A GPU-based all-weather satellite autonomous monitoring system for sea and air space

Abstract

The present application relates to the technical field of satellite autonomous monitoring, and relates to a GPU-based all-weather real-time satellite autonomous monitoring system for sea and air space, which solves the problem of efficient monitoring in the absence of a real-time monitoring scheme for sea and air space; specifically comprising a target identification system for infrared and visible light image fusion on the satellite, a VGG16 network-based super-resolution reconstruction of the fused infrared and visible light images, and simultaneously obtaining a super-resolution target image; a target identification module for target identification according to the generated super-resolution image, cutting the target, packing the target with auxiliary data, and returning the target to the ground for analysis; a ground target classification and early warning system for analyzing the returned data packet, classifying the target image (military ship, aircraft carrier, airbus, fighter jet, etc.), identifying the target danger level, and determining whether to trigger an alarm. The system can achieve all-weather real-time autonomous monitoring, and the satellite data transmission data volume can be reduced by more than 80%, and the equipment maintenance cost is low.

Description

Technical Field

[0001] This invention relates to the field of satellite autonomous monitoring technology, and in particular to a GPU-based all-weather satellite autonomous monitoring system for sea and airspace. Background Technology

[0002] In the existing technology, the main methods for monitoring the sea and airspace rely on sea surface radar, acoustic and optical driven equipment, AIS (Automatic Identification System) and the like. All of these are deployed on the sea and land, with a relatively small monitoring range. Large-area monitoring requires the deployment of a wide range and a large number of devices, resulting in high maintenance costs. The entire software system is complex to build, and the system is easily affected by the quality of individual devices.

[0003] In addition, most existing satellites are used for monitoring land resources, coastlines, and forest fire prevention. Their main working method is to take concentrated pictures at a certain time, store the image data on the satellite, and transmit the images to the ground when the satellite reaches a certain position relative to the earth. Ground agencies then process the images and obtain analysis results. However, no more effective technical measures have been proposed for real-time monitoring of sea and airspace. Summary of the Invention

[0004] Regarding solutions for real-time monitoring of sea and airspace, this invention aims to address the problem that existing technologies lack efficient and wide-range monitoring measures, and proposes a GPU-based all-weather satellite autonomous monitoring system for sea and airspace.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] A GPU-based all-weather satellite autonomous monitoring system for sea and airspace includes:

[0007] The on-board super-resolution reconstruction target recognition system is based on a GPU system that can acquire a super-resolution image of target 1 and obtain the corresponding information of the super-resolution image through a recognition process.

[0008] Among them, target 1 refers to offensive vehicles such as ships and aircraft based on the sea surface and airspace;

[0009] The method for acquiring the super-resolution image of target 1 is as follows: the visible light image and infrared image of target 1 obtained by an image acquisition system are fused and reconstructed to complete the acquisition of the super-resolution image;

[0010] The ground-based data analysis and autonomous early warning system can receive the super-resolution image and corresponding information packaged and transmitted back by the target recognition system reconstructed by the on-board super-resolution system, and present it in the form of recognition information groups.

[0011] The satellite data analysis and autonomous early warning system can perform an analysis process on the identification information group to classify the target 1 and obtain the danger level of the target 1, and issue a corresponding early warning based on the danger level.

[0012] More specifically, the on-board super-resolution reconstruction target recognition system includes a super-resolution reconstruction system 10 and a target recognition system 20, and both can be loaded into a GPU system;

[0013] The super-resolution reconstruction system 10 and the target recognition system 20 have independent threads.

[0014] More specifically, the fusion and reconstruction process is as follows:

[0015] The super-resolution reconstruction system 10 is equipped with a VGG16 network, and obtains the super-resolution image of the target 1 through a super-resolution reconstruction process via the VGG16 network.

[0016] More specifically, the target recognition system 20 is capable of performing a target recognition process and can obtain the super-resolution image and corresponding information of the target 1 from the on-board super-resolution reconstruction system 10, package it in the form of the recognition information group, and transmit it back.

