Satellite and ship target on-orbit detection device determination method, device and detection method

Abstract

The present application provides a method for determining an on-orbit detection device for satellite and ship targets, a device, and a detection method, which relates to the technical field of on-orbit detection of ship targets and satellites. The method comprises obtaining a target detection model, determining weight values ​​and activation values; sorting channels from large to small according to importance according to the L1 norm; dividing the sorted channels into m equal parts, and performing quantization and pruning to obtain a quantized and pruned target detection model; using a layer fusion method to fuse the convolution + BN layer and the fully connected + BN layer of the quantized and pruned target detection model to obtain a layer fusion target detection model; determining the FPGA hardware operators required in the forward reasoning model to implement the layer fusion target detection model; and determining the required FPGA board resources based on the required FPGA hardware operators and the target detection frame rate. The present application combines lightweight ship target detection with on-orbit ship target detection through collaborative software and hardware to achieve efficient real-time detection of target ships at sea by on-orbit satellites.

Description

Technical Field

[0001] The present application relates to the technical field of on-orbit detection of ship targets and satellites, and in particular to a method for determining an on-orbit detection device for satellite and ship targets, a device and a detection method. Background Art

[0002] Real-time monitoring and identification of ship targets is crucial for global maritime security, resource development, and environmental monitoring. Synthetic Aperture Radar (SAR) satellites, with their all-weather, all-day observation capabilities, have become a core tool for maritime target monitoring. However, due to limited resources, traditional SAR satellites cannot carry large server computing resources, making them incapable of real-time on-orbit detection of ship targets. Summary of the Invention

[0003] The purpose of this application is to provide a method for determining an on-orbit detection device for satellite ship targets, a device and a detection method, which can realize real-time on-orbit detection of ship targets.

[0004] To achieve the above objectives, this application provides the following solutions:

[0005] In a first aspect, the present application provides a method for determining an on-orbit detection device for a satellite or ship target, comprising:

[0006] Get the trained target detection model;

[0007] Determine the weight value and activation value of each convolution kernel layer in the target detection model;

[0008] Calculate the L1 norm based on the weight value and the activation value;

[0009] Using the L1 norm as the importance evaluation value of the channel corresponding to the convolution kernel;

[0010] Sort the channels by importance from greatest to least;

[0011] Dividing the sorted channel into six parts;

[0012] For the first channel after division, retain the original floating-point number; for the second channel after division, quantize the original floating-point number to 32 bits; for the third channel after division, quantize the original floating-point number to 16 bits; for the fourth channel after division, quantize the original floating-point number to 8 bits; for the fifth channel after division, quantize the original floating-point number to 4 bits; for the sixth channel after division, prune to obtain a quantized pruned object detection model;

[0013] The convolution+BN layer and the fully connected+BN layer of the quantized pruning target detection model are fused using a layer fusion method to obtain a layer fusion target detection model;

[0014] Determining the FPGA hardware operators required to implement the forward reasoning model of the layer fusion target detection model;

[0015] The required FPGA board resources are determined based on the required FPGA hardware operators combined with the target detection frame rate.

[0016] In a second aspect, the present application provides an on-orbit detection device for a satellite ship target, comprising: a hardware acceleration module; the hardware acceleration module is used to be carried on an on-orbit SAR satellite;

[0017] The hardware acceleration module includes: a hardware management system, a top-level file of an inference model, and a plurality of inference operators; the hardware management system is connected to the plurality of inference operators via the top-level file of the inference model;

[0018] The inference model top-level file is used to control the construction of the multiple inference operators and load the layer fusion target detection model;

[0019] The steps of acquiring the layer fusion target detection model include:

[0020] Get the trained target detection model;

[0021] Determine the weight value and activation value of each convolution kernel layer in the target detection model;

[0022] Calculate the L1 norm based on the weight value and the activation value;

[0023] Using the L1 norm as the importance evaluation value of the channel corresponding to the convolution kernel;

[0024] Sort the channels by importance from greatest to least;

[0025] Dividing the sorted channel into six parts;

[0026] For the first channel after division, retain the original floating-point number; for the second channel after division, quantize the original floating-point number to 32 bits; for the third channel after division, quantize the original floating-point number to 16 bits; for the fourth channel after division, quantize the original floating-point number to 8 bits; for the fifth channel after division, quantize the original floating-point number to 4 bits; for the sixth channel after division, prune to obtain a quantized pruned object detection model;

[0027] The convolution+BN layer and the fully connected+BN layer of the quantized pruning target detection model are fused using a layer fusion method to obtain a layer fusion target detection model.

