Research On Deep Hash Image Retrieval Algorithm Base On Semantic Preservation

With the continuous upgrading of storage devices and network bandwidth,the image data in the Internet has exploded in recent years.In the case of convenient user inquiries,the problem brought about is how to quickly and accurately retrieve what users need image.Image retrieval technology based on deep hashing has brought solutions to such problems.Due to the low storage cost,fast retrieval speed and high retrieval accuracy of hash retrieval,more and more research achievements have been made on image retrieval technology based on deep hashing in recent years.How to extract a compact,highprecision hash code has become the primary problem to be solved in image retrieval based on deep hashing.Therefore,this thesis has made relevant improvements to image retrieval methods and applied them to image retrieval related data sets.The main research results are as follows:At present,the mainstream deep hashing methods generally have the problem of missing semantic information,and the generated hash codes can not well retain the semantic similarity between similar images.In order to improve the semantic information contained in hash codes,this thesis proposes a new deep hash retrieval network.The deep convolutional neural network based on semantic preservation can learn the binary code of the image by supervising information,which can maintain the semantic information between similar images,so that the generated binary code can also effectively retain the semantic information of the image.At the same time,these binary codes have also achieved better performance in classification tasks.The hash code generated by the model can ensure the retrieval accuracy while making the binary code generated from similar images have a shorter Hamming distance.There are two evaluation indicators commonly used in image retrieval: mean average Precision(m AP)and precision of the top k retrieved images.In addition,this thesis uses the idea of weighted voting in ensemble learning,combine it with image retrieval technology,and put forward an image retrieval algorithm based on neural network ensemble.It takes the value of the verification accuracy multiplied by the normalized Hamming distance in the training process of each model as the weight of each retrieval result.Then re-sort the previously retrieved images according to the obtained values,and finally calculate the two evaluation index values for the sorted images.The two deep hashing algorithms proposed in this thesis are tested on different data sets,and the experimental results are analyzed.Both methods have significantly improved the image retrieval evaluation indicators,while also ensuring the semantic information in the original image data.In addition,this thesis applies the two algorithms proposed in the actual scene,designs and completes a deep hash image retrieval system,and sets the method in this thesis into two different retrieval modes,which can use different algorithms to retrieve according to the user's choice.