A low-power consumption space target image acquisition and real-time detection integrated method and system

Abstract

The application provides a low-power space target image acquisition and real-time detection integrated method and system. The system is based on a ZYNQ platform, acquires spacecraft images through an image sensor, and performs denoising and low-light enhancement preprocessing using an FPGA. The detection algorithm replaces the SSD backbone network with MobileNetV4, introduces a CBAM attention mechanism, and optimizes anchor boxes using K-means++, thereby improving small target detection accuracy. The model is pruned and quantized, deployed on an FPGA for accelerated execution, and an ARM is responsible for task scheduling. The application has the advantages of high integration, low power consumption, strong real-time performance, and good low-light adaptability, with a total power consumption of only about 7.3W and a single-frame detection latency of less than 10ms, and is suitable for resource-constrained spaceborne embedded environments.

Description

Technical Field

[0001] This invention belongs to the field of spatial target perception and embedded intelligent vision technology, and in particular, it is a low-power integrated method and system for spatial target image acquisition and real-time detection. Background Technology

[0002] With the development of aerospace technology, the detection of spacecraft targets in space has become a crucial means of ensuring flight safety and mission success. In space missions, accurate detection of spacecraft targets is fundamental for space navigation, rendezvous and docking, and situational awareness. However, space platforms typically have limited computing resources. Traditional deep learning methods based on image processing or large GPU platforms, such as YOLOv5 and Faster R-CNN, have high demands for computing power and memory, and face problems such as high power consumption, large size, and poor real-time performance in space, making them difficult to deploy directly on resource-constrained space platforms.

[0003] MobileNetV4, a lightweight and efficient feature extraction neural network, boasts a small number of parameters and high computational efficiency, while the SSD (Single Shot MultiBox Detector) algorithm possesses strong object detection capabilities, enabling end-to-end detection. The combination of the two holds great promise for embedded applications. However, the combined algorithm still suffers from insufficient detection accuracy in spatial image scenarios with complex backgrounds, small target sizes, and weak features. Therefore, it is necessary to enhance the network's feature representation capabilities and optimize the anchor box matching strategy of the detection head to improve the detection performance of small targets.

[0004] The ZYNQ (Programmable SoC) platform integrates ARM processors and FPGA logic resources, offering advantages such as high performance, low power consumption, and customizability, making it an ideal choice for deploying neural network models in space platforms. However, deploying deep learning models to the ZYNQ platform still faces challenges such as limited hardware resources, bandwidth constraints, and power consumption limitations. Therefore, joint optimization of the entire process, from algorithm structure to hardware deployment, is necessary. Summary of the Invention

[0005] The purpose of this invention is to address the problems and deficiencies of existing technologies by providing a low-power integrated method and system for spatial target image acquisition and real-time detection.

[0006] The technical solution for achieving the objective of this invention is: a low-power integrated method for spatial target image acquisition and real-time detection, the method comprising:

[0007] Step 1: Initialize the image sensor parameters and load the model parameters;

[0008] Step 2: Acquire spacecraft target image data using an image sensor;

[0009] Step 3: Preprocess the acquired target image data;

[0010] Step 4: Use spacecraft detection algorithms to perform target detection on the preprocessed image;

[0011] Step 5: Output the location and category information of the spacecraft target;

[0012] Step 6: Store and display the image data and detection results.

[0013] Furthermore, step 1 specifically includes:

[0014] The processor configures the image sensor registers and transmits the configuration parameters to the logic control unit via the bus interface.

[0015] The logic control unit drives the image sensor to write configuration information in order to control the sensor's working state and acquisition parameters;

[0016] The pre-trained lightweight detection model weight parameters are stored in external memory.

[0017] Furthermore, the preprocessing in step 3 includes:

[0018] In the logic control unit, the received raw target image data is processed by signal decoding, byte alignment, data width conversion, gamma correction, and noise reduction.

[0019] An IP core for a low-light enhancement algorithm for compressed images is generated using high-level synthesis HLS.

