Research On Image Retrieval Technology Based On Deep Learning

With the development of information technology and the increasing of image data,it has become an essential research to solve the storage problem of massive image data scientifically.Meanwhile,it is also very important to quickly discover and extract the features of images and to deal with the rapid retrieval of massive data.A storage platform based on NoSQL is proposed because traditional relational databases cannot meet the needs of big data storage.Because of the remarkable application of deep learning in computer vision,it has become the most important research to apply deep learning to image retrieval.This paper studies the image retrieval based on deep learning.The main work is as follows:(1)The three-level retrieval structure based on SQL Server+HBase+HDFS is designed to combine the advantages of both SQL Server and HBase,which realize the fast image retrieval.Based on the Hadoop-based image storage platform,the storage structure of combining HBase with SQL Server is designed.The RowKey in SQL Server corresponds to the primary key RowKey in HBase.The image file stored in the HDFS is quickly located by redesigning the FileName in Hbase to Blockid+ Fileid+ Offset.(2)Under the Hadoop platform,we compare the storage efficiency of the three-level structure and HBase+HDFS structure on ILSVRC2012.Experimental results show that HDFS runs faster than SQL Server with the increase of storage.Although the efficiency of three-level retrieval structure is slightly lower than that of HBase+ HDFS retrieval structure,it can quickly locate the RowKey of the corresponding image,which can fast retrieve and extract images.(3)A fourteen-layer convolutional neural network is implemented under TensorFlow framework to extract image features.To verify the parallel performance of the network,the executing efficiency of the neural network under different cluster nodes is compared.Experimental results show that with the increase of cluster nodes,the training time decreases,the speedup ratio increases,and the efficiency increases.At the same time,the effects of different iterations,number of batch training samples and convolution kernel size on network performance are analyzed according to the structural characteristics of convolutional neural networks.(4)An image retrieval prototype system based on deep learning is designed and implemented.The system extracts features from the images to be searched and determines the category of the retrieved images.Then the corresponding features are queried from SQL Server according to the category.And the features are matched with the image features stored in HBase.At last,the actual image storage locations are acquired through the features and the target image is extracted.To test the performance of the image retrieval system,the time from the input image to the return of the corresponding similar images is used to evaluate the retrieval performance.The retrieval speed of the three-level retrieval structure and the HBase+ HDFS retrieval structure is compared.Experimental results showthat the three-level search structure greatly improves the speed of image retrieval.