[0017] More specifically, the super-resolution reconstruction system 10 and the target recognition system 20 are located in an integrated network, which is a VGG16 network + YOLOv3 combined network architecture.

[0018] More specifically, the super-resolution reconstruction process is as follows:

[0019] The super-resolution reconstruction layer of the super-resolution reconstruction system 10 is embedded in a preset model for target 1 recognition based on YOLO v3, and the preset model is deployed to the GPU+ARM system. The GPU+ARM system loads the data of the preset model into the GPU system upon startup, and inputs the image into the GPU system according to the real-time acquired image.

[0020] The image in question is a super-resolution image.

[0021] More specifically, the target recognition process is as follows: the image is divided into blocks, which are then sequentially input into the recognition network, and target information is output.

[0022] The target information effect diagram test uses the DOTA dataset;

[0023] The identification results yielded integrated information about the target;

[0024] Subsequently, by cropping the image, the pixels of the cropped target 1 and the integrated information are extracted, and the recognition result is obtained based on a preset protocol;

[0025] The target recognition unit 20 is capable of recognizing multiple targets 1 in the super-resolution image based on the YOLO module to obtain the target image and information containing the image;

[0026] The identification results include: the target's location information, category information, similarity information, and latitude and longitude.

[0027] More specifically, the integrated network has a network training process, specifically as follows:

[0028] Freeze the YOLO model parameters, train a VGG16-based super-resolution reconstruction network to form the first network, and then freeze the super-resolution network parameters of the VGG16 network.

[0029] Next, a recognition network based on YOLO v3 is trained to form a second network. When training the YOLO network, the parameters of the last three layers of the YOLO network need to be frozen first, the parameters of the Darknet53 backbone network need to be trained, and finally the parameters of the last three layers of the network need to be unfrozen to perform unified training of the YOLO network.

[0030] Finally, all parameters of the first and second networks are unfrozen to obtain an integrated network through training;

[0031] The training of the preset model uses a pre-labeled dataset of visible light images and infrared images;

[0032] The preset model is deployed to a preset system with GPU and ARM.

[0033] More specifically, the process of cropping the image is as follows:

[0034] The image of target 1 is cropped, the latitude and longitude information of target 1 is calculated based on the cropping position, and the target similarity value output by target 1 is packaged together according to the preset protocol and transmitted back to the ground system.

[0035] More specifically, the parsing process is as follows:

[0036] By parsing the data according to a preset packaging format, the data is analyzed and used for real-time ground-based early warning monitoring.

[0037] The target classification system 30 performs a classification operation on the identification results, determines the target danger level based on the classification results, and determines whether an alarm should be sounded.

[0038] Specifically, it consists of a classification module and a warning information generation module;

[0039] The ground target classification system uses the VGG16 target classification model, which classifies the target 1 on the satellite to determine the type of the offensive vehicle and generate a corresponding alarm level signal.

[0040] The present invention has the following beneficial effects:

[0041] The system and method provided by this invention are mainly aimed at low-cost operation and monitoring of satellite payloads, deploying geostationary orbit satellites in a fixed area of ​​the sea, and are innovatively designed in three main aspects:

[0042] Firstly, it enables real-time monitoring of the area around the clock;

[0043] Secondly, this system actually generates less load, less data (reduced by 80%), and lower data transmission costs (reduced by 50%); specifically, this includes reduced power consumption of data transmission equipment, reduced thermal control costs, reduced power supply costs, and reduced satellite launch costs.

[0044] Thirdly, this system and method integrate the visible light and infrared super-resolution reconstruction network with the identification network into an integrated network. Compared with traditional sea and airspace monitoring systems, this system has advantages such as a wider monitoring area, higher efficiency, fewer deployed devices, and higher equipment stability. Attached Figure Description

[0045] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0046] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0047] Figure 2 This is a schematic diagram of the on-board system architecture of the present invention;

[0048] Figure 3 This is a schematic diagram of the ground data parsing software architecture of the present invention;

[0049] Figure 4 This is the interface of the ground target identification and early warning software of the present invention.

