An online recommendation method for remote sensing image processing algorithms

Abstract

The present invention discloses an online recommendation method for remote sensing image processing algorithms, which belongs to the field of remote sensing image processing. The evaluation method includes: constructing a standard data set required for algorithm evaluation; receiving a remote sensing image processing algorithm and a description file uploaded by a user; parsing the description file of the remote sensing image processing algorithm; matching a set of corresponding standard input data according to the algorithm type, using each standard input data as an input parameter to drive the remote sensing image processing algorithm to execute once; recording resource usage information, algorithm execution time, and algorithm output parameters, and calculating the algorithm accuracy; performing weighted summation on the resource usage information, algorithm execution time, and algorithm accuracy obtained from previous algorithm executions to obtain a comprehensive score for the algorithm. Based on the evaluation, the recommendation method constructs an algorithm optimization deep learning model and recommends the most suitable algorithm to improve image processing efficiency.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing image processing, and in particular relates to an online recommendation method for remote sensing image processing algorithms. Background Art

[0002] With the development of fields like commercial aerospace, the number of remote sensing satellites in orbit has increased, and the quality of remote sensing data has also improved. All-weather, multi-angle, multi-temporal, multi-source, and multi-resolution Earth observation has become a reality. Remote sensing data has been widely applied in a variety of industries, including military, environmental, land, national defense, agriculture, forestry, water conservancy, transportation, and disaster reduction, promoting the rapid development of the national economy.

[0003] The effectiveness of satellite remote sensing systems in various fields strongly depends on the performance of remote sensing algorithms. Due to the vast volume of remote sensing data, manual interpretation alone is inefficient. The application of remote sensing data in various industries requires the use of a wide variety of complex remote sensing image processing algorithms. The performance of remote sensing image processing algorithms directly impacts the effectiveness of their applications. Numerous studies have shown that different remote sensing algorithms have varying processing effects on different data types. In actual data processing, algorithms must be optimized based on the specific image processing task.

[0004] At present, due to the lack of a scientific evaluation system for remote sensing image processing algorithms, the evaluation of remote sensing image processing algorithms faces problems such as lack of objectivity and lack of unified standards. As a result, there is also a lack of algorithm optimization models, and it is impossible to optimize algorithms according to different image processing tasks. Summary of the Invention

[0005] In light of this, this paper proposes an online recommendation method for remote sensing image processing algorithms. This method categorizes remote sensing image processing algorithms and constructs a standardized evaluation dataset, which can improve the scientificity, objectivity, and accuracy of remote sensing image processing algorithm evaluation. Furthermore, a deep learning model for algorithm optimization is constructed, which can recommend optimal algorithms based on different data processing tasks, thereby improving processing performance.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] An online recommendation method for remote sensing image processing algorithms, running on a server, includes the following steps:

[0008] Step 1: Build an algorithm optimization deep learning model; the input data of the algorithm optimization deep learning model is the evaluation data set and the algorithm description file, and the output data is the algorithm evaluation result;

[0009] Step 2: Construct a training dataset. The training dataset is divided into multiple subsets. Each subset corresponds to a remote sensing image processing algorithm. Each training data set includes the algorithm description file of the remote sensing image processing algorithm corresponding to the subset, a set of images to be processed, and a comprehensive score. The comprehensive score is obtained as follows:

[0010] Step 201: input an image to be processed into a remote sensing image processing algorithm, and drive the remote sensing image processing algorithm to execute once;

[0011] Step 202, repeating step 201 until all the images to be processed in the training data are exhausted; recording resource usage information, algorithm execution time, and algorithm output parameters during each algorithm execution; wherein the resource usage information includes CPU utilization, memory utilization, network bandwidth utilization, and hard disk utilization;

[0012] Step 203, comparing the difference between the output parameters in each execution and the corresponding standard result data, and calculating the algorithm accuracy of the execution;

[0013] Step 204: performing a weighted summation of resource usage information, algorithm execution time, and algorithm accuracy obtained from previous algorithm executions to obtain a comprehensive score for the algorithm;

[0014] Step 3: Use the training data set to train the algorithm optimization deep learning model, and deploy the trained algorithm optimization deep learning model on the server; the specific method is:

[0015] Step 301, set the loss function to comprehensive score The mean square error of: , where || || represents the 2-norm, Indicates the evaluation results of the algorithm's optimal deep learning model output;

[0016] Step 302: During each training, the evaluation data set and the algorithm description file are input into the model respectively, and the evaluation result error is used to perform reverse gradient transfer on the model to obtain the model parameters;

[0017] Step 303: Repeat the training process until the loss function is less than 5%. When , stop training;

[0018] Step 4: Receive an image to be processed uploaded by the user and algorithm description files of multiple optional remote sensing image processing algorithms, and use the algorithm optimization deep learning model to make algorithm optimization recommendations.