[0028] Optionally, the hardware acceleration module further includes: a detection information sending module; the detection information sending module is connected to the hardware management system;

[0029] The detection information sending module is used to send the results of the on-orbit detection of the ship target to the ground receiving station.

[0030] Optionally, the hardware acceleration module further includes: a to-be-detected image cache unit; the to-be-detected image cache unit is connected to the hardware management system;

[0031] The to-be-detected image cache unit is used to cache pre-processed data of the to-be-detected satellite real-time image.

[0032] Optionally, the hardware acceleration module is an FPGA hardware platform.

[0033] In a third aspect, the present application provides a method for detecting satellite and ship targets on-orbit according to the above-mentioned device for detecting satellite and ship targets on-orbit, comprising:

[0034] Using on-orbit satellites to obtain real-time satellite image data;

[0035] Using a lightweight on-orbit ship target detection model, ship target detection is performed on the real-time satellite image data to obtain a detection result; the detection result includes: a ship target mark image, a target frame, a target position, and image imaging time information;

[0036] The detection result is sent to a ground receiving station.

[0037] Optionally, after sending the detection result to the ground receiving station, the method further includes:

[0038] The real-time satellite image data corresponding to the detection result is deleted.

[0039] Optionally, performing ship target detection on the satellite real-time image data using the lightweight on-orbit ship target detection model to obtain a detection result specifically includes:

[0040] Performing data preprocessing on the satellite real-time image data to obtain preprocessed data;

[0041] Utilizing a cache unit to be processed to acquire and store the pre-processed data;

[0042] The pre-processed data is input into a lightweight ship target on-orbit detection model to perform ship target detection and obtain a detection result.

[0043] Optionally, the training process of the ship target on-orbit detection model includes:

[0044] Acquire historical satellite image data with labels, wherein the labels are ship target labels;

[0045] The historical satellite image data is used to train an on-orbit ship target detection network to obtain an on-orbit ship target detection model.

[0046] Optionally, the ship target on-orbit detection network is a lightweight ship target detection model; the lightweight ship target detection model is a Yolo target detection network or an RT-DETR target detection network.

[0047] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0048] The present application provides a method, device, and detection method for determining an on-orbit detection device for satellite and ship targets. The method comprises: obtaining a trained target detection model; determining weight values ​​and activation values ​​of each convolution kernel layer in the target detection model; calculating an L1 norm based on the weight values ​​and activation values; using the L1 norm as an importance evaluation value of a channel corresponding to the convolution kernel; sorting the channels from greatest to least importance; and dividing the sorted channels into six parts.

[0049] For the first channel after division, the original floating-point number is retained; for the second channel after division, the original floating-point number is quantized to 32 bits; for the third channel after division, the original floating-point number is quantized to 16 bits; for the fourth channel after division, the original floating-point number is quantized to 8 bits; for the fifth channel after division, the original floating-point number is quantized to 4 bits; for the sixth channel after division, pruning is performed to obtain a quantized pruned target detection model; the convolutional + BN layer and the fully connected + BN layer of the quantized pruned target detection model are fused using a layer fusion method to obtain a layer fusion target detection model; the FPGA hardware operators required for implementing the forward inference model of the layer fusion target detection model are determined; and the required FPGA board resources are determined based on the required FPGA hardware operators and the target detection frame rate. In the existing technology, due to limitations such as limited on-orbit computing resources, complex detection models, and insufficient hardware support, it is difficult for SAR satellites to achieve on-orbit real-time detection of ship targets. In this application, after lightweight processing and layer fusion optimization of the on-orbit ship target detection model, it is loaded into the hardware acceleration module and implemented using several hardware inference operators. This combines the lightweight ship target detection and on-orbit ship target detection of software and hardware collaboration to achieve efficient detection of maritime target ships by satellites in real time. This application can achieve real-time on-orbit detection of ship targets, can monitor ocean dynamics in real time and efficiently, and provide timely and reliable maritime security. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0051] Figure 1 A schematic flow chart of a method for determining an on-orbit detection device for a satellite or ship target provided in one embodiment of the present application;

