Remote sensing image processing method and device, electronic equipment and medium

Abstract

The invention discloses a remote sensing image processing method and device, electronic equipment and a medium. The method comprises the steps that an image processing request input by a user is acquired; analyzing the image processing request through a pre-constructed image processing large model to generate a task sequence and tool parameters; calling a target processing tool for processing the task sequence from a pre-constructed image processing tool library; setting corresponding tool parameters for the target processing tool; and controlling each target processing tool to execute the corresponding image processing task to obtain an image processing result. Therefore, through combination of automatic analysis of the task sequence, intelligent scheduling of the processing tool and automatic setting of parameters, the whole image processing process is intelligent, the use threshold of image processing is reduced, efficient and accurate image processing is realized, and timeliness and flexibility of image processing are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing, and in particular to a remote sensing image processing method and device, electronic equipment and a medium. BACKGROUND

[0002] In the rapid development of remote sensing technology, how to efficiently and accurately process remote sensing image data is crucial for the application of natural disaster detection, environmental change analysis and other technical fields. At present, remote sensing image processing mainly relies on professional desktop software systems. Such software is expensive and needs to be installed and run on the user's local computer, which not only has high requirements for hardware configuration (such as processor performance, memory capacity, storage space, etc.), but also is complex to operate, requiring users to have remote sensing professional knowledge and operation skills, resulting in a high threshold for image processing and limiting the application of remote sensing technology.

[0003] In addition, in application scenarios such as natural disaster monitoring, environmental change dynamic tracking, emergency assessment and real-time data fusion analysis that require rapid response and flexible processing, the local installation of software cannot achieve access and processing of remote sensing data anytime and anywhere, resulting in low timeliness of information acquisition.

[0004] Therefore, how to improve the timeliness, flexibility and processing efficiency of remote sensing image processing, and reduce the threshold and economic barriers of remote sensing data processing, is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] Therefore, one aspect of the present application provides a remote sensing image processing method, which comprises: obtaining an image processing request input by a user; analyzing the image processing request by a pre-constructed image processing large model to generate a task sequence and tool parameters; calling a target processing tool for processing the task sequence from a pre-constructed image processing tool library; and setting corresponding tool parameters for the target processing tool; controlling each target processing tool to execute a corresponding image processing task to obtain an image processing result.

[0006] Optionally, constructing the image processing large model comprises the following steps: pre-constructing a lightweight initial large model; obtaining a triple training sample; wherein the triple training sample comprises a task prompt word, a processing instruction and a processing tool sequence; fine-tuning the initial large model by the triple training sample to obtain the image processing large model.

[0007] Optionally, the control each target processing tool to execute the corresponding image processing task, comprising: Resolving the dependency between each image processing task in the task sequence; According to the dependency, a directed acyclic graph representing the image processing path is constructed; Based on the directed acyclic graph, the target processing tool is controlled to execute the corresponding image processing task.

[0008] Optionally, the control each target processing tool to execute the corresponding image processing task based on the directed acyclic graph, comprising: Get the current computing resources, and the pre-constructed resource image database; the resource image database includes the corresponding relationship between processing tools and resource images; Based on the resource image database, determine the target computing resources required by each directed acyclic graph; According to the current computing resources and the target computing resources, assign corresponding computing resources to each directed acyclic graph, and formulate an execution strategy for each directed acyclic graph; Based on the execution strategy, call the corresponding computing resources, and control the target processing tool to execute the corresponding image processing task.

[0009] Optionally, according to the current computing resources and the target computing resources, formulate an execution strategy for each directed acyclic graph, comprising: Different directed acyclic graphs, the same processing tool formulates a parallel execution strategy; Assign high bandwidth resources to input-output intensive tasks in the directed acyclic graph, and formulate a parallel execution strategy; For the calculation intensive task in the directed acyclic graph, allocate GPU resource pool, and reuse the GPU resource pool by LRU principle.

[0010] Optionally, the processing method of remote sensing image further comprises: Through the pre-constructed quality evaluation model, the quality score of the image processing result is obtained, and the quality score of the historical image processing data is obtained; Get the feedback information for representing the user's satisfaction with the image processing result; The historical image processing data is divided into first historical data and second historical data; wherein the first historical data is the data corresponding to the quality score greater than the threshold value and the feedback information representing satisfaction; the second historical data is the data corresponding to the quality score not greater than the threshold value and the feedback information representing dissatisfaction; Get the user input modification information about the second historical data; According to the first historical data, the second historical data, and the modification information, the image processing large model is fine-tuned.