[0019] Furthermore, in step 1, the algorithm preferably selects a deep learning model that uses the VGG16 model and the CBOW model, wherein the VGG16 model is responsible for processing the input image data, and the CBOW model is responsible for processing the input algorithm description file; the results of the VGG16 model and the CBOW model are regressed to the output layer after multiple layers of convolution.

[0020] Furthermore, the specific method of step 4 is:

[0021] Step 401: Input the image to be processed and multiple algorithm description files into the algorithm optimization deep learning model;

[0022] Step 402: Outputting a comprehensive score of each algorithm for the current image to be processed by the algorithm-optimized deep learning model;

[0023] Step 403: recommend the algorithm with the highest comprehensive score to the customer.

[0024] The present invention has the following beneficial effects:

[0025] 1. This invention builds an algorithm classification system, adopts different evaluation strategies based on the characteristics of different types of algorithms, and achieves a more scientific evaluation;

[0026] 2. The present invention constructs an algorithm optimization deep learning model, which can recommend the optimal algorithm according to different image processing tasks and improve processing efficiency.

[0027] In summary, the present invention classifies remote sensing image processing algorithms into categories, constructs a standard evaluation data set, and constructs an algorithm optimization deep learning model based on a scientific, objective, and accurate evaluation of remote sensing image processing algorithms. It can recommend optimal algorithms for different image processing tasks and improve processing efficiency. DETAILED DESCRIPTION

[0028] The present invention is further described in detail below with reference to specific embodiments.

[0029] An online evaluation method for a remote sensing image processing algorithm is provided, which runs on a server and includes the following steps:

[0030] Step 1: Construct the standard dataset required for algorithm evaluation. The standard dataset consists of standard input data and standard result data. The standard dataset is managed in a classified manner and contains standard data for different types of remote sensing image processing algorithms.

[0031] Step 2: Receive the remote sensing image processing algorithm and description file uploaded by the user. Specifically, the remote sensing image processing algorithm and description file must meet the following rules:

[0032] (1) Algorithm form: The algorithm should be an executable file in the form of EXE, DLL, JAR, PYTHON script file, or MATLAB script file;

[0033] (2) Algorithm type: divided into four types: single value calculation, distributed intensity calculation, binary calculation, and enumerable multivariate calculation. Among them, single value calculation refers to the algorithm that performs single value statistical calculation on a certain type of information in the remote sensing image, such as the area of ​​water in the image; distributed intensity calculation is an algorithm that performs pixel-by-pixel intensity calculation based on a certain feature in the remote sensing image, such as the shading index of the remote sensing image; binary calculation refers to the algorithm that performs binary division calculation on the remote sensing image, such as the extraction of roads in the image; enumerable multivariate calculation refers to the algorithm that performs multivariate division calculation on the remote sensing image, such as the ground feature classification algorithm.

[0034] (3) Algorithm input and output parameters: The input parameters and output parameters are divided into three major types: single-band images, multi-band images, and numerical values. Among them, single-band images are divided into two types: single-band binary images and single-band intensity images, and multi-band images are divided into two types: multi-band binary images and multi-band intensity images.

[0035] (4) Algorithm description file: It is an XML file with two main categories of content: mandatory elements and auxiliary elements. The mandatory elements include the algorithm form, algorithm type, algorithm input parameters, and algorithm output parameters; the auxiliary elements include the algorithm version, algorithm author, algorithm default parameters, and algorithm description.

[0036] In specific implementation, a user Web interaction interface can be provided for the client based on the BS architecture. The user uploads the remote sensing image processing algorithm through human-computer interaction, and the system transmits the remote sensing image algorithm from the client to the server.

[0037] Standard datasets consist of two sets of data: standard input data and standard result data, which are in a one-to-one correspondence. Standard input data primarily consists of single-band and multi-band images, categorized according to the input parameter classification scheme. Standard result data is categorized according to the output parameter classification scheme. The specific number of standard datasets can be constructed based on the application scenario, such as standard data for vegetation indices and water body indices.