[0052] Figure 2 A schematic structural diagram of an on-orbit detection device for satellite and ship targets provided in one embodiment of the present application;

[0053] Figure 3 A flowchart of an on-orbit detection method for satellite and ship targets provided in one embodiment of the present application;

[0054] Figure 4 A flowchart of an on-orbit detection method for satellite and ship targets provided in one embodiment of the present application;

[0055] Figure 5 A schematic diagram of the on-orbit ship target detection process provided in one embodiment of the present application. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0057] This application proposes a satellite ship target on-orbit detection device and method, which detects whether there is a ship target in the image by analyzing the SAR satellite imaging information in real time on-orbit, and sends the detected ship target mark image, target frame (W, H), target position (LON, LAT), and image imaging time and other information to the ground receiving station, thereby achieving efficient detection of ship targets while effectively reducing satellite communication bandwidth. Among them, this article proposes multiple exemplary embodiments, which can be selected by those skilled in the art. Among them, the attached Figure 4 This is a schematic diagram of the overall process of this application, which completes the overall on-orbit detection process. Figure 1 and Figure 5 yes Figure 4 The supplements represent specific processes of different embodiments.

[0058] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0059] In an exemplary embodiment, Figure 1 As shown, a method for determining an on-orbit detection device for a satellite or ship target is provided, comprising:

[0060] Get the trained target detection model;

[0061] Determine the weight value and activation value of each convolution kernel layer in the target detection model;

[0062] Calculate the L1 norm based on the weight value and the activation value;

[0063] Using the L1 norm as the importance evaluation value of the channel corresponding to the convolution kernel;

[0064] Sort the channels by importance from greatest to least;

[0065] Dividing the sorted channel into six parts;

[0066] For the first channel after division, retain the original floating-point number; for the second channel after division, quantize the original floating-point number to 32 bits; for the third channel after division, quantize the original floating-point number to 16 bits; for the fourth channel after division, quantize the original floating-point number to 8 bits; for the fifth channel after division, quantize the original floating-point number to 4 bits; for the sixth channel after division, prune to obtain a quantized pruned object detection model;

[0067] The convolution+BN layer and the fully connected+BN layer of the quantized pruning target detection model are fused using a layer fusion method to obtain a layer fusion target detection model;

[0068] Determining the FPGA hardware operators required to implement the forward reasoning model of the layer fusion object detection model;

[0069] The required FPGA board resources are determined based on the required FPGA hardware operators combined with the target detection frame rate.

[0070] In this embodiment, the method for implementing the process of "dividing the sorted channel into six parts; retaining the original floating-point number for the first channel after the division; quantizing the original floating-point number for the second channel after the division into 32 bits; quantizing the original floating-point number for the third channel after the division into 16 bits; quantizing the original floating-point number for the fourth channel after the division into 8 bits; quantizing the original floating-point number for the fifth channel after the division into 4 bits; and pruning the sixth channel after the division to obtain a quantized pruned target detection model" can be:

[0071] Dividing the sorted channel into m equal parts;

[0072] For the first The channel after division retains the original floating point number; The channel after the division is divided, and the original floating point number is quantized into 32 bits; The channel after the division is divided, and the original floating point number is quantized into 16 bits; The channel after the division is divided, and the original floating point number is quantized into 8 bits; The channel after the division is divided, and the original floating point number is quantized into 4 bits; The channels are divided and pruned to obtain a quantized pruned target detection model; wherein 0<n1<n2<n3<n4<n5<m.

[0073] In this embodiment, obtaining a trained target detection model specifically includes:

[0074] Set the target detection model's indicators of precision, recall, and detection frame rate.

[0075] Then, a target detection network is designed based on this, and the designed target detection network is trained using historical ship image data to obtain a target detection model.

[0076] Furthermore, in this embodiment, after determining the required FPGA board resources based on the required FPGA hardware operators and the target detection frame rate, it is necessary to test whether the aforementioned precision, recall, and detection frame rate are met. If not, the target detection network design step is returned to and readjusted (primarily adjusting the channel division and the different processing for different channels) until the requirements are met.

[0077] In an exemplary embodiment, Figure 2As shown, a satellite and ship target on-orbit detection device is provided, which includes: a hardware acceleration module; the hardware acceleration module is used to be carried on an on-orbit SAR satellite; the hardware acceleration module is an FPGA hardware platform.