[0050] Figure 5 This is a schematic diagram of the target recognition results of the present invention;

[0051] Figure 6 This is a schematic diagram illustrating the target cropping, early warning information generation, and data packaging process of the present invention.

[0052] Figure 7 This is a schematic diagram of the system process of the present invention;

[0053] Figure 8 This is a schematic interface showing the results of an alarm embodiment of the test of the present invention;

[0054] Figure 9 This is a schematic diagram illustrating the identification process of the present invention to obtain information corresponding to the super-resolution image;

[0055] Figure 10 This is a schematic diagram of the fusion and reconstruction process of the present invention;

[0056] Figure 11 This is a schematic diagram of the multi-objective working process of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. It should be noted that, for ease of description, in this application, "left side" is referred to as "first end", "right side" as "second end", "upper side" as "first end", and "lower side" as "second end" in the current view. The purpose of such description is to clearly express the technical solution and should not be construed as an improper limitation of the technical solution of this application.

[0058] The specific meaning of the technology involved in this application is as follows:

[0059] (1) GPU, specifically known as Graphic Processing Unit, is a graphics processor;

[0060] (2) Arm, specifically named Advanced RISC Machines, is an advanced reduced instruction set processor;

[0061] (3)Yolo v3, specifically You Only Look Once, version 3.

[0062] (4) AIS (Automatic Identification System): Specifically, Automatic Identification System

[0063] Regarding real-time monitoring solutions for sea and airspace, this invention aims to address the lack of an efficient and large-scale monitoring technology in existing technologies. The proposed GPU-based all-weather satellite autonomous monitoring system for sea and airspace is detailed in the appendix. Figure 1 , 2As shown in 3 and 7, it specifically includes: an on-board super-resolution reconstruction target recognition system and an off-board data analysis and autonomous early warning system, which can be understood as the on-board part and the off-board part;

[0064] Specifically, the on-board super-resolution reconstruction target recognition system is based on a GPU system that can acquire a super-resolution image of target 1 and can acquire corresponding information of the super-resolution image through a recognition process.

[0065] Among them, Target 1 refers to offensive vehicles such as ships and aircraft based on the sea surface and airspace;

[0066] The method for acquiring the super-resolution image of target 1 is as follows: the visible light image and infrared image of target 1 obtained by an image acquisition system are fused and reconstructed to complete the acquisition of the super-resolution image.

[0067] The ground-based data analysis and autonomous early warning system can receive super-resolution images and corresponding information packaged and transmitted back by the target recognition system reconstructed by the satellite, and present them in the form of recognition information groups.

[0068] The satellite data analysis and autonomous early warning system can perform an analysis process on the identified information group to classify Target 1 and obtain the danger level of Target 1, and issue corresponding early warning prompts based on the danger level.

[0069] The on-board portion specifically refers to the super-resolution reconstruction system 10 and the target recognition system 20, which have a priority relationship. The super-resolution reconstruction system 10 works by fusing two types of images. When the on-board system is working, the super-resolution reconstruction system 10 and the target recognition system 20 can be loaded into the GPU system. The image acquisition system can acquire visible light and infrared images of target 1. The visible light and infrared images are processed through a VGG16 network of the super-resolution reconstruction system 10 to perform a super-resolution reconstruction process and obtain a super-resolution image of target 1. The target recognition system 20 performs a target recognition process and can obtain the super-resolution image of target 1 and corresponding information from the on-board super-resolution reconstruction system 10, package it in the form of recognition information groups, and transmit it back. The ground portion includes a target classification system 30 and a warning information system 40. The target classification system 30 is used to receive the recognition information groups and perform a parsing process to obtain the classification of target 1 and the danger level of target 1. The warning information generation system 40 can issue warnings based on the danger level.

