Research On Large-Scale Image Retrieval Technology Based On Deep Hashing Learning

With the development of information technology,the image data presents an explosive growth trend.How to quickly and accurately find the user-interested image in largescale image dataset has become a hot spot in the area of multimedia information retrieval.Traditional content-based image retrieval algorithms mainly use the high-dimensional features of images to search and match.The main disadvantage of these algorithms is their high feature dimension,large storage space and low retrieval efficiency.Using hash-based image retrieval algorithm can effectively make up for these shortcomings.Hash learning map the data to a binary code,it can not only significantly reduce the data dimension,but also reduce the data storage,which can significantly improve the performance of image retrieval system.However,the traditional method of hash learning learn the hash function from the hand-crafted features of images.These hand-crafted features are extracted by unsupervised methods and can not correctly preserve the semantic similarity between images.And these methods split feature extraction and hash function learning to two parts,resulting in low retrieval accuracy of these methods,for these issues,the main work of this paper is as follows:An end-to-end deep hashing learning network structure is designed,and the deep convolutional neural network is used for both image feature extraction and hash function learning.Compared with hand-crafted features,the features extracted by deep convolutional neural networks can retain the semantic similarity between images.A loss function is designed for hash learning task,which reduces the impact of positive and negative sample imbalance on retrieval performance.At the same time,adding a binary constraint regular term to the loss function,it can reduce the error caused by quantization and avoid the loss of information caused by the traditional method due to ”relaxation” strategy.Adding a semantics-preserving layer,using the representation of the image semantics,to optimize the generation of the hash code,let the learned hash code is more able to retain the semantic similarity and improves the performance of retrieval.For the problem that the number of label data in an image dataset is far more less than the number of unlabeled data and label data is difficult to get.Enlighten by the generative adversarial networks and semi-supervised learning.A semi-supervised hash learning network based on generative adversarial network is designed.A generator is added before the network input layer to generate a ”fake” picture.In this way,both labeled and unlabeled samples can be regarded as ”real” samples.The samples generated by the generator are regarded as ”fake” samples.At the same time,the hash network is regarded as a discriminator,and a discriminant node is added on the output layer to discriminate authentic samples.The network takes full advantage of untagged data and improves the performance of the hash network.In order to verify the effectiveness of the proposed algorithm,a comparative experiment was carried out on multiple public data sets for each innovative point.The experimental results show that these innovative points can improve the accuracy of image retrieval.At the same time,combining these innovative points and compared with the other hash algorithm,experiments show that this hash-based learning algorithm has better performance in image retrieval than the current hash method.