[0078] The hardware acceleration module includes: a hardware management system, a top-level file of an inference model and a plurality of inference operators; the hardware management system is connected to the plurality of inference operators through the top-level file of the inference model.

[0079] The inference model top-level file is used to control the construction of the multiple inference operators and load the layer fusion target detection model;

[0080] The steps of acquiring the layer fusion target detection model include:

[0081] Get the trained target detection model;

[0082] Determine the weight value and activation value of each convolution kernel layer in the target detection model;

[0083] Calculate the L1 norm based on the weight value and the activation value;

[0084] Using the L1 norm as the importance evaluation value of the channel corresponding to the convolution kernel;

[0085] Sort the channels by importance from greatest to least;

[0086] Dividing the sorted channel into m equal parts;

[0087] For the first The channel after division retains the original floating point number; The channel after the division is divided, and the original floating point number is quantized into 32 bits; The channel after the division is divided, and the original floating point number is quantized into 16 bits; The channel after the division is divided, and the original floating point number is quantized into 8 bits; The channel after the division is divided, and the original floating point number is quantized into 4 bits; The channels are pruned to obtain a quantized pruned target detection model; wherein 0<n1<n2<n3<n4<n5<m;

[0088] The convolution+BN layer and the fully connected+BN layer of the quantized pruning target detection model are fused using a layer fusion method to obtain a layer fusion target detection model.

[0089] Wherein, the hardware acceleration module further includes: a detection information sending module; the detection information sending module is connected to the hardware management system.

[0090] The detection information sending module is used to send the results of the on-orbit detection of the ship target to the ground receiving station.

[0091] The hardware acceleration module further includes: an image cache unit to be detected; the image cache unit to be detected is connected to the hardware management system.

[0092] The to-be-detected image cache unit is used to cache the real-time satellite image data to be detected.

[0093] This embodiment provides a satellite ship target on-orbit detection device, comprising: a hardware acceleration module; the hardware acceleration module is configured to be carried on an on-orbit SAR satellite; the hardware acceleration module comprises: a hardware management system, an inference model top-level file, and several inference operators; the hardware management system is connected to the several inference operators via the inference model top-level file; the inference model top-level file is configured to control the several inference operators to construct a target detection inference model and load a lightweight ship target on-orbit detection model; before loading, the parameters of the lightweight ship target on-orbit detection model are optimized using a layer fusion method; the lightweight ship target on-orbit detection model is a ship target on-orbit detection model that has undergone network model pruning; the network model pruning is performed by performing unstructured pruning on the ship target on-orbit detection model using a data-driven model pruning method; the ship target on-orbit detection model is configured to perform on-orbit detection of ship targets based on real-time satellite image data. In the prior art, due to limitations such as limited on-orbit computing resources, complex detection models, and insufficient hardware support, SAR satellites have difficulty in achieving real-time on-orbit detection of ship targets. In this application, after lightweight processing and layer fusion optimization of the on-orbit ship target detection model, it is loaded into the hardware acceleration module and implemented using several inference operators. This combines the lightweight ship target detection and on-orbit ship target detection of software and hardware collaboration, and realizes the efficient detection of maritime ship targets by SAR satellites in real time on orbit. This application can realize real-time on-orbit detection of ship targets, can monitor ocean dynamics in real time and efficiently, and provide timely and reliable maritime security.

[0094] In another exemplary embodiment of the present application, Figure 3 As shown, a method for detecting satellite and ship targets on-orbit is provided. The method is based on the above-mentioned device for detecting satellite and ship targets on-orbit. The method for detecting satellite and ship targets on-orbit includes:

[0095] S1. Use in-orbit satellites to obtain real-time satellite image data.

[0096] S2. Use a lightweight on-orbit ship target detection model to perform ship target detection on the real-time satellite image data to obtain a detection result. The detection result includes: a ship target mark image, a target frame, a target position, and image imaging time information.

[0097] The specific steps include:

[0098] The real-time satellite image data is preprocessed to obtain preprocessed data. The preprocessing process in this embodiment is to remove sea clutter noise from the real-time satellite image data.

[0099] Utilizing a cache unit to be processed to acquire and store the pre-processed data;

[0100] The pre-processed data is input into a lightweight ship target on-orbit detection model to perform ship target detection and obtain a detection result.