[0070] The high efficiency of this technical solution is specifically reflected in "integrating the infrared and visible light super-resolution reconstruction network and the target recognition network together, which can reduce the time for GPU and CPU memory exchange and greatly increase the execution efficiency of the entire process"; furthermore, its technical advantages are reflected in the following three aspects:

[0071] Firstly, this technical solution enables real-time monitoring of the area around the clock;

[0072] Secondly, this system actually generates less load, less data (reduced by 80%), and lower data transmission costs (reduced by 50%); specifically, this includes reduced power consumption of data transmission equipment, reduced thermal control costs, reduced power supply costs, and reduced satellite launch costs.

[0073] Thirdly, this system and method integrate the visible light and infrared super-resolution reconstruction network with the identification network into an integrated network. Compared with traditional sea and airspace monitoring systems, this system has advantages such as a wider monitoring area, higher efficiency, fewer deployed devices, and higher equipment stability.

[0074] In one specific embodiment, please refer to the appendix. Figure 2 As shown, the on-board super-resolution reconstruction target recognition system includes a super-resolution reconstruction system 10 and a target recognition system 20, and both can be loaded into a GPU system.

[0075] Among them, the super-resolution reconstruction system 10 and the target recognition system 20 have independent threads;

[0076] Specifically, the super-resolution reconstruction system and the target recognition system are a large network. The super-resolution reconstruction system and the target recognition system are integrated into a large network, which is loaded into the GPU and is independent of the CPU thread.

[0077] The infrared and visible light images from the satellite are fused into a target recognition system, namely the super-resolution reconstruction system 10 and the target recognition system 20, using a VGG16 network + YOLOv3 combined network architecture. The VGG16 uses an unsupervised learning method, employing overall image noise, image brightness, and image structure similarity algorithms as loss functions. The fused super-resolution image is input into the YOLOv3 network module for target recognition, and the recognized image is cropped and auxiliary data is packaged and transmitted back to the ground.

[0078] In one specific embodiment, please refer to the appendix. Figure 1 , 2 As shown in Figures 3 and 4, the fusion reconstruction process is as follows: The super-resolution reconstruction system 10 is equipped with a VGG16 network, and the super-resolution image of target 1 is obtained through a super-resolution reconstruction process via the VGG16 network.

[0079] In one specific embodiment, please refer to the appendix. Figure 1 , 2As shown in Figures 3 and 4, the target recognition system 20 is capable of performing a target recognition process and can obtain the super-resolution image and corresponding information of target 1 from the on-board super-resolution reconstruction system 10, package it in the form of recognition information groups, and transmit it back.

[0080] The super-resolution reconstruction system 10 and the target recognition system 20 are located in an integrated network, which is a VGG16 network + YOLO v3 combined network architecture.

[0081] The specific working process is as follows: First, the super-resolution reconstruction and target recognition model is loaded into the GPU. Then, visible light and infrared images are input into the GPU for super-resolution reconstruction. Finally, the target recognition result is obtained and input into the target information parsing thread for target parsing and generating early warning information. Finally, the super-resolution image is cropped according to the target location, which can be done by cropping according to the target bounding box. The packaged information only requires the target image. It will be packaged according to the protocol and transmitted to the ground. The ground parsing process analyzes the data and classifies the target type, such as passenger planes, fighter jets, warships, aircraft carriers, and cruise ships. Based on the target danger level, early warning information is generated, and it is determined whether to trigger the alarm output of the alarm device.

[0082] Specifically, the target classification system 30 and the early warning information generation system 40 use the VGG19 classification network to classify the returned images and identify the specific types of targets (civilian merchant ships, warships, aircraft carriers, civilian Airbus, fighter jets, bombers, etc.). Early warning information is automatically generated based on the specific type of the target, and a danger level is generated to determine whether an alarm should be triggered.