[0038] Step 3: Perform format verification and integrity verification on the remote sensing image processing algorithm, and execute subsequent processes after passing the verification;

[0039] Step 4: parse the description file of the remote sensing image processing algorithm and analyze its algorithm type and required input parameters;

[0040] Step 5: Match a set of corresponding standard input data according to the algorithm type, use each standard input data as an input parameter, and drive the remote sensing image processing algorithm to execute once until all the matched standard input data are used up;

[0041] Step 6: Record resource usage information, algorithm execution time, and algorithm output parameters during each algorithm execution; resource usage information includes CPU utilization, memory utilization, hard disk utilization, and network bandwidth utilization;

[0042] Step 7: Compare the difference between the output parameters of each execution and the corresponding standard result data to calculate the algorithm accuracy of that execution; specifically:

[0043] (1) Single numerical calculation: directly compare the numerical results of the algorithm to be evaluated Standard results corresponding to standard data ,by As a measure of algorithm accuracy;

[0044] (2) Distributed intensity calculation type: The output of this type of algorithm is a single-band intensity image;

[0045] (3) Binary calculation type: The result of this type of algorithm is a single-band binary image;

[0046] (4) Enumerable multivariate computation class: The results of this type of algorithm are multiple binary images or an integer-valued image, and the two results are essentially equivalent.

[0047] Step 8: The resource usage information, algorithm execution time, and algorithm accuracy obtained from all previous algorithm executions are weighted and summed to obtain a comprehensive score for the algorithm. Specifically, resource usage is collected directly from system information, and the collected parameters include CPU utilization, memory utilization, hard disk utilization, and network bandwidth utilization, all in percentages.

[0048] The resource usage score is calculated as (1-u)×100, where u is the corresponding resource utilization; the scores of CPU utilization, memory utilization, hard disk utilization, and network bandwidth utilization are 、 、 、 ;

[0049] The algorithm execution time t is in seconds, and its score With data size (Unit is GB), the specific formula is ;

[0050] The algorithm accuracy value is designed to be The number in the interval, its score Directly multiply the algorithm accuracy by 100;

[0051] Finally, calculate the comprehensive score: ,in, is the weight coefficient, the sum of which is 1, and can be set according to the actual situation. In particular, the algorithm accuracy coefficient The value should not be lower than 0.7.

[0052] This method constructs an algorithm classification system and adopts different evaluation strategies according to the characteristics of different types of algorithms, thus achieving a more scientific evaluation of remote sensing image processing algorithms.

[0053] Using the scoring method in the above evaluation method, an online recommendation method for remote sensing image processing algorithms can also be provided. The method runs on a server and includes the following steps:

[0054] Step 1: Construct an algorithm-optimized deep learning model. The algorithm-optimized deep learning model is a deep learning model that uses the VGG16 model and the CBOW model. The VGG16 model is responsible for processing the input image data, and the CBOW model is responsible for processing the input algorithm description file. The results of the VGG16 model and the CBOW model are regressed to the output layer after multiple layers of convolution.

[0055] Step 2: Construct a training dataset. The training dataset is divided into multiple subsets. Each subset corresponds to a remote sensing image processing algorithm. Each training data set includes the algorithm description file of the remote sensing image processing algorithm corresponding to the subset, a set of images to be processed, and a comprehensive score. The comprehensive score is obtained as follows:

[0056] Step 201: input an image to be processed into a remote sensing image processing algorithm, and drive the remote sensing image processing algorithm to execute once;

[0057] Step 202, repeating step 201 until all the images to be processed in the training data are exhausted; recording resource usage information, algorithm execution time, and algorithm output parameters during each algorithm execution; wherein the resource usage information includes CPU utilization, memory utilization, network bandwidth utilization, and hard disk utilization;

[0058] Step 203, comparing the difference between the output parameters in each execution and the corresponding standard result data, and calculating the algorithm accuracy of the execution;

[0059] Step 204: performing a weighted summation of resource usage information, algorithm execution time, and algorithm accuracy obtained from previous algorithm executions to obtain a comprehensive score for the algorithm;

[0060] Step 3: Use the training data set to train the algorithm optimization deep learning model, and deploy the trained algorithm optimization deep learning model on the server; the specific method is:

[0061] (1) Definition of model input and output data: In the training phase, the model input data is the evaluation dataset and algorithm description document, and the output data is the algorithm evaluation result. In the inference phase, the model input data is the dataset to be processed and the algorithm, and the output is the algorithm optimization result;

[0062] (2) Model composition: A deep learning model based on the VGG16 model and the CBOW set is used. The VGG16 model is responsible for processing the input image data, and the CBOW model is responsible for processing the input algorithm description document. The results of VGG16 and CBOW are regressed to the output layer after multi-layer convolution. The loss function is the mean square error of the algorithm score: , where || || represents the 2-norm, Indicates the evaluation results of the algorithm's optimal deep learning model output;

[0063] (3) Model training: Different algorithms are used to process the same image data to obtain training data consisting of images, algorithms, and evaluation results. The algorithm is used to optimize the deep learning model and regress the evaluation results. During each training, the evaluation data set and algorithm description file are input into the model respectively. The model is reversely transferred with the error of the evaluation result to obtain the model parameters. The training process is repeated until the loss function accounts for less than 5%, that is, When , stop training and get the final model parameters.

