High-resolution remote sensing image cloud computing optimization distributed retrieval method

Abstract

According to the high-resolution remote sensing image cloud computing optimization distributed retrieval method, on the basis of a self-defined MapReduce architecture IO method, a Hadoop-based high-resolution remote sensing image distributed retrieval architecture with offline library building and online retrieval is realized; a target detection method is integrated into a retrieval architecture, a method and parameters of an optimal alternative region extraction effect are determined through comparison and analysis of multiple homogeneous region extraction methods and different parameter settings and calculation of time consumption and result precision, and target positioning during retrieval is realized; the method comprises the following steps of: customizing a remote sensing image reading format and an image partitioning mode by adopting a GDAL, extracting various features of an original remote sensing image in a Map stage, searching an optimal feature combination, generating a visual word bag model through a clustering algorithm based on MapReduce, constructing a high-resolution remote sensing image database and a feature database based on Hadoop, carrying out distributed storage on image data and feature data, and storing the image data and the feature data. Retrieval acceleration is realized through parallel computing; the high-resolution remote sensing image retrieval speed is high, and the precision is high.

Description

Technical Field

[0001] The present application relates to a cloud computing retrieval method for high-resolution remote sensing images, and in particular to a cloud computing optimized distributed retrieval method for high-resolution remote sensing images, belonging to the technical field of remote sensing image retrieval. Background Art

[0002] Massive amounts of remote sensing image data provide a rich, high-quality data source for a variety of major applications based on high-resolution remote sensing imagery, playing a vital role in forestry, agriculture, surveying and mapping, marine science, economics, urban construction, and land planning. Due to the current lack of ground-based data processing capabilities, the problem of "massive data and overwhelming information" has become increasingly prominent. The current characteristics of remote sensing data, such as multiple platforms, high spatial and spectral resolution, short revisit cycles, and wide image swaths, have led to an explosive growth in the total volume of data. Over time, various departments, companies, and research institutions have accumulated massive amounts of remote sensing data. Faced with the demands of massive data-intensive computing, the traditional "standalone storage and decentralized management" approach to remote sensing data storage faces significant challenges. Therefore, how to leverage emerging technologies to distribute and manage massive amounts of high-resolution remote sensing imagery and achieve rapid location and efficient retrieval of target areas of interest from high-resolution remote sensing images, which contain rich spectral, textural, and spatial relationship information about ground objects, is an urgent challenge that needs to be addressed in the application of remote sensing imagery.

[0003] Open-source cloud computing platforms, such as Hadoop, offer high efficiency, high fault tolerance, high scalability, and low latency, making them ideal cloud platforms for storing and processing massive amounts of data. They have already achieved impressive results in areas such as online shopping, social networking, e-commerce, and e-government. Currently, Hadoop-based research focuses primarily on processing textual and symbolic data, such as logs. Hadoop itself lacks native interfaces suitable for processing massive amounts of image data, and Java is inefficient in image processing. Therefore, research on how to efficiently use Hadoop to process massive amounts of high-resolution remote sensing imagery will help optimize and enhance the storage, processing, and retrieval of these images. Combining the HDFS distributed file system and MapReduce high-performance computing architecture with the storage and processing requirements of massive amounts of high-resolution remote sensing imagery, leveraging the advantages of cluster scalability and parallel computing, and employing content-based remote sensing image retrieval techniques, a Hadoop-based distributed retrieval architecture for high-resolution remote sensing imagery is crucial for the storage, management, analysis, and utilization of massive amounts of high-resolution image data.

[0004] The problems that need to be solved in the existing high-resolution remote sensing image retrieval technology and the technical difficulties of the subject matter of this application include:

[0005] (1) Current content-based remote sensing image retrieval still has the following shortcomings:

[0006] (1) Traditional remote sensing image management methods use relational databases to manage metadata and stand-alone file systems to directly store and manage remote sensing images. However, with the diversification of remote sensing platforms, the continuous improvement of image resolution, and the continuous shortening of data acquisition cycles, remote sensing image data has exploded. Traditional storage methods have the problem of poor scalability. Once the storage space of a single machine is insufficient, the storage space can only be expanded by increasing the number of hard disks. At this time, the motherboard hard disk interface becomes the bottleneck of this storage method. Cluster technology represented by Hadoop is used to take advantage of its high scalability, high reliability, and ease of use to achieve distributed storage and management of high-resolution remote sensing image data.

[0007] (2) The size of the entire high-resolution remote sensing image is large, and the pixels of the target area to be retrieved only occupy a small proportion of the pixels of the entire remote sensing image. It is meaningless to extract feature vectors for the entire image. The high-dimensional feature vectors make the retrieval efficiency continuously reduced. The large number of complex texture features and color features in the useless background make the retrieval results confusing and invalid, and often fail to retrieve the correct target. To solve this problem, it is inappropriate to downsample the large-scale entire remote sensing image into a unified smaller-scale image. The downsampling process reduces the image resolution and causes the loss of image texture information and shape information. For the entire remote sensing image, downsampling greatly reduces the number of pixels occupied by relatively small targets such as airplanes and cars. Therefore, a good homogeneous region extraction algorithm can avoid the information loss caused by downsampling and can preliminarily eliminate some target-free areas during feature extraction, thereby reducing the feature extraction area, reducing the amount of feature data, and speeding up the efficiency of feature library construction. The target area can be located while completing the retrieval task.

[0008] (3) Traditional features need to be well designed to describe the target more accurately. The application scope of a single feature is narrow and it is often designed to solve the needs of specific application scenarios. Its generalization ability is weak. For example, the HOG feature has achieved great success in the field of pedestrian detection, but its accuracy drops significantly when processing noisy data. Deep features based on neural networks can achieve a good description of the target area. By inputting a large amount of sample data to achieve network training and optimize network parameters, the network can better describe the ground target. Therefore, combining traditional features with deep features is one of the ways to improve retrieval accuracy.

[0009] (4) In a single-machine environment, the hardware configuration largely determines the efficiency of feature clustering. Cluster-based parallel computing can greatly improve the clustering speed. Multiple feature files are processed simultaneously on multiple nodes, reducing the data processing pressure on a single node and achieving real-time or near real-time feature clustering and data processing.

[0010] (2) The use of Hadoop platform can significantly improve the efficiency of processing massive remote sensing image data. HDFS can effectively perform distributed storage and management of massive remote sensing images, and its flexible storage block strategy can better utilize hard disk space.

[0011] However, remote sensing image retrieval based on Hadoop still has the following problems:

[0012] (1) Hadoop was originally designed for the storage and processing of massive text and log data. It does not have a built-in image processing interface, so its support for processing massive image data is poor. Currently, there are few Hadoop image processing interfaces available, and they mainly support natural image formats. Compared with natural images, remote sensing images have the characteristics of large single-scene data volume, geographic coordinate information, complex textures, complex landform shapes, and complex spectra. Existing technologies convert remote sensing images into commonly used compression formats for natural images such as JPEG and PNG, and use the natural image interface of the Java language to perform Hadoop-based remote sensing image processing. Although this method solves the problem of Hadoop's support for image formats to a certain extent, image formats such as JPEG and PNG compress high-resolution remote sensing images, reducing the resolution of the original image. In addition, image formats such as JPEG and PNG do not save geographic coordinate information. The method of converting remote sensing images into formats such as JPEG does not fully utilize all the information of high-resolution remote sensing images.

[0013] (2) Homogeneous region extraction is still performed in a stand-alone environment, and the homogeneous region extraction step is not put into the cluster to run to fully utilize the advantages of the cluster. Hadoop-based image processing still physically divides image data into overlapping or non-overlapping regular regions, which causes data loss and the fragmentation of homogeneous regions. Directly reading the entire remote sensing image into a parallel computing architecture to extract features is not suitable for content-based remote sensing image retrieval, because the target to be retrieved usually only occupies a few pixels in the entire image. Directly performing feature extraction on the entire image will greatly increase the number of features and reduce the efficiency of feature extraction. A large number of objects in the image that are not related to the target add "noise" to the features, reducing the retrieval accuracy. The development of a Hadoop-based homogeneous region extraction method has practical significance for improving retrieval efficiency.