[0011] Optionally, the fine-tuning training of the image processing large model according to the first historical data, the second historical data, and the modification information comprises: The historical processing requests, the historical task sequences, and the historical processing tools in the first historical data are constructed into first triple samples; According to the second historical data and the modification information, second triple samples are generated; The image processing large model is fine-tuned through the first triple samples and the second triple samples.

[0012] Another aspect of the present application provides a processing device for remote sensing images, the device comprising: A request acquisition module is configured to acquire an image processing request input by a user; A request analysis module is configured to analyze the image processing request by using a pre-constructed image processing large model to generate a task sequence and tool parameters; A processing tool acquisition module is configured to call target processing tools for processing the task sequence from a pre-constructed image processing tool library, and set corresponding tool parameters for the target processing tools; A tool control module is configured to control the target processing tools to execute corresponding image processing tasks to obtain image processing results.

[0013] Another aspect of the present application provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor executes the program to implement the steps of the processing method for remote sensing images.

[0014] Another aspect of the present application provides a computer readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to implement the steps of the processing method for remote sensing images.

[0015] The processing method, device, electronic device, and medium for remote sensing images provided by the present application have the following beneficial effects: based on a pre-constructed image processing large model, online analysis of an image task sequence is realized, and corresponding processing tools are called for online processing, so that a user does not need to master professional remote sensing knowledge or programming skills, and can complete complex image processing through simple processing instructions, thereby reducing the use threshold of image processing. In addition, the combination of automatic analysis of a task sequence, intelligent scheduling of processing tools, and automatic setting of parameters makes the entire image processing process intelligent, realizes efficient and accurate image processing, and guarantees the timeliness and flexibility of image processing. Attached Figure Description

[0016] Figure 1 A schematic flowchart illustrating a remote sensing image processing method provided in an embodiment of this application; Figure 2 A schematic diagram of the structure of a remote sensing image processing system provided in an embodiment of this application; Figure 3 A schematic diagram illustrating the principle of constructing a large image processing model provided in an embodiment of this application; Figure 4 A schematic diagram illustrating the principle of a remote sensing image processing method provided in an embodiment of this application; Figure 5 A schematic diagram of the structure of a remote sensing image processing apparatus provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0017] The reference numerals in the attached diagram are as follows: 50 is the request acquisition module, 51 is the request parsing module, 52 is the processing tool acquisition module, 53 is the tool control module, 60 is the memory, 61 is the processor, 62 is the display screen, 63 is the input / output interface, 64 is the communication interface, 65 is the power supply, 66 is the communication bus, 601 is the computer program, 602 is the operating system, and 603 is the data. Detailed Implementation

[0018] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0019] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0020] Figure 1 This is a schematic flowchart illustrating a remote sensing image processing method provided in an embodiment of this application, as shown below. Figure 1 As shown, the method includes: S10: Obtain the image processing request input by the user; Figure 2 This is a schematic diagram of the structure of a remote sensing image processing system provided in an embodiment of this application. In an optional embodiment, the remote sensing image processing method provided in this application can be applied to... Figure 2 The remote sensing image processing system shown is composed of Figure 2 As can be seen, the system includes: a user interface, an intelligent agent, an execution engine, and an intelligent scheduling module. In a specific embodiment, the user interface, the intelligent agent, the execution engine, and the intelligent scheduling module are sequentially connected for communication.

[0021] In a specific embodiment, such as Figure 2 As shown, the user interface receives the image processing request input by the user and displays the processing result corresponding to the current image processing request, i.e., it provides a human-computer dialogue window. It should be noted that the image processing request can be in the form of text, image, and voice, etc., and this application does not limit it.

[0022] In one alternative embodiment, the user can input an image processing request through a user interface using natural language descriptions. For example, the image processing request could be "Generate a map of vegetation cover changes in Changchun City over the past five years".

[0023] S11: Parse the image processing request using a pre-built large image processing model to generate task sequences and tool parameters; To achieve online processing of remote sensing images, such as Figure 2 As shown, the remote sensing image processing system provided in this application deploys an intelligent agent, which is the core processing unit of the system. It can call a pre-built large image processing model to parse the acquired image processing requests and generate corresponding task sequences and tool parameters of the tools to be called.