[0101] S3. Send the detection result to a ground receiving station.

[0102] After sending the detection result to the ground receiving station, the method further includes:

[0103] The real-time satellite image data corresponding to the detection result is deleted.

[0104] The training process of the on-orbit ship target detection model includes:

[0105] Acquire historical satellite image data with labels, wherein the labels are ship target labels;

[0106] The historical satellite image data is used to train an on-orbit ship target detection network to obtain an on-orbit ship target detection model. The on-orbit ship target detection network is a lightweight ship target detection model, such as a Yolo target detection network or an RT-DETR target detection network.

[0107] In another exemplary embodiment of the present application, Figure 4 As shown, a method for detecting satellite and ship targets on orbit is provided, and the specific steps are as follows:

[0108] A1: Acquire Synthetic Aperture Radar (SAR) satellite imaging data, mark ship targets, and construct training samples and test libraries for on-orbit ship target detection network training and hardware platform testing.

[0109] A2: Using an FPGA-based software-hardware collaborative approach, we design satellite-based lightweight ship target detection hardware and an on-orbit ship target detection model.

[0110] A3: Send the SAR satellite carrying the ship target hardware platform and the ship target on-orbit detection model into the predetermined orbit.

[0111] A4: Obtain real-time SAR satellite image data.

[0112] A5: Since the real-time SAR satellite image data usually obtained contains a lot of noise, certain data preprocessing is required to prepare for subsequent target detection.

[0113] A6: The pre-processed SAR satellite real-time image data is stored in a buffer unit to be processed, waiting for the hardware to read it.

[0114] A7: Perform ship detection on the images in the cached units in step A6 one by one. The images here are the images in the SAR satellite real-time image data that have been pre-processed, such as Figure 5 shown.

[0115] The steps for training the on-orbit ship target detection model and building a lightweight target detection module are as follows:

[0116] B1: Set detection indicators according to target detection requirements.

[0117] B2: Based on the SAR ship target imaging characteristics, a lightweight ship target detection network is designed, such as Yolo, RT-DETR and other detection models.

[0118] B3: Due to the large size of the network model, it is difficult to directly deploy it on a resource-constrained FPGA hardware platform. Using a data-driven model pruning method, we perform unstructured pruning on the model to reduce the model's computational and storage requirements.

[0119] B4: High-precision computing is expensive and requires significant hardware resources. FPGAs, however, enable custom precision design (such as 16-bit, 8-bit, or 4-bit). By incorporating quantization into the model training phase, this significantly reduces computational and memory requirements during deployment while maximizing model accuracy.

[0120] B8: To accelerate model inference and reduce computational costs, the derived model parameters are optimized through techniques such as layer fusion. For example, multiple consecutive neural network layers are fused into an equivalent operation, thereby reducing the hardware computation steps and data transmission overhead.

[0121] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0122] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0123] The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, and the like.

[0124] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0125] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for determining an on-orbit detection device for a satellite or ship target, characterized in that: The method for determining the on-orbit detection device for satellite and ship targets includes: Get the trained target detection model; Determine the weight value and activation value of each convolution kernel layer in the target detection model; Calculate the L1 norm based on the weight value and the activation value; Using the L1 norm as the importance evaluation value of the channel corresponding to the convolution kernel; Sort the channels by importance from greatest to least; Dividing the sorted channel into m equal parts; For the first The channel after division retains the original floating point number; The channel after the division is divided, and the original floating point number is quantized into 32 bits; The channel after the division is divided, and the original floating point number is quantized into 16 bits; The channel after the division is divided, and the original floating point number is quantized into 8 bits; The channel after the division is divided, and the original floating point number is quantized into 4 bits; The channels are pruned to obtain a quantized pruned target detection model; wherein 0<n1<n2<n3<n4<n5<m; The convolution+BN layer and the fully connected+BN layer of the quantized pruning target detection model are fused using a layer fusion method to obtain a layer fusion target detection model; Determining the FPGA hardware operators required to implement the forward reasoning model of the layer fusion object detection model; Determine the required FPGA board resources based on the required FPGA hardware operators and the target detection frame rate; Obtain the trained target detection model, including: Set the target detection model's indicators of precision, recall, and detection frame rate; Then, a target detection network is designed based on this data, and the designed target detection network is trained using historical ship image data to obtain a target detection model. The target detection network is the Yolo target detection network or the RT-DETR target detection network; Using on-orbit satellites to obtain real-time satellite image data; Performing data preprocessing on the satellite real-time image data to obtain preprocessed data; Utilizing a cache unit to be processed to acquire and store the pre-processed data; Inputting the preprocessed data into a lightweight on-orbit ship target detection model to perform ship target detection, and obtaining a detection result; the detection result includes: a ship target mark image, a target frame, a target position, and image imaging time information; Sending the detection results to a ground receiving station; The training process of the ship target on-orbit detection model includes: Acquire historical satellite image data with labels, wherein the labels are ship target labels; The historical satellite image data is used to train an on-orbit ship target detection network to obtain an on-orbit ship target detection model.