[0014] (3) Massive high-resolution remote sensing images provide a rich, high-quality data source for various major application scenarios. Faced with massive data, insufficient ground processing capabilities have made the problem of "massive data and information drowning" increasingly prominent. How to quickly locate and efficiently retrieve targets or areas of interest from high-resolution remote sensing images that contain rich spectral, texture and structural information of ground objects has become a difficult problem that urgently needs to be solved in the use of remote sensing images. The existing technology lacks target detection alternative area extraction algorithms to quickly obtain homogeneous areas in large-scale remote sensing images, lacks Hadoop-based architecture to extract multiple features and build visual word bag models, lacks distributed storage and parallel computing of remote sensing images and feature files, and the contradiction between the rapid growth of remote sensing data and the serious lack of ground data processing capabilities is prominent. There is a lack of design and implementation of a Hadoop-based high-resolution remote sensing image distributed retrieval architecture for offline database construction and online retrieval based on the customized MapReduce architecture IO method. The target detection method has not been integrated into the retrieval architecture, and the method and parameters for extracting the best candidate area cannot be determined; the GDAL custom remote sensing image reading format and image segmentation method are missing, making it impossible to extract multiple features of the original remote sensing image in the Map stage and find the best feature combination. It is impossible to build a high-resolution remote sensing image database and feature database based on Hadoop, and it is impossible to distribute and store image data and feature data. There is a lack of parallel computing to accelerate the retrieval process; high-resolution remote sensing image retrieval is time-consuming, inefficient, and has poor accuracy. Summary of the Invention

[0015] Massive high-resolution image data faces challenges of large data volumes and insufficient processing power on single machines. This application proposes a distributed retrieval architecture for high-resolution remote sensing images based on Hadoop, enabling "offline database construction and online retrieval," cluster construction, and retrieval experiments and analysis. The remote sensing image library is constructed using the HDFS distributed file system, and custom image input methods such as GDAL support multiple high-resolution remote sensing image formats. Single-scene high-resolution image data is treated as a task slice, and the image is logically partitioned using an alternative region extraction algorithm in the Map phase. This avoids the traditional Hadoop method of partitioning images into overlapping or non-overlapping regular regions, resolving the data loss problem caused by image partitioning in Hadoop. Homogeneous region extraction experiments are conducted to select the optimal algorithm for the architecture. The MapReduce parallel computing architecture is used to improve feature extraction and clustering efficiency. SIFT, Zernike, GLCM, and deep features are extracted, clustering comparison analysis is performed, and a parallel median iteration algorithm is used to process SIFT features to construct a visual bag-of-words model. Finally, an HBase-based image feature library is constructed. Finally, retrieval experiments are conducted based on this application's architecture using multiple features and different feature combinations to evaluate accuracy and identify the optimal feature combination. Experiments have shown that this architecture effectively reduces database construction time, and achieves fast and accurate retrieval of high-resolution remote sensing images.

[0016] To achieve the above technical effects, the technical solutions adopted in this application are as follows:

[0017] This paper proposes a distributed retrieval method for high-resolution remote sensing image cloud computing optimization. It uses a target detection candidate region extraction algorithm to quickly obtain homogeneous regions in large-scale remote sensing images. Based on the Hadoop architecture, it extracts multiple features and constructs a visual word bag model. It improves the distributed storage and parallel computing of remote sensing images and feature files, and establishes a distributed retrieval of high-resolution remote sensing images based on Hadoop. First, based on the customized MapReduce architecture IO method, it implements a Hadoop-based distributed retrieval architecture for high-resolution remote sensing images with offline database construction and online retrieval. Second, it integrates the target detection method into the retrieval architecture. Through comparative analysis of multiple homogeneous region extraction methods and different parameter settings, it calculates the time consumption and result accuracy to determine the method and parameters for the best candidate region extraction effect, and realizes target positioning during retrieval. Third, it uses the GDAL customized remote sensing image reading format and image segmentation method to extract multiple features of the original remote sensing image in the Map stage, find the optimal feature combination, and generate a visual word bag model through a MapReduce-based clustering algorithm. It builds a high-resolution remote sensing image database and feature database based on Hadoop, distributes the image data and feature data, and accelerates the retrieval process through parallel computing.

[0018] First, a distributed retrieval architecture for high-resolution remote sensing images improved by Hadoop is established, and the architecture is divided into two stages: offline database construction and online retrieval, and the cluster construction steps are constructed; then, the specific methods and parameters for architecture image library construction, homogeneous area extraction, feature extraction, feature clustering, and feature library construction are designed according to the offline database construction sequence, and the better algorithms and parameters are selected through comparison; then, massive whole-scene remote sensing images are stored in HDFS for distributed storage and management, and the image reading format in Hadoop is customized, and each scene remote sensing image is passed into a cluster node as a single task slice for processing, avoiding the problem of target fragmentation and information loss caused by parallel computing data division. The SelectiveSearch algorithm is first used in MapReduce. Extract candidate areas and reduce the image area processed by subsequent algorithms, so that feature extraction is only targeted at candidate areas of high-resolution images, reducing the impact of background noise on the accuracy of feature description of target areas; extract grayscale co-occurrence features GLCM, local features SIFT, depth features and Zernike features based on MapReduce architecture, store grayscale co-occurrence features, depth features and Zernike features directly in HBase, perform parallel accelerated clustering of SIFT features based on median iteration to construct a visual word bag model, and finally pass codebook data and data dictionary into HBase, and fuse them with grayscale co-occurrence features, depth features and Zernike features to construct a high-resolution remote sensing image feature library, so as to quickly combine features to meet the retrieval requirements based on multiple features.

[0019] Preferably, the retrieval architecture is designed as follows: the first to fifth steps are offline construction methods for high-resolution remote sensing image database and feature database, and the sixth to tenth steps are online retrieval methods for high-resolution remote sensing images:

[0020] Step 1: Store high-resolution remote sensing images in the HDFS distributed file system to build a remote sensing image database;

[0021] Step 2: Modify the MapReduce architecture function and customize the image input format class. The RSInputFormat class inherits from FileInputFormat.<Text,BytesWritable> , modify the isSplitable() method to set its return value to false to avoid data loss caused by forced strip-type partitioning of high-resolution remote sensing images, and implement the Map task by taking the entire high-resolution remote sensing image as a task slice input; when reading high-resolution remote sensing images from HDFS, use the inherited RecordReader<Text,BytesWritable> RSRecordReader class reads high-resolution remote sensing images in .tif format, supports multiple high-resolution remote sensing image data formats, uses GDAL to read images and serializes them into<key,value> In the key-value pair, the image name is stored in the key value as a Text data type, and the image data itself is stored in the value value as a Bytes type;

[0022] Step 3: Quickly obtain the target area to be retrieved from the large-scale high-resolution remote sensing image, use the candidate region extraction algorithm in the target detection field to extract the homogeneous area of ​​the remote sensing image, use GDAL based on the MapReduce architecture to read the virtual image data constructed by the data stream from the memory, and use the Selective Search algorithm to obtain the corresponding high-resolution remote sensing image candidate region coordinates (x, y, w, h) at each computing node, which respectively represent the screen coordinates (x, y) of the upper left point of a candidate region and the width and height of the candidate region;

[0023] Step 4: Extract features in parallel based on the Hadoop cluster, cluster the extracted features of the candidate image regions using a distributed median iterative clustering algorithm based on MapReduce, output the cluster centers to build a data dictionary, and output the frequency calculation of the cluster centers corresponding to each candidate region as the codebook for each candidate region;

[0024] Step 5: Store the data dictionary and codebook application files constructed in step 4 into HBase to implement distributed data storage management based on NoSQL database and build a feature database;

[0025] Step 6: During retrieval, the user submits the query target image;

[0026] Step 7: Extract features online and build the corresponding codebook based on the data dictionary;

[0027] Step 8: Use the distance metric function to compare the query target image codebook with each codebook in the feature database and return the similarity ranking;

[0028] Step 9: According to the codebook approximation ranking, obtain the attribute information, i.e. the candidate area (x, y, w, h) and the corresponding high-resolution remote sensing image name;

[0029] Step 10: Based on the attribute information obtained in step 9, draw the target area frame on the corresponding high-resolution remote sensing image. Finally, return the high-resolution remote sensing image with the target area frame drawn according to the similarity ranking of the candidate areas to achieve target retrieval and target positioning.

[0030] Preferably, Hadoop optimizes distributed cluster construction: Install and configure ClouderaManager and CDH on the CentOS system to complete the construction of the experimental cluster. The following are the steps for cluster construction:

[0031] Step 1: Install CentOS 8.2 system on several computers in the cluster;

[0032] Step 2: Modify the name of each node and the corresponding IP address;

[0033] Step 3: Turn off the firewall;

[0034] Step 4: Uninstall the system's native Java and install Oracle version Java 1.7;

[0035] Step 5: Install MySQL database on the master node;

[0036] Step 6: Configure inter-cluster SSH communication service;

[0037] Step 7: Configure Cloudera Manager Server on the master node;

[0038] Step 8: Synchronously configure Cloudera Manager Agent on each node in the cluster;

[0039] Step 9: Copy the CDH5 related offline package to the master node, start the Cloudera ManagerServer on the master node, and start the Cloudera Manager Agent on each node;

[0040] Step 10: Access port 7180 of the master node to visualize the cluster configuration and installation;

[0041] Step 11: Configure the cluster ntp service;

[0042] Step 12: Start the cluster.