[0024] It is worth noting that for remote sensing images acquired by different satellite sensors, the tool parameters used differ when different processing tools are invoked for image processing. Therefore, in a specific embodiment, the large-scale image processing model can parse the image processing request to determine the satellite sensor corresponding to the current request and generate the tool parameters corresponding to the processing tools used for different image processing tasks in the task sequence. In an optional embodiment, if the satellite sensor cannot be determined from the current parsing results, default tool parameters can be used.

[0025] S12: Retrieve the target processing tool for processing the task sequence from the pre-built image processing tool library; and set the corresponding tool parameters for the target processing tool; Furthermore, such as Figure 2As shown, the agent transmits the generated task sequence and tool parameters to the execution engine. At this time, the execution engine calls the target processing tool that can be used to process the task sequence from the pre-constructed image processing tool library.

[0026] In specific embodiments, the task sequence can include multiple image processing tasks, which can include but are not limited to image retrieval tasks, atmospheric correction tasks, image segmentation tasks, and image target detection tasks. In order to improve image processing efficiency and accuracy, different processing tools can be called to execute tasks for different image processing tasks. Correspondingly, the processing tools can include but are not limited to image retrieval tools, atmospheric correction tools, image segmentation tools, and image target detection tools.

[0027] After calling the target processing tool, the corresponding parameters of the target processing tool need to be set, i.e., the parameters of different target processing tools are set to the tool parameters generated by the image processing large model, thereby ensuring the image processing effect.

[0028] S13: Control each target processing tool to execute the corresponding image processing task to obtain an image processing result.

[0029] Further, as shown in Figure 1 and Figure 2 , control each target processing tool to execute the corresponding image processing task to obtain an image processing result. After obtaining the image processing result, it can be fed back to the user through the user interface as shown in Figure 2

[0030] Thus, the processing method for remote sensing images provided by the embodiments of the present application realizes online analysis of image task sequences based on pre-constructed image processing large models, and calls corresponding processing tools for online processing. Users do not need to master professional remote sensing knowledge or programming skills, and can complete complex image processing through simple processing instructions, thereby reducing the use threshold of image processing. In addition, the combination of automatic analysis of task sequences, intelligent scheduling of processing tools, and automatic setting of parameters makes the entire image processing process intelligent, realizes efficient and accurate image processing, and guarantees the timeliness and flexibility of image processing.

[0031] Figure 3 A principle diagram for constructing an image processing large model provided by an embodiment of the present application, in an optional embodiment, constructing an image processing large model includes the following steps: Pre-construct a lightweight initial large model; Obtain a triple training sample; wherein the triple training sample includes a task prompt word, a processing instruction, and a processing tool sequence; Through the triple training sample, fine-tune the initial large model to obtain an image processing large model.​

[0032] It can be understood that, in specific embodiments, the image processing large model is used to analyze the image processing request input by the user and generate a corresponding task sequence. In order to improve the analysis accuracy of the image processing large model on the request, thereby improving the accuracy of online image processing.

[0033] In an optional embodiment, as shown in Figure 3 , a lightweight initial large model with a parameter amount of 7B level is constructed in advance. The initial large model can be Qwen-7B or DeepSeek-R1 7B. The selection of the initial large model is not limited in the present application.

[0034] At the same time, source data about remote sensing images is obtained. The source data can include but is not limited to open source image databases, image papers, image software manuals and image expert manuals. Based on the source data, a triple training sample used to train the initial large model is constructed.

[0035] Since the final image processing large model is used to analyze the request and generate a task sequence, the triple training sample includes a task prompt word, a processing instruction and a processing tool sequence. The task prompt word is used to represent the current image processing request, the processing instruction is the specific processing instruction corresponding to the image processing request included in the task prompt word, and the processing tool sequence is the combination of processing tools called by different processing instructions. In order to facilitate understanding, examples will be given below.

[0036] For example, the triple training sample is (generate NDVI map, calculate normalized vegetation index, call atmospheric correction tool, band calculation tool), the task prompt word is to generate an NDVI (Normalized Difference Vegetation Index) map, the processing instruction is to calculate NDV, and the processing tool sequence includes an atmospheric correction tool and a band calculation tool. NDVI is an index used to evaluate vegetation coverage and growth conditions using remote sensing image data. It quantifies the "greenness" or health of vegetation by calculating the difference in reflectance between the near-infrared band (NIR) and the red band (Red).