[0083] In one specific embodiment, please refer to the appendix. Figure 1-4 As shown, the super-resolution reconstruction process is as follows:

[0084] The super-resolution reconstruction layer of the super-resolution reconstruction system 10 is embedded in a preset model for target 1 recognition based on YOLO v3, and the preset model is deployed to the GPU+ARM system. When the GPU+ARM system starts, it loads the data of the preset model into the GPU system and inputs the image into the GPU system according to the real-time acquired image.

[0085] The image is a super-resolution image; in the optional embodiment, embedded specifically, the hardware adopts and uses the Yulong810C, i.e., the Yulong810C smart chip, which has the functions of supporting deep learning and development libraries.

[0086] In one specific embodiment, please refer to the appendix. Figure 4 , 5As shown in Figure 6, the target recognition process is as follows: the image is divided into blocks, which are then input into the recognition network in sequence, and the target information is output.

[0087] The target information effect diagram test uses the DOTA dataset;

[0088] The identification results yielded integrated information about the target;

[0089] Subsequently, by cropping the image, the pixels of the cropped target 1 and the integrated information are extracted, and the recognition result is obtained based on the preset protocol;

[0090] Among them, the target recognition unit 20 can identify multiple targets 1 in the super-resolution image based on the YOLO module to obtain the target image and information containing the image;

[0091] The identification results include: the target's location information, category information, similarity information, and latitude and longitude.

[0092] In one specific embodiment, please refer to the appendix. Figure 8 , 9 As shown in Figure 10, the integrated network has a network training process, specifically as follows:

[0093] Freeze the YOLO model parameters, train a VGG16-based super-resolution reconstruction network to form the first network, and then freeze the super-resolution network parameters of the VGG16 network.

[0094] Next, a recognition network based on YOLO v3 is trained to form a second network. When training the YOLO network, the parameters of the last three layers of the YOLO network need to be frozen first, the parameters of the Darknet53 backbone network need to be trained, and finally the parameters of the last three layers of the network need to be unfrozen to perform unified training of the YOLO network.

[0095] Finally, the parameters of the first and second networks are unfrozen to obtain the integrated network through training;

[0096] The training of the preset model uses a dataset of pre-labeled visible light and infrared images;

[0097] Deploy the preset model onto a preset system with GPU and ARM.

[0098] In the existing research and development directions of the first network and the second network, no one has trained the two networks together. However, the technical solution of this invention provides a specific and feasible training method. Specifically, the purpose of merging the two networks is to address the current application scenario and, from a real-time perspective, avoid increasing the overall network execution time by having the two networks exchange data through the CPU.

[0099] The network training process, based on the training concept of YOLOv3, first freezes the last three layers of the network, trains the Darknet53 backbone network, trains 50 epochs, and then unfreezes and trains the entire network. Borrowing from this idea, the two networks are merged and trained separately. First, the YOLOv3 network parameters are frozen, and the infrared and visible light super-resolution reconstruction networks are trained. Then, the super-resolution reconstruction network is frozen, the YOLOv3 Darknet53 backbone network is unfrozen, and the last three layers of the YOLOv3 network are frozen. After training, all networks are unfrozen and trained as a whole, resulting in the final integrated network.

[0100] The specific principle can be understood as follows: the super-resolution reconstruction network and the object recognition network are merged into a large network. The main purpose of this method is to ensure that both super-resolution reconstruction and object recognition are performed in the GPU, without wasting running time by interacting with the CPU in memory, thus greatly improving the running efficiency.

[0101] The VGG16-based super-resolution reconstruction module primarily employs unsupervised learning methods, using algorithms that evaluate overall image noise, image brightness, and image structural similarity as loss functions.

[0102] The main training process is divided into three parts:

[0103] (1) Freeze the weight parameters of the YOLO v3 object recognition network and train the parameters of the VGG16 super-resolution reconstruction network.