[0043] Preferably, the remote sensing image management is optimized based on Hadoop: upload high-resolution remote sensing image files in the .tif format to the HDFS distributed file system in the cluster, divide each remote sensing image into 64M blocks, distribute them in each data node of the cluster for storage and management, and store them according to the actual size if the size is less than 64M;

[0044] Based on the construction of high-resolution remote sensing image feature library: First, homogeneous regions are extracted from the high-resolution image data in the image library. MapReduce itself does not support raster image data formats. This application uses GDAL to customize the image input format class in MapReduce to implement reading and writing support for raster remote sensing images;

[0045] Customize the RSImageInputFormat class, which inherits from FileInputFormat<Text,BytesWritable> , responsible for reading remote sensing images in HDFS, determining the slicing strategy and passing the image data to the Map function for processing. The return value of the function isSplitable(JobContextcontext, Path filename) is set to false, that is, the entire remote sensing image is sent as a task slice to a Map task for processing. The corresponding record reader RSImageRecordReader inherits from RecordReader<Text,BytesWritable> , the key value stores the name of the remote sensing image, the data format is Text, and the value value stores the remote sensing image data, the format is BytesWritable, that is, the remote sensing image data is stored in the value value in the form of a binary stream;

[0046] Customize RSImageWritable to serialize the binary stream of remote sensing image data, and transmit image coordinate parameters, name, width, height, image data type, image band information, and image data in sequence;

[0047] When the Map function calls a remote sensing image, it reads the image data from HDFS in the form of a binary stream, uses GDAL to build a virtual image file in memory, reads the image file in band order, and stores the image data of one scene in an array.

[0048] Preferably, homogeneous regions are extracted based on SelectiveSearch: in this stage, the generated remote sensing image array is received, and each computing node in the cluster receives a Map task to perform homogeneous region extraction processing on a remote sensing image. According to the homogeneous region extraction effect of MeanShift and SelectiveSearch algorithms, Selective Search is improved as a homogeneous region extraction algorithm for high-resolution remote sensing image distributed retrieval architecture based on Hadoop, and the Map function in the MapReduce architecture is used for calling. The four parameters, namely the screen coordinates (x, y) of the upper left point of each rectangular homogeneous region and the length and width (w, h) of the candidate region, are passed to the image feature extraction algorithm, so that the feature extraction area is limited to the extracted homogeneous region, reducing the computing pressure of the cluster.

[0049] Preferably, high-resolution remote sensing image feature extraction: in the Hadoop cluster Map stage, feature vector extraction is performed on the homogeneous areas extracted by Selective Search;

[0050] In this stage, SIFT features, Zernike features, GLCM features, and depth features are extracted and stored in HDFS. Each rectangular homogeneous region corresponds to a .txt feature file. The feature files are named as follows: type-name-xywh.txt, where type represents the name of the feature extraction method, name represents the name of the original image where the homogeneous region is located, and xywh represents the screen coordinates of the homogeneous region in the original full-scene remote sensing image and the length and width pixel values ​​of the homogeneous region, respectively. The SIFT feature dimension of each homogeneous region is n×128, the Zernike feature dimension is 1×656, the GLCM feature dimension is 1×5, and the depth feature uses the feature vector extracted by the fc7 layer in AlexNet, with a dimension of 1×4096.

[0051] The four extracted features are used for weighted combination to find the best feature combination.

[0052] Preferably, median iterative feature clustering based on MapReduce: a bag-of-words model for remote sensing image features is constructed through feature clustering to reduce the computational complexity of feature vector distance measurement during retrieval. SIFT feature files in HDFS are clustered using median iteration, wherein the cluster center value is set to 256, and a data dictionary with 256 "words" is constructed. Each "word" has a dimension of 1×128. For each homogeneous area, the distance between the corresponding feature file and the cluster center is measured, and the frequency of the feature vector falling in each cluster center is calculated to form a "word frequency histogram" of the feature vector, and a codebook file with a dimension of 1×256 is constructed.

[0053] First, we compare the three clustering algorithms of median iteration, AP clustering, and spectral clustering, and adopt the median iteration algorithm with obvious speed advantage. We also implement the parallel median iteration based on MapReduce to establish feature clustering to construct the image feature library, and realize the rapid clustering of remote sensing image features to construct the bag-of-words model.

[0054] Preferably, median iterative feature clustering based on MapReduce:

[0055] Algorithm: Median Iterative Clustering Algorithm Based on MapReduce

[0056] Input: File inputFile that stores feature vector data

[0057] InputPath, the storage path for feature vector data on HDFS

[0058] Initialize the storage path centerPath of the cluster center file on HDFS

[0059] ClusterPath is the storage path of clustering results on HDFS.

[0060] The number of cluster centers k

[0061] Output: k classes

[0062] Start: Read inputPath, randomly select k points as initial cluster center points, and write the k cluster center point information to centerPath

[0063] While the clustering termination condition is not met

[0064] In the Mapper stage, read the inputPath, for<key,value> The point corresponding to the key in the key-value pair traverses all cluster centers, calculates the distance, and updates the key value with the number of the nearest cluster center

[0065] Pass the key value of the changed point to the Reducer

[0066] In the Reducer stage, the key values ​​passed by the Mapper stage are received and merged, and the results are written to the clusterPath

[0067] Read clusterPath, update the cluster center using the mean of all points within the cluster center range, write the result to centerPath, and repeat this step until all cluster centers are completely updated

[0068] endwhile

[0069] Finish

[0070] The MapReduce architecture highly abstracts iterative problems into two processes, Map and Reduce, improving the efficiency of text data processing. It combines the median iterative hard-core iterative algorithm with an efficient parallel computing architecture to increase the clustering speed of massive feature vector data.

[0071] Preferably, the feature library based on HBase is optimized and constructed: this architecture stores feature data, combined feature data, data dictionary, and codebook data in HBase, where the data dictionary and codebook data are obtained by SIFT clustering, denoted as BoVW. The dimension of the data dictionary is 1×128×256, where 256 indicates that there are 256 "topic words" in the data dictionary. The dimension of each codebook data is 1×256, indicating the frequency with which the feature vector in the corresponding SIFT file falls on the "topic word" in the data dictionary.

[0072] Based on the Zernike features stored in HBase, the gray-level co-occurrence moment GLCM features and the deep features extracted from the fc7 layer in AlexNet, denoted as Deep, there are two combination feature settings: (1) GLCM and BoVW combination, with the weights of 10000 and 1 respectively; (2) GLCM+BoVW+Deep combination, with the weights of 10000, 1, and 15 respectively;

[0073] In HBase, a table is created for each rectangular homogeneous region. The table name is the homogeneous region name. In each table, the column field stores the feature name, and the value field stores the corresponding column category feature data extracted from the homogeneous region. In addition, the timestamp information in the table indicates the time when the data is operated.

[0074] Preferably, median iterative feature clustering based on MapReduce: remote sensing image target retrieval using a high-resolution remote sensing image distributed retrieval architecture based on Hadoop:

[0075] Phase 1: Upload Dataset 2 to HDFS to build a remote sensing image library;

[0076] Phase 2: Read high-resolution remote sensing images from HDFS and pass each image as a single task slice to a cluster node;

[0077] Phase 3: In the Map phase, SelectiveSearch is used to extract homogeneous regions and the screen coordinates of the upper left point and the length and width pixel values ​​of each candidate region are recorded;

[0078] Phase 4: In the Map phase, GLCM features, SIFT features, AlexNet-based fc7 features, and Zernike matrix features are extracted from the candidate regions. The feature files are named in the following format: type-name-xywh.txt. SIFT features are stored in HDFS, and the other three features are stored in HBase.

[0079] Phase 5: Start a new MapReduce task and use median iteration to perform parallel clustering on SIFT features. Set the number of cluster centers to 256, each with a dimension of 1×128. Use the cluster centers as the visual data dictionary. Store the generated codebook application and data dictionary corresponding to each homogeneous region in HBase. At this step, the feature database and visual bag-of-words model are constructed.

[0080] Stage 6: During retrieval, a single machine extracts four features of the image to be retrieved, compares the SIFT features with the cluster centers, and generates codebook data;

[0081] Phase 7: Measure the similarity of features and feature combinations, return the similarity ranking of homogeneous regions, and draw the homogeneous regions on the original high-resolution remote sensing image and return it to the user UI.