[0037] Further, as shown in Figure 3 , the initial large model is supervised fine-tuning (SFT) trained through the triple training sample, thereby obtaining an image processing large model that can accurately generate a task sequence. In an optional embodiment, in order to further improve the accuracy of the image processing large model, on the basis of the above embodiment, as shown in Figure 3As shown, a triple test sample can also be constructed according to the source data. The precision of the current trained image large model is detected through the triple test sample. If the precision reaches the expectation, the image large model is directly applied to Figure 2 As shown in the intelligent agent, if the expectation is not reached, the triple training sample is used to continue iterative fine-tuning training of the model.

[0038] It should be noted that in an optional embodiment, in the test of the triple test sample, only at least one of the task prompt word and the processing instruction is input into the image processing large model, and after the image processing large model generates the processing tool sequence, the generated processing tool sequence and the tool sequence in the triple test sample are transmitted to the quality evaluation large model for quality evaluation, so as to determine whether the image processing large model meets the expectation according to the quality score output by the quality evaluation large model.

[0039] It can be understood that in addition to the task sequence, the image processing large model also needs to generate tool parameters. Therefore, in another optional embodiment, a parameter training data set is constructed, which includes the correspondence between the tool parameters and the satellite sensors. The image processing large model is trained through the parameter training data set, so that the large model learns the tool parameters corresponding to the settings of different satellite sensors.

[0040] Therefore, the processing method of the remote sensing image provided by the present application not only provides automation, but also provides intelligent process planning and parameter suggestion based on prior knowledge, which improves the accuracy and reliability of the image processing result. In addition, the use of lightweight large models reduces the training and inference cost, and expands the commercial application prospect.

[0041] In an optional embodiment, the control of each target processing tool to execute the corresponding image processing task includes: analyzing the dependency relationship between each image processing task in the task sequence; constructing a directed acyclic graph representing an image processing path according to the dependency relationship; controlling the target processing tool to execute the corresponding image processing task based on the directed acyclic graph.

[0042] It can be understood that in specific embodiments, multiple task sequences can be generated for different image processing requests, and multiple task sequences can also be generated for the same image processing request. There is an execution order between different task sequences, that is, different image processing paths can be constructed in different task sequences.

[0043] In an alternative embodiment, in order to improve the processing efficiency, the dependency relationship between different image processing tasks in the task sequence is analyzed, and based on the dependency relationship, a directed acyclic graph (DAG) representing the image processing path is constructed, wherein the DAG is a directed acyclic graph structure composed of a group of vertices (nodes) and directed edges (arrows), and any path will not return to itself (i.e. there is no loop). In order to facilitate understanding, examples will be given below.

[0044] For example, for target recognition and target counting, the target needs to be recognized first, and then the target counting can be performed, so in this embodiment, the execution order of the target recognition tool is prior to the execution order of the target technology tool.

[0045] Further, after obtaining the DAG execution path, the target processing tool corresponding to each node in the DAG is controlled to perform the corresponding image processing task.

[0046] Figure 4 The principle diagram of the remote sensing image processing method provided by the embodiments of the present application is based on the above embodiments, and in an alternative embodiment, based on the directed acyclic graph, the target processing tool is controlled to perform the corresponding image processing task, comprising: Obtain the current computing resource and the pre-constructed resource portrait database; the resource portrait database includes the corresponding relationship between the processing tool and the resource portrait; Based on the resource portrait database, determine the target computing resource required by each directed acyclic graph; According to the current computing resource and the target computing resource, assign corresponding computing resources to each directed acyclic graph, and formulate an execution strategy for each directed acyclic graph; Based on the execution strategy, call the corresponding computing resource, and control the target processing tool to perform the corresponding image processing task.

[0047] In order to further improve the online image processing efficiency and save computing resources, in addition to the path planning of the processing task (i.e. constructing the DAG), in an alternative embodiment, different DAGs also need to be assigned corresponding computing resources.

[0048] Specifically, as shown in Figure 4 In a specific embodiment, the intelligent scheduling module obtains the task sequence, first constructs the DAG according to the dependency relationship between different image processing tasks in the task sequence, and further determines the target computing resource required by different DAGs based on the pre-constructed resource portrait database.