[0064] Step 4: Receive an image to be processed uploaded by the user and algorithm description files of multiple optional remote sensing image processing algorithms, and use the algorithm optimization deep learning model to make algorithm optimization recommendations; the specific method is as follows:

[0065] Step 401: Input the image to be processed and multiple algorithm description files into the algorithm optimization deep learning model;

[0066] Step 402: Outputting a comprehensive score of each algorithm for the current image to be processed by the algorithm-optimized deep learning model;

[0067] Step 403: recommend the algorithm with the highest comprehensive score to the customer.

[0068] This method uses a deep learning model based on the VGG16 model and the CBOW set to perform regression calculations on the algorithm evaluation results. It can automatically recommend optimal algorithms for different data processing tasks and improve processing efficiency.

[0069] In summary, this paper selects appropriate standard data for evaluation based on the algorithm type, employing different evaluation methods for different algorithm types to achieve scientific, objective, and accurate evaluation of remote sensing image processing algorithms. Based on this algorithm evaluation, a deep learning model for algorithm optimization is constructed to regress the algorithm evaluation results. This model recommends the most suitable algorithm for a given image processing task, thereby improving image processing efficiency.

Claims

1. A remote sensing image processing algorithm online recommendation method, characterized in that: Running on the server, including the following steps: Step 1: Build an algorithm optimization deep learning model; the input data of the algorithm optimization deep learning model is the evaluation data set and the algorithm description file, and the output data is the algorithm evaluation result; Step 2: Construct a training dataset. The training dataset is divided into multiple subsets. Each subset corresponds to a remote sensing image processing algorithm. Each training data set includes the algorithm description file of the remote sensing image processing algorithm corresponding to the subset, a set of images to be processed, and a comprehensive score. The comprehensive score is obtained as follows: Step 201: input an image to be processed into a remote sensing image processing algorithm, and drive the remote sensing image processing algorithm to execute once; Step 202, repeating step 201 until all the images to be processed in the training data are exhausted; recording resource usage information, algorithm execution time, and algorithm output parameters during each algorithm execution; wherein the resource usage information includes CPU utilization, memory utilization, network bandwidth utilization, and hard disk utilization; Step 203, comparing the difference between the output parameters in each execution and the corresponding standard result data, and calculating the algorithm accuracy of the execution; Step 204: performing a weighted summation of resource usage information, algorithm execution time, and algorithm accuracy obtained from previous algorithm executions to obtain a comprehensive score for the algorithm; Step 3: Use the training data set to train the algorithm optimization deep learning model, and deploy the trained algorithm optimization deep learning model on the server; the specific method is: Step 301, set the loss function to comprehensive score The mean square error is: , where || || represents the 2-norm, Indicates the evaluation results of the algorithm's optimal deep learning model output; Step 302: During each training, the evaluation data set and the algorithm description file are input into the model respectively, and the evaluation result error is used to perform reverse gradient transfer on the model to obtain the model parameters; Step 303: Repeat the training process until the loss function is less than 5%. When , stop training; Step 4: Receive an image to be processed uploaded by the user and algorithm description files of multiple optional remote sensing image processing algorithms, and use the algorithm optimization deep learning model to make algorithm optimization recommendations.

2. The online recommendation method for remote sensing image processing algorithm according to claim 1, characterized in that: In step 1, the algorithm preferably selects a deep learning model that uses the VGG16 model and the CBOW model, where the VGG16 model is responsible for processing the input image data, and the CBOW model is responsible for processing the input algorithm description file; the results of the VGG16 model and the CBOW model are regressed to the output layer after multi-layer convolution.

3. The online recommendation method for remote sensing image processing algorithm according to claim 1, characterized in that: The specific method of step 4 is: Step 401: Input the image to be processed and multiple algorithm description files into the algorithm optimization deep learning model; Step 402: Outputting a comprehensive score of each algorithm for the current image to be processed by the algorithm-optimized deep learning model; Step 403: recommend the algorithm with the highest comprehensive score to the customer.

Images