[0082] Compared with the existing technology, the innovation and advantages of this application are:

[0083] (1) Faced with massive amounts of data, insufficient ground processing capabilities have led to an increasingly prominent problem of "massive data and overwhelming information." This application aims to achieve rapid positioning and efficient retrieval of targets or regions of interest from high-resolution remote sensing images containing rich spectral, texture, and structural information of ground objects. This application uses a target detection candidate region extraction algorithm to rapidly obtain homogeneous regions in large-scale remote sensing images. Based on the Hadoop architecture, multiple features are extracted and a bag-of-visual-words model is constructed. Distributed storage and parallel computing of remote sensing images and feature files are improved, and a distributed retrieval of high-resolution remote sensing images based on Hadoop is established. Based on the contradiction between the rapid growth of remote sensing data and the severe lack of ground data processing capabilities, this application focuses on designing and implementing a distributed retrieval architecture for high-resolution remote sensing images based on Hadoop, with offline database construction and online retrieval, based on a customized MapReduce architecture IO method. The target detection method is integrated into the retrieval architecture. Through comparative analysis of various homogeneous region extraction methods and different parameter settings, the time consumption and result accuracy are calculated to determine the method and parameters for the best candidate region extraction effect, and realize target positioning during retrieval. The GDAL custom remote sensing image reading format and image segmentation method are used to extract multiple features of the original remote sensing image in the Map stage, find the best feature combination, and generate a visual word bag model through a MapReduce-based clustering algorithm. A high-resolution remote sensing image database and feature database based on Hadoop are constructed, and image data and feature data are distributedly stored. The retrieval process is accelerated through parallel computing. The Hadoop-based distributed retrieval of high-resolution images is designed and completed, which effectively improves the accuracy and efficiency of high-resolution remote sensing image retrieval with less time consumption.

[0084] (2) Massive high-resolution image data faces the problems of large data volume and insufficient single-machine processing capabilities. This application focuses on proposing and implementing a distributed retrieval architecture for high-resolution remote sensing images based on Hadoop, realizing "offline database construction and online retrieval", building clusters, and conducting retrieval experiments and analysis. The remote sensing image library is constructed using the HDFS distributed file system, and a variety of high-resolution remote sensing image formats are supported through custom image input methods such as GDAL. The single-scene high-resolution image data is used as a task slice, and the image is logically divided into blocks through the Map stage alternative region extraction algorithm, avoiding the traditional Hadoop image division method of overlapping or non-overlapping regular regions, and solving the data loss problem caused by image block in Hadoop. Homogeneous region extraction experiments are conducted to select the best algorithm for the architecture. The MapReduce parallel computing architecture is used to improve the efficiency of feature extraction and clustering. SIFT, Zernike, GLCM, and deep features are extracted respectively, clustering comparative analysis is performed, and the parallel median iteration algorithm is used to process the SIFT features to construct a visual word bag model, and an image feature library based on HBase is constructed. Finally, a retrieval experiment is conducted based on the architecture of this application using multiple features and different feature combinations, and accuracy evaluation is performed to find the best feature combination method. Experiments have shown that this architecture effectively reduces database construction time, and the retrieval speed of high-resolution remote sensing images is fast and the accuracy is high.

[0085] (3) This application proposes a content-based distributed retrieval architecture for high-resolution remote sensing images based on Hadoop, constructs a database of high-resolution remote sensing images to be retrieved and a feature database based on HDFS and HBase, and realizes distributed storage and management of image files and feature files. It combines homogeneous region extraction, multiple feature extraction and combination, and feature clustering algorithm with MapReduce architecture, uses cluster parallel computing technology to improve image and feature processing efficiency, speed up retrieval, and determine the target location during retrieval. To enhance Hadoop's support for remote sensing image processing in multiple formats, this application uses GDAL to design and implement customization of Writable type input, output, and block division methods in the Hadoop architecture. It combines the target detection concept with the traditional remote sensing image homogeneous region extraction, and replaces the commonly used overlapping or non-overlapping regular region division method in Hadoop with alternative region extraction, effectively avoiding data loss caused by the traditional Hadoop-based image block division strategy, improving retrieval accuracy, and obtaining the target bounding box coordinates during retrieval. It introduces parallel computing architecture into homogeneous region feature extraction to improve feature extraction efficiency. In addition, it extracts multiple traditional features and deep features based on neural networks at the same time, conducts comparative experiments to verify the effect of each type of feature on the experimental data set, and selects the optimal feature combination method. Based on MapReduce, parallel feature clustering is implemented to construct an image feature bag-of-words model. The data dictionary and codebook data corresponding to candidate regions are stored in HBase to build a distributed image feature library. When a user submits a search image, the image features to be retrieved are first extracted, compared with the data dictionary, and the corresponding codebook is output. This codebook is then compared with the codebook data in the image feature library based on distance, and a proximity ranking is returned. Finally, based on the proximity ranking, the candidate regions are plotted on the original remote sensing image and returned to the user. This significantly improves the efficiency of high-resolution remote sensing image retrieval, enables distributed storage and management of massive amounts of high-resolution remote sensing images, and enables rapid location and efficient retrieval of target areas of interest from high-resolution remote sensing images that contain rich spectral, texture, and spatial relationship information. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 This is the architecture diagram of the distributed retrieval of high-resolution remote sensing images improved by Hadoop.

[0087] Figure 2 This is a diagram of the average running time and average standard deviation of the three clustering algorithms.

[0088] Figure 3 This is a schematic diagram of how feature data is stored in HBase.

[0089] Figure 4 This is an example image of a remote sensing image dataset.

[0090] Figure 5It is the image of the aircraft target to be retrieved.

[0091] Figure 6 This is a diagram of the sample data of 6 scenes in dataset 2.

[0092] Figure 7 This is a schematic diagram of the image library and feature library construction process.

[0093] Figure 8 It is a schematic diagram of the user retrieval stage.

[0094] Figure 9 This is a diagram showing the average time consumption for retrieval based on different features.

[0095] Figure 10 This is a schematic diagram of the precision rate of searching six types of land features using multiple features.

[0096] Figure 11 This is a schematic diagram of the completeness rate of searching six types of land features using multiple features. DETAILED DESCRIPTION

[0097] The following, in conjunction with the accompanying drawings, further describes the technical solution of the high-resolution remote sensing image cloud computing optimized distributed retrieval method provided by this application, so that those skilled in the art can better understand this application and implement it.

[0098] The speed and number of remote sensing satellite launches have entered an accelerated stage, and the use of drones has greatly accelerated the updating speed of high-resolution remote sensing images. Massive high-resolution remote sensing images provide a rich and high-quality data source for various major application scenarios. Faced with massive data, insufficient ground processing capabilities have made the problem of "massive data and overwhelming information" increasingly prominent. How to achieve rapid positioning and efficient retrieval of targets or areas of interest from high-resolution remote sensing images that contain rich spectral, texture and structural information of ground objects has become a difficult problem that urgently needs to be solved in the use of remote sensing images. This application uses a target detection alternative region extraction algorithm to quickly obtain homogeneous areas in large-scale remote sensing images, extracts multiple features based on the Hadoop architecture and constructs a visual word bag model, improves the distributed storage and parallel computing of remote sensing images and feature files, and establishes a distributed retrieval of high-resolution remote sensing images based on Hadoop improvements.

[0099] In response to the contradiction between the rapid growth of remote sensing data and the severe shortage of ground data processing capabilities, based on a customized MapReduce architecture IO method, this paper focuses on designing and implementing a Hadoop-based distributed retrieval architecture for high-resolution remote sensing images with offline database construction and online retrieval. This application covers the following aspects:

[0100] (1) Integrate the target detection method into the retrieval architecture, conduct comparative analysis of multiple homogeneous region extraction methods and different parameter settings, calculate the time consumption and result accuracy to determine the best candidate region extraction method and parameters, and achieve target positioning during retrieval;

[0101] (2) Using GDAL to customize the remote sensing image reading format and image segmentation method, extract multiple features of the original remote sensing image in the Map stage, find the best feature combination, generate a visual word bag model based on the MapReduce clustering algorithm, build a high-resolution remote sensing image database and feature database based on Hadoop, perform distributed storage of image data and feature data, and accelerate the retrieval process through parallel computing;

[0102] (3) Taking aircraft targets as an example, a distributed retrieval system for high-resolution images based on Hadoop was designed and implemented. Experiments have shown that this architecture effectively improves the retrieval accuracy of high-resolution remote sensing images with minimal time consumption.