[0049] Meanwhile, current computing resources available in the current system are acquired, and at this time, according to the available current computing resources and the target computing resources, corresponding computing resources can be allocated to different DAGs, and meanwhile, corresponding execution strategies are formulated for each DAG.

[0050] In an optional embodiment, when the resource profile database is constructed, a performance profiling tool can be set, and each image processing tool is tested and analyzed offline, and the resource usage information of each image processing tool, such as average CPU occupation, memory consumption, whether GPU is used, running time, etc., is recorded, and meanwhile, the computing intensity score of each processing tool is calculated based on the above resource usage information, which is used to represent the resource usage of the processing tool, and the higher the computing intensity score is, the greater the resource usage of the corresponding processing tool is.

[0051] Further, the correspondence between the computing intensity score, the processing tool and the resource profile is stored in the resource profile database. The resource profile includes but is not limited to the computing type, the estimated time consumption, the memory requirement index and the running tool context. In an optional embodiment, the scheduling can adopt secondary development of an open source workflow engine (such as Apache Airflow) so that it can query the resource profile database when scheduling.

[0052] Correspondingly, in an optional embodiment, based on the resource profile database, when the target computing resources required by each DAG are determined, the processing tools included in the DAG can be determined first, and the computing intensity score of the corresponding processing tool is acquired from the resource profile database, and further, the target computing density score of the current DAG is calculated according to the computing intensity score of each processing tool.

[0053] Therefore, according to the target computing density score of different DAGs and the available current computing resources, resource allocation and execution strategy formulation are performed for different DAGs.

[0054] As shown in Figure 4 , after the intelligent scheduling module constructs the DAG, allocates computing resources to the DAG and formulates the execution strategy, it is transmitted to the execution engine. At this time, the execution engine calls the corresponding computing resources based on the execution strategy, and controls the target processing tool to execute the corresponding image processing task.

[0055] On the basis of the above embodiment, as an optional embodiment, according to the current computing resources and the target computing resources, the execution strategy is formulated for each directed acyclic graph, which includes: The same processing tool is formulated and parallel execution strategy between different directed acyclic graphs; The input-output intensive task in the directed acyclic graph is allocated with high bandwidth resources, and parallel execution strategy is formulated; The GPU resource pool is allocated to the calculation-intensive task in the directed acyclic graph, and the LRU principle is used to reuse the GPU resource pool.

[0056] It can be understood that, in order to ensure the efficiency of online image processing, tasks that can be performed simultaneously can be formulated and parallel execution strategies can be formulated. For example, for different DAGs that include the same processing tools, that is, include the same processing tasks, the same processing tools between different DAGs can be controlled to perform tasks in parallel. In order to facilitate understanding, examples will be given below.

[0057] For example, image download and image segmentation are included in DAG1, and image download and image target recognition are included in DAG2. Therefore, when performing tasks, the image download tool in DAG1 and the image download tool in DAG2 can be controlled to perform tasks in parallel.

[0058] In another optional embodiment, for input / output (I / O) intensive tasks, high-bandwidth resources can be allocated, and parallel execution tasks can be formulated for I / O intensive tasks. Specifically, CPU / GPU is often idle, waiting for the disk, network, database, API to send data or send away, for example, as shown in Figure 4 The data download in step T1 is an I / O intensive task.

[0059] In still another optional embodiment, for calculation-intensive tasks, a GPU resource pool can be allocated, and the LRU principle (Least Recently Used) can be used to reuse the GPU resource pool, for example, as shown in Figure 4 Atmospheric correction and NDVI calculation both need to allocate a GPU resource pool. Specifically, CPU / GPU is fully loaded, and memory / disk / network is mostly idle, for example, tasks that use deep learning models.

[0060] The LRU principle is a cache eviction algorithm and resource management strategy. The core idea is that if data has been accessed recently, it is more likely to be accessed in the future; otherwise, the least used data is most likely to be discarded. It should be noted that the execution strategy is not limited in the present application, and other parallel strategies can also be included, for example, Figure 4 Data download and atmospheric correction can be performed. In order to facilitate understanding, the following will be described in conjunction with Figure 4 .