[0104] (2) Freeze the VGG16 network parameters and freeze the weights of the last three layers of YOLO v3 network, and train the Darknet53 backbone network; then unfreeze all the YOLO v3 network weights and train the entire YOLO v3 network weights.

[0105] (3) Unfreeze the entire network and train it as a whole to obtain the infrared and visible light fusion target recognition network.

[0106] In one specific embodiment, please refer to the appendix. Figure 8 , 9 As shown in Figure 10, the process of cropping the image is as follows:

[0107] The image of target 1 is cropped, the latitude and longitude information of target 1 is calculated based on the cropping position, and the target similarity value output by target 1 is packaged together according to the preset protocol and transmitted back to the ground system.

[0108] The parsing process is as follows:

[0109] By parsing the data according to a preset packaging format, the data is analyzed and used for real-time ground-based early warning monitoring.

[0110] The target classification system 30 classifies the identification results, determines the target danger level based on the classification results, and determines whether to sound an alarm.

[0111] Specifically, it consists of a classification module and a warning information generation module;

[0112] The ground target classification system uses the VGG16 target classification model, which classifies the target 1 on the satellite to determine the type of offensive vehicle and generate the corresponding alarm level signal.

[0113] The advantages of this technical solution compared with existing technologies are:

[0114] (1) Traditional satellite payload systems require all high-resolution images to be transmitted to the ground for processing, which fails to achieve the purpose of real-time onboard identification and monitoring, and cannot achieve real-time monitoring of sea and airspace. In contrast, the system and method proposed in this application perform real-time target identification on the satellite and transmit the identified target location coordinates and cropped target images to the ground monitoring system. Compared with high-resolution images, the amount of data transmitted is less, and it includes the main target information.

[0115] (2) Visible light and infrared cameras acquire images of the same area at the same time. Under conditions such as rainy days, snowy days, and fog, the infrared camera will not be affected by the fog and can acquire images normally. However, when the infrared camera is used alone to acquire images, the infrared camera is affected by the diffraction limit and the imaging resolution is not high. In this application, a visible light camera is used to capture images and perform super-resolution reconstruction of visible light and infrared images. The two captured images are then transmitted to the recognition and task analysis system module to perform super-resolution reconstruction and target recognition operations.

[0116] In summary, this application proposes a system primarily designed for real-time monitoring applications in maritime and airspace, enabling low-cost and efficient real-time monitoring of target areas. This paper proposes a network that integrates a deep learning-based super-resolution reconstruction network and a target recognition network. By combining these two networks, infrared and visible light images are directly input into the network for super-resolution reconstruction and target recognition tasks. Integrating these tasks into a single GPU environment helps save time and increase resource utilization.

[0117] The GPU-based real-time airspace monitoring system proposed in this paper only requires the deployment of geostationary satellites. Multiple systems can be installed on one satellite and operate in parallel. The ground system automatically classifies targets, obtains classification results, generates target hazard levels, and determines whether an alarm should be triggered. The system autonomously and intelligently completes all operations without human intervention.

[0118] Ground developers can dynamically modify the software program of the satellite identification system according to their needs, and dynamically modify the super-resolution reconstruction and target recognition models to optimize the entire system in real time.