[0103] 1. Hadoop-improved distributed retrieval architecture for high-resolution remote sensing images

[0104] (1) Retrieval architecture design

[0105] The search process is as follows Figure 1 As shown in the figure, the first to fifth steps are the offline construction methods of high-resolution remote sensing image database and feature database, and the sixth to tenth steps are the online retrieval methods of high-resolution remote sensing images:

[0106] Step 1: Store high-resolution remote sensing images in the HDFS distributed file system to build a remote sensing image database;

[0107] Step 2: Modify the MapReduce architecture function and customize the image input format class. The RSInputFormat class inherits from FileInputFormat.<Text,BytesWritable> , modify the isSplitable() method to set its return value to false to avoid data loss caused by forced strip-type partitioning of high-resolution remote sensing images, and implement the Map task by taking the entire high-resolution remote sensing image as a task slice input; when reading high-resolution remote sensing images from HDFS, use the inherited RecordReader<Text,BytesWritable> RSRecordReader class reads high-resolution remote sensing images in .tif format, supports multiple high-resolution remote sensing image data formats, uses GDAL (Geospatial Data Abstraction Library) to read images and serializes them into<key,value> In the key-value pair, the image name is stored in the key value as a Text data type, and the image data itself is stored in the value value as a Bytes type;

[0108] Step 3: Quickly obtain the target area to be retrieved from the large-scale high-resolution remote sensing image, use the candidate region extraction algorithm in the target detection field to extract the homogeneous area of ​​the remote sensing image, use GDAL (Geospatial Data Abstraction Library) based on the MapReduce architecture to read the virtual image data constructed by the data stream from the memory, and use the Selective Search algorithm to obtain the coordinates (x, y, w, h) of the candidate area of ​​the corresponding high-resolution remote sensing image at each computing node, which respectively represent the screen coordinates (x, y) of the upper left point of a candidate area and the width and height of the candidate area;

[0109] Step 4: Extract features in parallel based on the Hadoop cluster, cluster the extracted features of the candidate image regions using a distributed median iterative clustering algorithm based on MapReduce, output the cluster centers to build a data dictionary, and output the frequency calculation of the cluster centers corresponding to each candidate region as the codebook for each candidate region;

[0110] Step 5: Store the data dictionary and codebook application files constructed in step 4 into HBase to implement distributed data storage management based on NoSQL database and build a feature database;

[0111] Step 6: During retrieval, the user submits the query target image;

[0112] Step 7: Extract features online and build the corresponding codebook based on the data dictionary;

[0113] Step 8: Use the distance metric function to compare the query target image codebook with each codebook in the feature database and return the similarity ranking;

[0114] Step 9: According to the codebook approximation ranking, obtain the attribute information, i.e. the candidate area (x, y, w, h) and the corresponding high-resolution remote sensing image name;

[0115] Step 10: Based on the attribute information obtained in step 9, draw the target area frame on the corresponding high-resolution remote sensing image. Finally, return the high-resolution remote sensing image with the target area frame drawn according to the similarity ranking of the candidate areas to achieve target retrieval and target positioning.

[0116] (2) Hadoop Optimization Distributed Cluster Construction

[0117] This application installs and configures Cloudera Manager and CDH on a CentOS system to complete the construction of an experimental cluster. The following are the steps for cluster construction:

[0118] Step 1: Install CentOS 8.2 system on several computers in the cluster;

[0119] Step 2: Modify the name of each node and the corresponding IP address;

[0120] Step 3: Turn off the firewall;

[0121] Step 4: Uninstall the system's native Java and install Oracle version Java 1.7;

[0122] Step 5: Install MySQL database on the master node;

[0123] Step 6: Configure inter-cluster SSH communication service;

[0124] Step 7: Configure Cloudera Manager Server on the master node;

[0125] Step 8: Synchronously configure Cloudera Manager Agent on each node in the cluster;

[0126] Step 9: Copy the CDH5 related offline package to the master node, start the Cloudera ManagerServer on the master node, and start the Cloudera Manager Agent on each node;

[0127] Step 10: Access port 7180 of the master node to visualize the cluster configuration and installation;

[0128] Step 11: Configure the cluster ntp service;

[0129] Step 12: Start the cluster.

[0130] (3) Optimized management of remote sensing images based on Hadoop

[0131] Upload high-resolution remote sensing image files in .tif format to the HDFS distributed file system in the cluster. Each remote sensing image is divided into 64MB blocks and distributed across the cluster's data nodes for storage and management. If the block size is less than 64MB, it will be stored according to the actual size.

[0132] Based on the construction of high-resolution remote sensing image feature library: First, homogeneous regions are extracted from the high-resolution image data in the image library. MapReduce itself does not support raster image data formats. This application uses GDAL to customize the image input format class in MapReduce to implement reading and writing support for raster remote sensing images;

[0133] Customize the RSImageInputFormat class, which inherits from FileInputFormat<Text,BytesWritable> , responsible for reading remote sensing images in HDFS, determining the slicing strategy and passing the image data to the Map function for processing. The return value of the function isSplitable(JobContextcontext, Path filename) is set to false, that is, the entire remote sensing image is sent as a task slice to a Map task for processing. The corresponding record reader RSImageRecordReader inherits from RecordReader<Text,BytesWritable> , the key value stores the name of the remote sensing image, the data format is Text, and the value value stores the remote sensing image data, the format is BytesWritable, that is, the remote sensing image data is stored in the value value in the form of a binary stream;

[0134] Customize RSImageWritable to serialize the binary stream of remote sensing image data, and transmit image coordinate parameters, name, width, height, image data type, image band information, and image data in sequence;

[0135] When the Map function calls a remote sensing image, it reads the image data from HDFS in the form of a binary stream, uses GDAL to build a virtual image file in memory, reads the image file in band order, and stores the image data of one scene in an array.

[0136] (IV) Extracting homogeneous regions based on SelectiveSearch

[0137] In this stage, the generated remote sensing image array is received. Each computing node in the cluster receives a Map task to extract homogeneous regions from a remote sensing image. Based on the homogeneous region extraction effects of the MeanShift and SelectiveSearch algorithms, the Selective Search algorithm is improved as a homogeneous region extraction algorithm for the distributed retrieval architecture of high-resolution remote sensing images based on Hadoop. The Map function in the MapReduce architecture is used to call it, and the four parameters, namely the screen coordinates (x, y) of the upper left point of each rectangular homogeneous region and the length and width (w, h) of the candidate region, are passed to the image feature extraction algorithm. This ensures that the feature extraction area is limited to the extracted homogeneous region, reducing the computing pressure of the cluster and improving the retrieval speed.

[0138] (5) Feature extraction of high-resolution remote sensing images

[0139] In the Hadoop cluster Map phase, feature vectors are extracted from the homogeneous regions extracted by Selective Search;

[0140] In this stage, SIFT features, Zernike features, GLCM features, and depth features are extracted and stored in HDFS. Each rectangular homogeneous region corresponds to a .txt feature file. The feature files are named as follows: type-name-xywh.txt, where type represents the name of the feature extraction method, name represents the name of the original image where the homogeneous region is located, and xywh represents the screen coordinates of the homogeneous region in the original full-scene remote sensing image and the length and width pixel values ​​of the homogeneous region, respectively. The SIFT feature dimension of each homogeneous region is n×128, the Zernike feature dimension is 1×656, the GLCM feature dimension is 1×5, and the depth feature uses the feature vector extracted by the fc7 layer in AlexNet, with a dimension of 1×4096.

[0141] The four extracted features are used for weighted combination to find the best feature combination.

[0142] (VI) Median Iterative Feature Clustering Based on MapReduce

[0143] A bag-of-words model for remote sensing image features is constructed through feature clustering to reduce the computational complexity of feature vector distance measurement during retrieval. SIFT feature files in HDFS are clustered using median iteration, with the cluster center value set to 256 (empirical value). A data dictionary with 256 "words" is constructed, each with a dimension of 1×128. For each homogeneous region, the distance between the corresponding feature file and the cluster center is measured, and the frequency of the feature vector falling within each cluster center is calculated to form a "word frequency histogram" of the feature vector. A codebook file with a dimension of 1×256 is constructed.

[0144] First, we compare the three clustering algorithms: median iteration, AP clustering, and spectral clustering. We use the SIFT feature file extracted from the "UC MercedLand Use Dataset". The dataset contains 21 land types, 100 images per category, and each image has a pixel size of 256*256. This dataset is small, with only 100 images per category. This dataset has the characteristics of large intra-class differences and small inter-class differences, which poses great challenges for feature extraction and clustering. Figure 2 The average running time and average standard deviation of the three clustering algorithms.

[0145] The median iteration algorithm with obvious speed advantage is adopted, and the parallel median iteration based on MapReduce is implemented to establish feature clustering and build image feature library, so as to realize the rapid clustering of remote sensing image features and build bag-of-words model.

[0146] Algorithm: Median Iterative Clustering Algorithm Based on MapReduce

[0147] Input: File inputFile that stores feature vector data

[0148] InputPath, the storage path for feature vector data on HDFS

[0149] Initialize the storage path centerPath of the cluster center file on HDFS

[0150] ClusterPath is the storage path of clustering results on HDFS.