[0061] As shown in Figure 4As shown, in one optional embodiment, the task sequence includes data download, atmospheric correction, image cropping, and NDVI calculation. After processing by the intelligent scheduling module, data download is an I / O-intensive task, which can be allocated high-bandwidth resources. Atmospheric correction is a computationally intensive task, which is allocated a GPU resource pool. After data download, step T2 can be executed for image cropping, and CPU resources are allocated for image cropping. Furthermore, the GPU resource pool is reused using the LRU principle; therefore, after atmospheric correction, the GPU resource pool is reused to execute step T3 for NDVI calculation. It should be noted that in a specific embodiment, if other tasks that require the use of the GPU resource pool are also included, the GPU resource pool can be reused once based on the dependencies between tasks after step T3 NDVI calculation.

[0062] Therefore, the remote sensing image processing method provided in this application embodiment realizes refined management of computing resources and parallel execution of tasks based on the intelligent scheduling module, solves the computing power bottleneck of massive remote sensing data, shortens waiting time, and improves the processing efficiency of remote sensing images.

[0063] In an optional embodiment, the remote sensing image processing method provided in this application further includes: By using a pre-built quality assessment model, the quality score of the image processing results is obtained, thus yielding the quality score of the historical image processing data. Obtain feedback information that characterizes user satisfaction with image processing results; The historical image processing data is divided into first historical data and second historical data. The first historical data consists of data with a quality score greater than a threshold and feedback information indicating satisfaction. The second historical data consists of data with a quality score less than a threshold and feedback information indicating dissatisfaction. Obtain user-inputted modification information regarding the second historical data; Based on the first historical data, the second historical data, and the modification information, the large-scale image processing model is fine-tuned and trained.

[0064] like Figure 2 As shown, based on the above embodiments, the remote sensing image processing system provided in this application also includes a feedback learning module, which is used to collect information such as user feedback to continuously optimize the image processing model and improve the image processing accuracy.

[0065] Specifically, in one alternative embodiment, such as Figure 2 As shown, a quality assessment model can be pre-built to score the quality of each generated image processing result, thus obtaining a quality score for the historical image processing data. A higher quality score indicates better historical image processing performance.

[0066] In addition, to further improve the accuracy of large-scale image processing models, such as Figure 2 As shown, the feedback learning module can also obtain user feedback information through the user interface. In an optional embodiment, the feedback information can be a tiered evaluation, including at least two levels: "satisfied" and "dissatisfied".

[0067] Furthermore, the feedback learning module divides historical image data into first historical data and second historical data. The first historical data is the data whose quality meets expectations and can be directly used as a fine-tuning dataset to fine-tune and train the large image processing model.

[0068] If the quality of the second historical dataset does not meet expectations, an interactive mechanism can be triggered to correct it. Specifically, user feedback on modifications to the second historical dataset can be collected; for example, users can reply to the user interface with "Please describe where modifications are needed." After obtaining this feedback, the dataset can be combined with the second historical dataset to generate a dataset that can be used to fine-tune the large image processing model.

[0069] Specifically, as an optional implementation, the image processing large model is fine-tuned and trained based on the first historical data, the second historical data, and the modification information, including: The historical processing requests, historical task sequences, and historical processing tools in the first historical data constitute the first triplet sample; Generate a second triplet sample based on the second historical data and modification information; The large image processing model is fine-tuned and trained using the first and second triplet samples.

[0070] Understandably, the first set of historical data is high-quality data. Therefore, historical processing requests, historical task sequences, and historical processing tools can be extracted from it to form the first triplet sample. Simultaneously, for the lower-quality second set of historical data, the acquired modification information is combined to generate the second triplet sample. Understandably, the second triplet sample also includes the three components: historical processing requests, historical task sequences, and historical processing tools.

[0071] Finally, the large-scale image processing model is fine-tuned and trained using the first and second triplet samples, resulting in increasingly higher accuracy in generating task sequences.

[0072] Therefore, the remote sensing image processing method provided in this application, with its dual-dimensional feedback and differentiated processing mechanism, enables the system to continuously learn from user interaction, becoming more intelligent with use, thus forming a technological barrier and ecological advantage.

[0073] In the above embodiments, the method for processing remote sensing images has been described in detail. This application also provides an embodiment of a remote sensing image processing apparatus.

[0074] Figure 5 This is a schematic diagram of the structure of a remote sensing image processing apparatus provided in an embodiment of this application, as shown below. Figure 5 As shown, the device includes: The request acquisition module 50 is used to acquire the image processing request input by the user; The request parsing module 51 is used to parse the image processing request using a pre-built image processing big model to generate task sequences and tool parameters; The processing tool acquisition module 52 is used to retrieve the target processing tool for processing the task sequence from a pre-built image processing tool library; and to set the corresponding tool parameters for the target processing tool. The tool control module 53 is used to control each target processing tool to execute the corresponding image processing task and obtain the image processing result.