[0119] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A GPU-based all-weather satellite autonomous monitoring system for sea and airspace, characterized in that, include: The on-board super-resolution reconstruction target recognition system is based on a GPU system that can acquire a super-resolution image of the target (1) and can acquire the corresponding information of the super-resolution image through a recognition process; Among them, the target (1) is an offensive vehicle such as a ship or aircraft based on the sea surface and airspace; The method for obtaining the super-resolution image of the target (1) is as follows: the visible light image and infrared image of the target (1) obtained by an image acquisition system are fused and reconstructed to complete the acquisition of the super-resolution image; The ground-based data analysis and autonomous early warning system can receive the super-resolution image and corresponding information packaged and transmitted back by the target recognition system reconstructed by the on-board super-resolution system, and present it in the form of recognition information groups. The satellite data analysis and autonomous early warning system can perform an analysis process on the identification information group to obtain the classification of the target (1), obtain the danger level of the target (1), and provide corresponding early warning prompts based on the danger level; The on-board super-resolution reconstruction target recognition system includes a super-resolution reconstruction system (10) and a target recognition system (20), and both can be loaded into a GPU system; The super-resolution reconstruction system (10) and the target recognition system (20) have independent threads; The fusion and reconstruction process is as follows: The super-resolution reconstruction system (10) is equipped with a VGG16 network, and the super-resolution image of the target (1) is obtained through a super-resolution reconstruction process via the VGG16 network. The target recognition system (20) is capable of performing a target recognition process and can obtain the super-resolution image and corresponding information of the target (1) from the on-board super-resolution reconstruction system (10), package it in the form of the recognition information group, and send it back. The super-resolution reconstruction system (10) and the target recognition system (20) are located in an integrated network, which is a VGG16 network + YOLO v3 combined network architecture. The super-resolution reconstruction process is as follows: The super-resolution reconstruction layer of the super-resolution reconstruction system (10) is embedded in a preset model for target (1) recognition based on YOLO v3, and the preset model is deployed to the GPU+ARM system. The GPU+ARM system starts to load the data of the preset model into the GPU system and inputs the image into the GPU system according to the real-time acquired image. The image in question is a super-resolution image.

2. The GPU-based all-weather satellite autonomous monitoring system for sea and airspace as described in claim 1, characterized in that, The target recognition process is as follows: the image is divided into blocks, which are then sequentially input into the recognition network, and the target information is output. The target information effect diagram test uses the DOTA dataset; The identification results yielded integrated information about the target; Subsequently, by cropping the image, the pixels of the cropped target (1) and the integrated information are extracted, and the recognition result is obtained based on the preset protocol; The target recognition unit (20) is capable of recognizing multiple targets (1) in the super-resolution image based on the YOLO module to obtain the target image and information containing the image; The identification results include: the target's location information, category information, similarity information, and latitude and longitude.

3. The GPU-based all-weather satellite autonomous monitoring system for sea and airspace as described in claim 2, characterized in that, The integrated network has a network training process, specifically: Freeze the YOLO model parameters, train a VGG16-based super-resolution reconstruction network to form the first network, and then freeze the super-resolution network parameters of the VGG16 network. Next, a recognition network based on YOLO v3 is trained to form a second network. When training the YOLO network, the parameters of the last three layers of the YOLO network need to be frozen first, the parameters of the Darknet53 backbone network need to be trained, and finally the parameters of the last three layers of the network need to be unfrozen to perform unified training of the YOLO network. Finally, all parameters of the first and second networks are unfrozen to obtain an integrated network through training; The training of the preset model uses a pre-labeled dataset of visible light images and infrared images; The preset model is deployed to a preset system with GPU and ARM.

4. The GPU-based all-weather satellite autonomous monitoring system for sea and airspace as described in claim 3, characterized in that, The process of cropping the image is as follows: The image of the target (1) is cropped, the latitude and longitude information of the current target (1) is calculated according to the cropping position, and the target similarity value output by the target (1) is packaged together according to the preset protocol and transmitted back to the ground system.

5. The GPU-based all-weather satellite autonomous monitoring system for sea and airspace as described in claim 2, characterized in that, The parsing process is as follows: By parsing the data according to a preset packaging format, the data is analyzed and used for real-time ground-based early warning monitoring. The target classification system (30) classifies the identification results, determines the target danger level based on the classification results, and determines whether to sound an alarm. Specifically, it consists of a classification module and a warning information generation module; The ground target classification system uses the VGG16 target classification model, which classifies the targets (1) on the satellite to determine the type of the offensive vehicle and generate a corresponding alarm level signal.

Images