[0151] The number of cluster centers k

[0152] Output: k classes

[0153] Start: Read inputPath, randomly select k points as initial cluster center points, and write the k cluster center point information to centerPath

[0154] While the clustering termination condition is not met

[0155] In the Mapper stage, read the inputPath, for<key,value> The point corresponding to the key in the key-value pair traverses all cluster centers, calculates the distance, and updates the key value with the number of the nearest cluster center

[0156] Pass the key value of the changed point to the Reducer

[0157] In the Reducer stage, the key values ​​passed by the Mapper stage are received and merged, and the results are written to the clusterPath

[0158] Read clusterPath, update the cluster center using the mean of all points within the cluster center range, write the result to centerPath, and repeat this step until all cluster centers are completely updated

[0159] endwhile

[0160] Finish

[0161] The MapReduce architecture highly abstracts the iterative problem into two processes, Map and Reduce, improving the efficiency of text data processing. It combines the median iterative hard-core iterative algorithm with an efficient parallel computing architecture, which can greatly improve the clustering speed of massive feature vector data and realize the median iterative clustering algorithm based on MapReduce.

[0162] (VII) Optimization and construction of feature library based on HBase

[0163] This architecture stores feature data, combined feature data, data dictionary, and codebook data in HBase. The data dictionary and codebook data are derived from SIFT clustering and are denoted as BoVW. The data dictionary dimension is 1 × 128 × 256, where 256 represents the number of "keywords" in the data dictionary. Each codebook data item has a dimension of 1 × 256, indicating the frequency with which the feature vector in the corresponding SIFT file falls on a data dictionary "keyword."

[0164] Based on the Zernike features stored in HBase, the gray-level co-occurrence moment GLCM features and the deep features extracted from the fc7 layer in AlexNet, denoted as Deep, there are two combination feature settings: (1) GLCM and BoVW combination, with the weights of 10,000 and 1 respectively; (2) GLCM+BoVW+Deep combination, with the weights of 10,000, 1, and 15 respectively.

[0165] In HBase, a table is created for each rectangular homogeneous region. The table name is the homogeneous region name. In each table, the column field stores the feature name, and the value field stores the corresponding column category feature data extracted from the homogeneous region. In addition, the timestamp information in the table indicates the time when the data is operated, such as Figure 3 shown.

[0166] This application first establishes a distributed retrieval architecture for high-resolution remote sensing images improved by Hadoop, divides the architecture into two stages: offline database construction and online retrieval, and constructs cluster building steps to provide a basic environment for subsequent experiments;

[0167] Then, according to the offline database construction sequence, the specific methods and parameters for architecture image library construction, homogeneous region extraction, feature extraction, feature clustering, and feature library construction are designed, and the better algorithms and parameters are selected through comparison;

[0168] This architecture stores massive whole-scene remote sensing images in HDFS for distributed storage and management, customizes the image reading format in Hadoop, and transfers each remote sensing image as a single task slice to a cluster node for processing, avoiding target fragmentation and information loss problems caused by parallel computing data partitioning. In MapReduce, the SelectiveSearch algorithm is first used to extract candidate areas, reducing the image area processed by subsequent algorithms, so that feature extraction is only targeted at candidate areas of high-resolution images, reducing the impact of background noise on the accuracy of feature description of target areas; based on the MapReduce architecture, grayscale co-occurrence features GLCM, local features SIFT, depth features and Zernike features are extracted, and the grayscale co-occurrence features, depth features and Zernike features are directly stored in HBase. The SIFT features are clustered in parallel based on median iteration to construct a visual word bag model. Finally, the codebook data and data dictionary are transferred to HBase and fused with the grayscale co-occurrence features, depth features and Zernike features to construct a high-resolution remote sensing image feature library, and feature combination is quickly performed to meet the retrieval requirements based on multiple features.

[0169] 2. High-resolution remote sensing image cloud computing distributed retrieval experiment

[0170] This study used a distributed retrieval architecture for high-resolution remote sensing images based on Hadoop to conduct aircraft target retrieval. In this experiment, the grayscale co-occurrence feature is denoted as GLCM, the bag-of-visual-words model generated by SIFT clustering of local features is denoted as BoVW, the Zernike feature is denoted as Zernike, and the fc7 layer feature in AlexNet is denoted as Deep. The "+" symbol indicates a combination of multiple features. For example, GLCM+BoVW represents a weighted combination of the grayscale co-occurrence feature and the bag-of-visual-words model generated by SIFT clustering.

[0171] (1) Experimental data

[0172] Dataset 1: This dataset contains 10 high-resolution remote sensing images of aircraft targets, which are used as user input data for retrieval. The resolution is 1 meter and the parameters are as follows: Figure 4 As shown, Figure 5 These are the 10 aircraft images.

[0173] Dataset 2: This dataset contains 102 remote sensing images with different sizes. The number of pixels in length and width are in the range of [990, 1400], and the resolution is between 0.5 meters and 1.5 meters. Figure 6 These are the 6 scene image examples in this dataset, which are divided into six categories.

[0174] (2) Experimental methods

[0175] An experiment on remote sensing image aircraft target retrieval was conducted using a Hadoop-based high-resolution remote sensing image distributed retrieval architecture. Figure 7 For the image library and feature library building process, Figure 8 For user retrieval stage:

[0176] Phase 1: Upload Dataset 2 to HDFS to build a remote sensing image library;

[0177] Phase 2: Read high-resolution remote sensing images from HDFS and pass each image as a single task slice to a cluster node;

[0178] Phase 3: In the Map phase, SelectiveSearch is used to extract homogeneous regions and the screen coordinates of the upper left point and the length and width pixel values ​​of each candidate region are recorded;

[0179] Phase 4: In the Map phase, GLCM features (5 dimensions), SIFT features (128 dimensions per row), AlexNet-based fc7 features (4096 dimensions), and Zernike matrix features (656 dimensions) are extracted from the candidate regions. The feature files are named in the following format: type-name-xywh.txt. SIFT features are stored in HDFS, and the other three features are stored in HBase.

[0180] Phase 5: Start a new MapReduce task and use median iteration to perform parallel clustering on SIFT features. Set the number of cluster centers to 256, each with a dimension of 1×128. Use the cluster centers as the visual data dictionary. Store the generated codebook application and data dictionary corresponding to each homogeneous region in HBase. At this step, the feature database and visual bag-of-words model are constructed.

[0181] Stage 6: During retrieval, a single machine extracts four features of the image to be retrieved, compares the SIFT features with the cluster centers, and generates codebook data;

[0182] Phase 7: Measure the similarity of features and feature combinations, return the similarity ranking of homogeneous regions, and draw the homogeneous regions on the original high-resolution remote sensing image and return it to the user UI.

[0183] (3) Experimental results and analysis

[0184] This experiment performed 10 searches for each aircraft target image to be retrieved, and the average value was calculated. The experiment used Dataset 2 to construct the image and feature libraries. For high-resolution remote sensing images with a certain resolution, the pixel size range of aircraft targets is deterministic. Therefore, this experiment performed a preliminary threshold screening on the homogeneous regions extracted by SelectiveSearch, limiting the length and width of the homogeneous regions to the interval [100, 400]. As a result, 3208 homogeneous regions were extracted from 102 high-resolution images. Of these, 431 homogeneous regions were considered to be aircraft targets, 340 were considered to be water surfaces, 997 were considered to be roads, 121 were considered to be beaches, 1094 were considered to be built-up areas, and 225 were considered to be vegetation.

[0185] Using dataset 1, the average time consumed for retrieval based on different features is as follows Figure 9 As shown;

[0186] The precision and recall indicators are used to evaluate the accuracy of the retrieval results.

[0187] The combination of GLCM+BoVW+Deep achieves the highest precision and recall, which are 56.8% and 18.2% respectively, but at the cost of a lot of time consumption.

[0188] The GLCM+BoVW combination uses less search time, achieving a recall of 18.1%, just 0.1% less than the GLCM+BoVW+Deep combination. While reducing search time by 22 seconds, it also achieves a precision of 45.3%, just below the 56.8% of the GLCM+BoVW+Deep combination and the 52.5% of the Deep feature.

[0189] In order to verify the precision and recall of multiple features or feature combinations in retrieval of different ground objects, images of each category were randomly selected from 3208 homogeneous areas as user input for retrieval experiments. The number of homogeneous areas of the correct category in the first 100 homogeneous areas returned was calculated. The retrieval precision of the six ground objects of airplane, water surface, road, bare soil, building, and vegetation is as follows: Figure 10 As shown, the recall rate is Figure 11 As shown in the figure, G+B represents the weighted combination of the grayscale co-occurrence feature GLCM and the visual word bag model BoVW generated after SIFT feature clustering, and G+B+D represents the weighted combination of the grayscale co-occurrence feature GLCM, the visual word bag model BoVW generated after SIFT feature clustering, and the features extracted from the fc7 layer in AlexNet (denoted as Deep).