2. A satellite and ship target on-orbit detection device, characterized in that: The satellite ship target on-orbit detection device includes: a hardware acceleration module; the hardware acceleration module is used to be carried on the on-orbit SAR satellite; The hardware acceleration module includes: a hardware management system, a top-level file of an inference model, and a plurality of inference operators; the hardware management system is connected to the plurality of inference operators via the top-level file of the inference model; The inference model top-level file is used to control the construction of the multiple inference operators and load the layer fusion target detection model; The steps of acquiring the layer fusion target detection model include: Get the trained target detection model; Determine the weight value and activation value of each convolution kernel layer in the target detection model; Calculate the L1 norm based on the weight value and the activation value; Using the L1 norm as the importance evaluation value of the channel corresponding to the convolution kernel; Sort the channels by importance from greatest to least; Dividing the sorted channel into m equal parts; For the first The channel after division retains the original floating point number; The channel after the division is divided, and the original floating point number is quantized into 32 bits; The channel after the division is divided, and the original floating point number is quantized into 16 bits; The channel after the division is divided, and the original floating point number is quantized into 8 bits; The channel after the division is divided, and the original floating point number is quantized into 4 bits; The channels are pruned to obtain a quantized pruned target detection model; wherein 0<n1<n2<n3<n4<n5<m; The convolution+BN layer and the fully connected+BN layer of the quantized pruning target detection model are fused using a layer fusion method to obtain a layer fusion target detection model; Obtain the trained target detection model, including: Set the target detection model's indicators of precision, recall, and detection frame rate; Then, a target detection network is designed based on this data, and the designed target detection network is trained using historical ship image data to obtain a target detection model. The target detection network is the Yolo target detection network or the RT-DETR target detection network; Using on-orbit satellites to obtain real-time satellite image data; Performing data preprocessing on the satellite real-time image data to obtain preprocessed data; Utilizing a cache unit to be processed to acquire and store the pre-processed data; Inputting the preprocessed data into a lightweight on-orbit ship target detection model to perform ship target detection, and obtaining a detection result; the detection result includes: a ship target mark image, a target frame, a target position, and image imaging time information; Sending the detection results to a ground receiving station; The training process of the ship target on-orbit detection model includes: Acquire historical satellite image data with labels, wherein the labels are ship target labels; The historical satellite image data is used to train an on-orbit ship target detection network to obtain an on-orbit ship target detection model.

3. The on-orbit detection device for satellite and ship targets according to claim 2, characterized in that: The hardware acceleration module further includes: a detection information sending module; the detection information sending module is connected to the hardware management system; The detection information sending module is used to send the results of the on-orbit detection of the ship target to the ground receiving station.

4. The on-orbit detection device for satellite and ship targets according to claim 2, characterized in that: The hardware acceleration module further includes: a to-be-detected image cache unit; the to-be-detected image cache unit is connected to the hardware management system; The to-be-detected image cache unit is used to cache the real-time satellite image data to be detected.

5. The on-orbit detection device for satellite and ship targets according to claim 2, characterized in that: The hardware acceleration module is an FPGA hardware platform.

6. The on-orbit detection device for satellite and ship targets according to claim 2, characterized in that: After sending the detection result to the ground receiving station, the method further includes: The real-time satellite image data corresponding to the detection result is deleted.

7. The on-orbit detection device for satellite and ship targets according to claim 2, characterized in that: The on-orbit ship target detection network is a lightweight ship target detection model; the lightweight ship target detection model is a Yolo target detection network or an RT-DETR target detection network.

Images