[0075] Furthermore, the remote sensing image processing apparatus provided in this application embodiment includes: The initial model building module is used to pre-build lightweight initial large models; The training sample acquisition module is used to acquire triplet training samples; wherein, the triplet training samples include task prompts, processing instructions and processing tool sequences; The first fine-tuning training module is used to fine-tune the initial large model using triplet training samples to obtain the large image processing model.

[0076] The dependency resolution module is used to resolve the dependencies between image processing tasks in a task sequence; The Directed Acyclic Graph (DAG) construction module is used to construct a DAG representing the image processing path based on dependencies. The first task execution module is used to control the target processing tool to execute the corresponding image processing task based on the directed acyclic graph.

[0077] The target acquisition module is used to acquire current computing resources and a pre-built resource profile database; the resource profile database includes the correspondence between processing tools and resource profiles. The target computing resource determination module is used to determine the target computing resources required for each directed acyclic graph based on the resource profile database. The target processing module is used to allocate corresponding computing resources to each directed acyclic graph based on the current computing resources and the target computing resources, and to formulate execution strategies for each directed acyclic graph. The second execution module is used to invoke corresponding computing resources based on the execution strategy and control the target processing tool to execute the corresponding image processing tasks.

[0078] The first strategy formulation module is used to formulate parallel execution strategies for the same processing tools among different directed acyclic graphs; The second strategy formulation module is used to allocate high-bandwidth resources for input-output intensive tasks in a directed acyclic graph and formulate parallel execution strategies. The third strategy formulation module is used to allocate GPU resource pools for computationally intensive tasks in a directed acyclic graph and reuse GPU resource pools using the LRU principle.

[0079] The quality scoring module is used to score the quality of image processing results using a pre-built quality assessment model, and obtain the quality score of historical image processing data. The feedback information acquisition module is used to acquire feedback information that characterizes the user's satisfaction with the image processing results; The data segmentation module is used to divide historical image processing data into first historical data and second historical data. The first historical data consists of data with a quality score greater than a threshold and feedback information indicating satisfaction. The second historical data consists of data with a quality score less than a threshold and feedback information indicating dissatisfaction. The modification information acquisition module is used to acquire modification information about the second historical data input by the user; The second fine-tuning training module is used to fine-tune the large image processing model based on the first historical data, the second historical data, and the modification information.

[0080] The first sample construction module is used to construct the first triplet sample from the historical processing requests, historical task sequences and historical processing tools in the first historical data. The second sample generation module is used to generate a second triplet sample based on the second historical data and modification information. The third fine-tuning training module is used to fine-tune the large image processing model using the first triplet samples and the second triplet samples.

[0081] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, such as... Figure 6 As shown, the electronic device includes: a memory 60 for storing computer programs; The processor 61 is configured to execute a computer program to implement the steps of the remote sensing image processing method as described in the above embodiments.

[0082] The electronic devices provided in this embodiment may include, but are not limited to, laptops or desktop computers.

[0083] The processor 61 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 61 may be implemented using at least one of the following hardware forms: Digital Signal Processor (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). The processor 61 may also include a main processor and a coprocessor. The main processor, also known as the Central Processing Unit (CPU), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 61 may integrate a Graphics Processing Unit (GPU), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 61 may also include an Artificial Intelligence (AI) processor, which is used to handle computational operations related to machine learning.

[0084] The memory 60 may include one or more computer-readable storage media, which may be non-transitory. The memory 60 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In this embodiment, the memory 60 is used to store at least the following computer program 601, which, after being loaded and executed by the processor 61, is capable of implementing the relevant steps of the remote sensing image processing method disclosed in any of the foregoing embodiments. In addition, the resources stored in the memory 60 may also include an operating system 602 and data 603, and the storage method may be temporary or permanent storage. The operating system 602 may include Windows, Unix, Linux, etc. The data 603 may include, but is not limited to, the relevant data involved in the remote sensing image processing method.

[0085] In some embodiments, the electronic device may further include a display screen 62, an input / output interface 63, a communication interface 64, a power supply 65, and a communication bus 66.