[0190] Experiments demonstrate that the GLCM+BoVW+Deep feature combination achieves the highest precision and recall across five ground object retrieval tasks. Texture features perform poorly when retrieval is performed for water surfaces, easily confusing them with bare soil. Deep networks pre-trained on ImageNet demonstrate significant advantages in describing building features, achieving significantly higher precision than other features. Deep features achieve the highest precision and recall for bare soil object retrieval. This is because the dataset used in this experiment contains a large number of sandy beaches within the bare soil, which can easily be confused with shallow sea surfaces. Shallow features perform poorly for this retrieval task, reducing feature description accuracy during feature combination, resulting in lower combined feature accuracy than deep features. Vegetation achieves the highest overall retrieval accuracy, as its texture and color are clearly distinguishable from other ground objects. When retrieval is performed for all six ground object categories using GLCM, vegetation features achieve significantly higher retrieval accuracy than the other five. BoVW feature retrieval results are generally poor, as the choice of cluster centers significantly influences the accuracy of feature descriptions in the bag-of-visual-words model.

[0191] Content-based image retrieval can greatly improve database construction efficiency and reduce the amount of manual annotation work. Combining content-based high-resolution remote sensing image retrieval with the Hadoop distributed computing architecture improves the efficiency of feature extraction and clustering during database construction, fully leveraging the advantages of cluster distributed storage to store and manage massive amounts of high-resolution remote sensing images and feature data.

[0192] This application uses query rate and recall rate as evaluation indicators. First, in a Hadoop cluster with 5 nodes, a variety of features and feature combinations are used to conduct aircraft target retrieval experiments based on the architecture proposed in this application. The retrieval time consumption of various features or feature combinations is calculated, and the results are analyzed. The number of aircraft in the first 100 candidate areas and the number of aircraft in the first 200 candidate areas in the returned retrieval results are calculated, and the precision rate and recall rate are calculated and analyzed. The candidate areas obtained after the SelectiveSearch processing of the data set 2 are divided into 6 land object categories. Retrieval experiments are conducted on the six land object categories to analyze the retrieval efficiency of the retrieval architecture of this application when facing the retrieval task of multiple land object categories of high-resolution remote sensing images using different features or feature combinations.

[0193] Experiments have demonstrated that the proposed distributed retrieval architecture for high-resolution remote sensing images based on Hadoop can effectively handle high-resolution image retrieval tasks for a variety of object categories, effectively storing and managing massive amounts of high-resolution remote sensing image data in a distributed manner. The use of the candidate region algorithm and MapReduce architecture significantly improves database construction efficiency.

Claims

1. A distributed retrieval method for high-resolution remote sensing image cloud computing optimization, characterized by: The target detection candidate region extraction algorithm is used to quickly obtain homogeneous regions in large-scale remote sensing images. Based on the Hadoop architecture, multiple features are extracted and a visual word bag model is constructed. The distributed storage and parallel computing of remote sensing images and feature files are improved, and a distributed retrieval of high-resolution remote sensing images based on Hadoop is established. First, based on the customized MapReduce architecture IO method, a distributed retrieval architecture for high-resolution remote sensing images based on Hadoop is realized with offline database construction and online retrieval. Second, the target detection method is integrated into the retrieval architecture. Through comparative analysis of multiple homogeneous region extraction methods and different parameter settings, the method and parameters for the best candidate region extraction effect are determined by calculating the time consumption and result accuracy, so as to achieve target positioning during retrieval. Third, the GDAL customized remote sensing image reading format and image segmentation method are used to extract multiple features of the original remote sensing image in the Map stage, find the best feature combination, generate a visual word bag model through the MapReduce-based clustering algorithm, build a high-resolution remote sensing image database and feature database based on Hadoop, distribute the image data and feature data, and accelerate the retrieval process through parallel computing. First, a distributed retrieval architecture for high-resolution remote sensing images improved by Hadoop is established, and the architecture is divided into two stages: offline database construction and online retrieval, and the cluster construction steps are constructed; then, the specific methods and parameters for architecture image library construction, homogeneous area extraction, feature extraction, feature clustering, and feature library construction are designed according to the offline database construction sequence, and the better algorithms and parameters are selected through comparison; then, massive whole-scene remote sensing images are stored in HDFS for distributed storage and management, and the image reading format in Hadoop is customized, and each scene remote sensing image is passed into a cluster node as a single task slice for processing, avoiding the problem of target fragmentation and information loss caused by parallel computing data division. The SelectiveSearch algorithm is first used in MapReduce. Extract candidate areas and reduce the image area processed by subsequent algorithms, so that feature extraction is only targeted at candidate areas of high-resolution images, reducing the impact of background noise on the accuracy of feature description of target areas; extract grayscale co-occurrence features GLCM, local features SIFT, depth features and Zernike features based on MapReduce architecture, store grayscale co-occurrence features, depth features and Zernike features directly in HBase, perform parallel accelerated clustering of SIFT features based on median iteration to construct a visual word bag model, and finally pass codebook data and data dictionary into HBase, and fuse them with grayscale co-occurrence features, depth features and Zernike features to construct a high-resolution remote sensing image feature library, so as to quickly combine features to meet the retrieval requirements based on multiple features.

2. The high-resolution remote sensing image cloud computing optimized distributed retrieval method according to claim 1 is characterized in that: Retrieval architecture design: Steps 1 to 5 are the offline construction methods of high-resolution remote sensing image database and feature database, and steps 6 to 10 are the online retrieval methods of high-resolution remote sensing images: Step 1: Store high-resolution remote sensing images in the HDFS distributed file system to build a remote sensing image database; Step 2: Modify the MapReduce architecture function and customize the image input format class. The RSInputFormat class inherits from FileInputFormat.<Text,BytesWritable> , modify the isSplitable() method to set its return value to false to avoid data loss caused by forced strip-type partitioning of high-resolution remote sensing images, and implement the Map task by taking the entire high-resolution remote sensing image as a task slice input; when reading high-resolution remote sensing images from HDFS, use the inherited RecordReader<Text,BytesWritable> RSRecordReader class reads high-resolution remote sensing images in .tif format, supports multiple high-resolution remote sensing image data formats, uses GDAL to read images and serializes them into<key,value> In the key-value pair, the image name is stored in the key value as a Text data type, and the image data itself is stored in the value value as a Bytes type; Step 3: Quickly obtain the target area to be retrieved from the large-scale high-resolution remote sensing image, use the candidate region extraction algorithm in the target detection field to extract the homogeneous area of ​​the remote sensing image, use GDAL based on the MapReduce architecture to read the virtual image data constructed by the data stream from the memory, and use the Selective Search algorithm to obtain the corresponding high-resolution remote sensing image candidate region coordinates (x, y, w, h) at each computing node, which respectively represent the screen coordinates (x, y) of the upper left point of a candidate region and the width and height of the candidate region; Step 4: Extract features in parallel based on the Hadoop cluster, cluster the extracted features of the candidate image regions using a distributed median iterative clustering algorithm based on MapReduce, output the cluster centers to build a data dictionary, and output the frequency calculation of the cluster centers corresponding to each candidate region as the codebook for each candidate region; Step 5: Store the data dictionary and codebook application files constructed in step 4 into HBase to implement distributed data storage management based on NoSQL database and build a feature database; Step 6: During retrieval, the user submits the query target image; Step 7: Extract features online and build the corresponding codebook based on the data dictionary; Step 8: Use the distance metric function to compare the query target image codebook with each codebook in the feature database and return the similarity ranking; Step 9: According to the codebook approximation ranking, obtain the attribute information, i.e. the candidate area (x, y, w, h) and the corresponding high-resolution remote sensing image name; Step 10: Based on the attribute information obtained in step 9, draw the target area frame on the corresponding high-resolution remote sensing image. Finally, return the high-resolution remote sensing image with the target area frame drawn according to the similarity ranking of the candidate areas to achieve target retrieval and target positioning.

3. The high-resolution remote sensing image cloud computing optimized distributed retrieval method according to claim 1 is characterized in that: Hadoop optimized distributed cluster construction: Install and configure Cloudera Manager and CDH on the CentOS system to complete the construction of the experimental cluster. The following are the steps for cluster construction: Step 1: Install CentOS 8.2 system on several computers in the cluster; Step 2: Modify the name of each node and the corresponding IP address; Step 3: Turn off the firewall; Step 4: Uninstall the system's native Java and install Oracle version Java 1.7; Step 5: Install MySQL database on the master node; Step 6: Configure inter-cluster SSH communication service; Step 7: Configure Cloudera Manager Server on the master node; Step 8: Synchronously configure Cloudera Manager Agent on each node in the cluster; Step 9: Copy the CDH5-related offline packages to the master node, start the Cloudera Manager Server on the master node, and start the Cloudera Manager Agent on each node; Step 10: Access port 7180 of the master node to visualize the cluster configuration and installation; Step 11: Configure the cluster ntp service; Step 12: Start the cluster.

