Research On Distributed Remote Sensing Image Retrieval Based On Deep Learning

With the development of remote sensing technology,the acquisition of high-resolution remote sensing images becomes more and more convenient and remote sensing data are exploding.How to use the new technology to realize the rapid location and efficient retrieval of the object or area of interest from the high-resolution remote sensing images containing abundant ground object information has become a research hotspot nowadays.On the one hand,remote sensing images have the characteristics of diversity and multi-scale.The accuracy and reliability of feature expression model directly affect the retrieval performance.How to apply the rising deep learning technology to feature extraction in remote sensing image retrieval is still a challenge.On the other hand,remote sensing images are massive.How to use cloud computing technology to store these image data efficiently and safely and to process them quickly to better serve remote sensing data applications is an urgent problem to be solved in remote sensing field.In addition,with the improvement of remote sensing image resolution,people pay more attention to object-based retrieval.How to select an appropriate image segmentation strategy to segment the whole remote sensing image based on the object,and then retrieve similar objects,is another problem that needs to be solved in the field of remote sensing image retrieval.Aiming at the above problems,this paper makes a thorough analysis and research on image feature extraction,remote sensing image segmentation,deep learning model and Hadoop/Spark distributed framework.And a distributed remote sensing image retrieval framework based on deep learning is proposed in this paper.Firstly,based on Hadoop/Spark framework,remote sensing image data is managed and distributed parallel processing is realized.Reliable storage and management of massive remote sensing image data is realized by distributed file system HDFS.After analyzing the image processing algorithm and the mechanism of distributed operation,the image feature extraction,object-based segmentation and clustering algorithms are improved in a distributed way.Finally,the running speed of the distributed algorithm under different number of nodes is compared by experiments.Secondly,this paper studies the distributed feature extraction algorithm based on deep learning.This paper analyzes the basic structure and training method of CNN network,constructs a network model for remote sensing images by using remote sensing image fine-tuning pre-training network.Meanwhile,based on BigDL framework,distributed feature extraction algorithm based on deep learning is realized.Finally,the efficiency between single and distributed cluster,the retrieval performance between hand-craft feature and CNN feature are compared through experiments.Third,object detection algorithm is integrated into the retrieval system to achieve object-level retrieval.Based on the idea of object detection algorithm,the candidate regions of remote sensing image are segmented based on the objects,the object database is established for the segmented candidate regions,and then similar retrieval objects are returned from the object database during retrieval.The distributed remote sensing image retrieval framework based on deep learning proposed in this paper has certain theoretical and practical significance for improving the processing speed and retrieval accuracy of massive remote sensing images.