieval of high-dimensional data has become a bottleneck factor in the use of these data.How to efficiently store and retrieve such data has attracted widespread attention in the academic and commercial industries.The hash method is an efficient indexing method.The use of hash codes to represent these high-dimensional data can reduce the storage space and improve the retrieval speed of data.The traditional hash method has achieved certain results in data retrieval,but it relies on the manual feature extraction process,which is a huge workload for large-scale data.In recent years,convolutional neural networks have achieved remarkable results in the field of image feature extraction.Researchers have combined the neural network and hashing methods to propose a new research direction deep learning to hash,and to a certain extent,have achieved good results effect.How to further improve the accuracy of hash code representation helps to improve the retrieval accuracy of data,so the research in this aspect has important significance and value.A new deep hash network model is proposed based on the combination of predecessors and their own understanding.The overall structure of the model is divided into two parts,namely the feature extraction part and the hash generation part.In the feature extraction part,the convolutional neural network is fully utilized in the field of image feature extraction.The overall framework is improved based on the AlexNet network model.The model input is a triplet,and the three have a ternary constraint relationship,making full use of the ternary Constraining the actual training,the classification effect is good,and the adaptability to the model is also good.The model uses the triplet loss function to train the model in both positive and negative directions to obtain the most expressive features of the sample.In the hash generation part,the constraint is used as a constraint,so that the Hamming distance of the hash code between the similar samples is smaller than the distance between the heterogeneous samples,and a hash code capable of retaining the image similarity and obtaining the optimal image of the sample is generated.After extracting the hash code through the deep hash network,two different optimizationalgorithms are designed for the traditional use of the Hamming distance granularity and the relationship between Hamming distance and image similarity.Weight retrieval retrieval rearrangement algorithm and quantization hash based retrieval rearrangement algorithm.In the feature weight-based rearrangement algorithm,each bit of the hash code is given a certain degree of importance.The idea of using the likelihood function is larger for a bit hash code,and its weight is larger.The quantification based on the quantized hash image retrieval is mainly applied to the hash code generated by the deep hash network.The feature will be more complete before the relaxation is hashed,and the search range is narrowed by the hash code index,and then the contrast is relaxed.The feature vector compares the similarity rearrangement output.Through theoretical research and related practices,this paper proposes a series of improvements on Hash learning image data retrieval technology,which makes the image retrieval based on hash code have better effect.