4. The high-resolution remote sensing image cloud computing optimized distributed retrieval method according to claim 1, characterized in that: Hadoop-based remote sensing image optimization management: Upload high-resolution remote sensing image files in .tif format to the HDFS distributed file system in the cluster. Each remote sensing image is divided into 64MB blocks and distributed across the cluster's data nodes for storage and management. If the block size is less than 64MB, it will be stored according to the actual size. Based on the construction of high-resolution remote sensing image feature library: First, homogeneous regions are extracted from the high-resolution image data in the image library. MapReduce itself does not support raster image data formats. This application uses GDAL to customize the image input format class in MapReduce to implement reading and writing support for raster remote sensing images; Customize the RSImageInputFormat class, which inherits from FileInputFormat<Text,BytesWritable> , responsible for reading remote sensing images in HDFS, determining the slicing strategy and passing the image data to the Map function for processing. The return value of the function isSplitable(JobContextcontext, Path filename) is set to false, that is, the entire remote sensing image is sent as a task slice to a Map task for processing. The corresponding record reader RSImageRecordReader inherits from RecordReader<Text,BytesWritable> , the key value stores the name of the remote sensing image, the data format is Text, and the value value stores the remote sensing image data, the format is BytesWritable, that is, the remote sensing image data is stored in the value value in the form of a binary stream; Customize RSImageWritable to serialize the binary stream of remote sensing image data, and transmit image coordinate parameters, name, width, height, image data type, image band information, and image data in sequence; When the Map function calls a remote sensing image, it reads the image data from HDFS in the form of a binary stream, uses GDAL to build a virtual image file in memory, reads the image file in band order, and stores the image data of one scene in an array.

5. The high-resolution remote sensing image cloud computing optimized distributed retrieval method according to claim 1, characterized in that: Extract homogeneous regions based on SelectiveSearch: This stage receives the generated remote sensing image array, and each computing node in the cluster receives a Map task to perform homogeneous region extraction processing on a remote sensing image. According to the homogeneous region extraction effect of MeanShift and SelectiveSearch algorithms, Selective Search is improved as a homogeneous region extraction algorithm for high-resolution remote sensing image distributed retrieval architecture based on Hadoop. The Map function in the MapReduce architecture is used for calling, and the four parameters of the upper left screen coordinates (x, y) of each rectangular homogeneous region and the length and width (w, h) of the candidate region are passed to the image feature extraction algorithm, so that the feature extraction area is limited to the extracted homogeneous region, reducing the computing pressure of the cluster.

6. The high-resolution remote sensing image cloud computing optimized distributed retrieval method according to claim 1, characterized in that: High-resolution remote sensing image feature extraction: In the Hadoop cluster Map phase, feature vectors are extracted from homogeneous areas extracted by Selective Search; In this stage, SIFT features, Zernike features, GLCM features, and depth features are extracted and stored in HDFS. Each rectangular homogeneous region corresponds to a .txt feature file. The feature files are named as follows: type-name-xywh.txt, where type represents the name of the feature extraction method, name represents the name of the original image where the homogeneous region is located, and xywh represents the screen coordinates of the homogeneous region in the original full-scene remote sensing image and the length and width pixel values ​​of the homogeneous region, respectively. The SIFT feature dimension of each homogeneous region is n×128, the Zernike feature dimension is 1×656, the GLCM feature dimension is 1×5, and the depth feature uses the feature vector extracted by the fc7 layer in AlexNet, with a dimension of 1×4096. The four extracted features are used for weighted combination to find the best feature combination.

7. The high-resolution remote sensing image cloud computing optimized distributed retrieval method according to claim 1, characterized in that: Median Iterative Feature Clustering Based on MapReduce: A bag-of-words model for remote sensing image features is constructed through feature clustering to reduce the computational complexity of feature vector distance measurement during retrieval. SIFT feature files in HDFS are clustered using median iteration. The cluster center value is set to 256, and a data dictionary with 256 "words" is constructed. Each "word" has a dimension of 1×128. For each homogeneous region, the distance between the corresponding feature file and the cluster center is measured. The frequency of feature vectors falling within each cluster center is calculated to form a "word frequency histogram" of the feature vectors, and a codebook file with a dimension of 1×256 is constructed. First, we compare the three clustering algorithms of median iteration, AP clustering, and spectral clustering, and adopt the median iteration algorithm with obvious speed advantage. We also implement the parallel median iteration based on MapReduce to establish feature clustering to construct the image feature library, and realize the rapid clustering of remote sensing image features to construct the bag-of-words model.

8. The high-resolution remote sensing image cloud computing optimized distributed retrieval method according to claim 1, characterized in that: Median Iterative Feature Clustering Based on MapReduce: Algorithm: Median Iterative Clustering Algorithm Based on MapReduce Input: File inputFile that stores feature vector data InputPath, the storage path for feature vector data on HDFS Initialize the storage path centerPath of the cluster center file on HDFS ClusterPath is the storage path of clustering results on HDFS. The number of cluster centers k Output: k classes Start: Read inputPath, randomly select k points as initial cluster center points, and write the k cluster center point information to centerPath While the clustering termination condition is not met In the Mapper stage, read the inputPath, for<key,value> The point corresponding to the key in the key-value pair traverses all cluster centers, calculates the distance, and updates the key value with the number of the nearest cluster center Pass the key value of the changed point to the Reducer In the Reducer stage, the key values ​​passed by the Mapper stage are received and merged, and the results are written to the clusterPath Read clusterPath, update the cluster center using the mean of all points within the cluster center range, write the result to centerPath, and repeat this step until all cluster centers are completely updated endwhile Finish The MapReduce architecture highly abstracts iterative problems into two processes, Map and Reduce, improving the efficiency of text data processing. It combines the median iterative hard-core iterative algorithm with an efficient parallel computing architecture to increase the clustering speed of massive feature vector data.

9. The high-resolution remote sensing image cloud computing optimized distributed retrieval method according to claim 1, characterized in that: Optimized construction of a feature library based on HBase: This architecture stores feature data, combined feature data, data dictionary, and codebook data in HBase. The data dictionary and codebook data are obtained by SIFT clustering and are denoted as BoVW. The data dictionary dimension is 1×128×256, where 256 indicates the number of "keywords" in the data dictionary. Each codebook data dimension is 1×256, indicating the frequency with which the feature vector in the corresponding SIFT file falls on the data dictionary "keyword". Based on the Zernike features stored in HBase, the gray-level co-occurrence moment GLCM features and the deep features extracted from the fc7 layer in AlexNet, denoted as Deep, there are two combination feature settings: (1) GLCM and BoVW combination, with the weights of 10000 and 1 respectively; (2) GLCM+BoVW+Deep combination, with the weights of 10000, 1, and 15 respectively; In HBase, a table is created for each rectangular homogeneous region. The table name is the homogeneous region name. In each table, the column field stores the feature name, and the value field stores the corresponding column category feature data extracted from the homogeneous region. In addition, the timestamp information in the table indicates the time when the data is operated.

10. The high-resolution remote sensing image cloud computing optimized distributed retrieval method according to claim 1, characterized in that: Median Iterative Feature Clustering Based on MapReduce: Remote Sensing Image Target Retrieval Using Hadoop-Based High-Resolution Remote Sensing Image Distributed Retrieval Architecture: Phase 1: Upload Dataset 2 to HDFS to build a remote sensing image library; Phase 2: Read high-resolution remote sensing images from HDFS and pass each image as a single task slice to a cluster node; Phase 3: In the Map phase, SelectiveSearch is used to extract homogeneous regions and the screen coordinates of the upper left point and the length and width pixel values ​​of each candidate region are recorded; Phase 4: In the Map phase, GLCM features, SIFT features, AlexNet-based fc7 features, and Zernike matrix features are extracted from the candidate regions. The feature files are named in the following format: type-name-xywh.txt. SIFT features are stored in HDFS, and the other three features are stored in HBase. Phase 5: Start a new MapReduce task and use median iteration to perform parallel clustering on SIFT features. Set the number of cluster centers to 256, each with a dimension of 1×128. Use the cluster centers as the visual data dictionary. Store the generated codebook application and data dictionary corresponding to each homogeneous region in HBase. At this step, the feature database and visual bag-of-words model are constructed. Stage 6: During retrieval, a single machine extracts four features of the image to be retrieved, compares the SIFT features with the cluster centers, and generates codebook data; Phase 7: Measure the similarity of features and feature combinations, return the similarity ranking of homogeneous regions, and draw the homogeneous regions on the original high-resolution remote sensing image and return it to the user UI.

Images