[0086] Those skilled in the art will understand that Figure 6 The structures shown do not constitute a limitation on electronic devices and may include more or fewer components than those shown.

[0087] The electronic device provided in this application includes a memory and a processor. When the processor executes the program stored in the memory, it can implement the remote sensing image processing method in the above embodiments.

[0088] It should be noted that although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and components in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

Claims

1. A method for processing a remote sensing image, characterized in that, The method comprises: acquiring a user inputted image processing request; analyzing the image processing request by a pre-constructed image processing large model to generate a task sequence and tool parameters; from a pre-constructed image processing tool library, calling a target processing tool for processing the task sequence; and setting corresponding tool parameters for the target processing tool; controlling each target processing tool to execute corresponding image processing tasks to obtain an image processing result.

2. The method of claim 1, wherein, Constructing the image processing large model comprises the following steps: pre-constructing a lightweight initial large model; acquiring a triple training sample; wherein the triple training sample comprises a task prompt word, a processing instruction and a processing tool sequence; fine-tuning the initial large model by the triple training sample to obtain the image processing large model.

3. The method of claim 1, wherein the step of processing the remote sensing image is performed by a computer. The control of each target processing tool to execute corresponding image processing tasks comprises: analyzing the dependency relationship between each image processing task in the task sequence; constructing a directed acyclic graph representing an image processing path according to the dependency relationship; controlling the target processing tool to execute corresponding image processing tasks based on the directed acyclic graph.

4. The method of claim 3, wherein the step of processing the remote sensing image is performed by a computer. The control of the target processing tool to execute corresponding image processing tasks based on the directed acyclic graph comprises: acquiring current computing resources and a pre-constructed resource portrait database; the resource portrait database comprises a corresponding relationship between processing tools and resource portraits; determining target computing resources required by each directed acyclic graph based on the resource portrait database; allocating corresponding computing resources to each directed acyclic graph and formulating an execution strategy for each directed acyclic graph according to the current computing resources and the target computing resources; based on the execution strategy, calling the corresponding computing resources and controlling the target processing tool to execute corresponding image processing tasks.

5. The method of claim 4, wherein the step of processing the remote sensing image is performed by a computer system. Formulating an execution strategy for each directed acyclic graph according to the current computing resources and the target computing resources comprises: different directed acyclic graphs have the same processing tool formulating a parallel execution strategy; allocating high-bandwidth resources to input-output intensive tasks in the directed acyclic graph and formulating a parallel execution strategy; allocating a GPU resource pool to compute-intensive tasks in the directed acyclic graph and reusing the GPU resource pool according to the LRU principle.

6. The method of claim 1, wherein the step of processing the remote sensing image is performed by a computer system. The method further comprises: scoring the quality of the image processing result by a pre-constructed quality evaluation model to obtain the quality score of historical image processing data; acquiring feedback information representing the user's satisfaction with the image processing result; dividing the historical image processing data into first historical data and second historical data; wherein the first historical data corresponds to data whose quality score is greater than a threshold value and whose feedback information represents satisfaction; and the second historical data corresponds to data whose quality score is not greater than the threshold value and whose feedback information represents dissatisfaction; acquiring user inputted modification information about the second historical data; The image processing large model is fine-tuned according to the first historical data, the second historical data, and the modification information.

7. The method of claim 6, wherein the step of processing the remote sensing image is performed by a computer system. The fine-tuning training of the image processing large model according to the first historical data, the second historical data, and the modification information comprises: constructing a first triple sample of a historical processing request, a historical task sequence, and a historical processing tool in the first historical data; generating a second triple sample according to the second historical data and the modification information; fine-tuning the image processing large model through the first triple sample and the second triple sample.

8. A processing device for remote sensing imagery, characterized in that, The device comprises: a request acquisition module configured to acquire an image processing request input by a user; a request analysis module configured to analyze the image processing request by using a pre-constructed image processing large model to generate a task sequence and tool parameters; a processing tool acquisition module configured to call a target processing tool for processing the task sequence from a pre-constructed image processing tool library, and set corresponding tool parameters for the target processing tool; a tool control module configured to control each target processing tool to execute a corresponding image processing task to obtain an image processing result.

9. An electronic device comprising a memory and a processor, said memory having stored thereon a computer program operable to run on the processor, characterized in that, The processor executes the program to implement the steps of the remote sensing image processing method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps of the remote sensing image processing method of any one of claims 1 to 7.

Images