[0020] Furthermore, the spacecraft detection algorithm described in step 4 uses MobileNetV4 as the backbone feature extraction network and integrates the SSD detection framework to obtain target detection results at different scales.

[0021] Furthermore, the MobileNetV4 backbone feature extraction network introduces a general inverted bottleneck block (UIB) and a Mobile MQA attention block, and selects the output of a specific layer as the prediction feature layer to enhance the small target detection effect.

[0022] Furthermore, in the MobileNetV4 backbone feature extraction network, a CBAM module is connected after the feature extraction convolutional layer to enhance the feature representation capability. The CBAM module includes a channel attention module and a spatial attention module.

[0023] Furthermore, the spacecraft detection algorithm uses the K-means++ clustering algorithm to analyze the bounding boxes in the spacecraft image dataset, generating multiple anchor box sizes and aspect ratios that match the dataset, to replace the default prior box settings.

[0024] Furthermore, the spacecraft detection algorithm performs model compression before inference, including pruning and quantizing the model, and performing point-to-point processing on the data.

[0025] Furthermore, the object detection in step 4 is implemented on a heterogeneous platform, with the following specific architecture:

[0026] On the processor side, complete the processing task scheduling and image frame buffering;

[0027] An acceleration core is deployed on the logic control unit to perform feature extraction, target regression classification prediction, and image enhancement operations.

[0028] Furthermore, in step 6, data storage adopts a ping-pong buffer structure for data transmission, and the detection results and image data are transmitted to the host computer for display in real time via Ethernet.

[0029] Compared with the prior art, the significant advantages of this invention are:

[0030] 1) High system integration: Image acquisition, enhancement and detection are completed on a single spaceborne platform, reducing system complexity.

[0031] 2) High real-time performance and low power consumption: Fully utilize the parallel computing capabilities of FPGA to significantly reduce processing latency and system power consumption.

[0032] 3) Strong low-light adaptability: Improved target detection reliability under complex lighting conditions through on-board image enhancement processing.

[0033] 4) Low communication burden: Supports transmitting only target detection results, reducing the bandwidth requirements for satellite-to-ground communication.

[0034] 5) Highly deployable: Suitable for resource-constrained spaceborne embedded environments, with good practical application value.

[0035] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0036] Figure 1 This is a flowchart of a low-power spatial target image acquisition and real-time detection integrated system in one embodiment.

[0037] Figure 2 This is a block diagram of a lightweight MobileNetV4-SSD spacecraft detection algorithm design in one embodiment.

[0038] Figure 3 This is a schematic diagram of the overall network structure of the MobileNetV4-SSD spacecraft detection algorithm in one embodiment.

[0039] Figure 4This is a deployment architecture diagram of the ZYNQ platform in one embodiment.

[0040] Figure 5 This is a software design block diagram for one embodiment.

[0041] Figure 6 This is a detection result image under low light conditions in one embodiment, wherein... Figure 6 Image (a) in the image shows the detection results for a simple detection scenario. Figure 6 (b) in the image shows the detection results for a densely packed detection scene with targets of varying sizes. Figure 6 (c) is relatively Figure 6 (b) shows the detection results for a more densely populated detection scenario. Figure 6 (d) in the diagram represents the detection result for the most complex detection scenario. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0043] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0044] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0045] In one embodiment, combined Figure 1 This paper provides a low-power integrated method for spatial target image acquisition and real-time detection, the method comprising:

[0046] Step 1: Initialize the image sensor parameters and load the model parameters;

[0047] Step 2: Acquire spacecraft target image data using an image sensor;

[0048] Step 3: Preprocess the acquired target image data;

[0049] Step 4: Use spacecraft detection algorithms to perform target detection on the preprocessed image;

[0050] Step 5: Output the location and category information of the spacecraft target;

[0051] Step 6: Store and display the image data and detection results.

[0052] Furthermore, in one embodiment, step 1 specifically includes:

[0053] The image sensor registers are configured via a processor (ARM), and the configuration parameters are transmitted to the logic control unit (FPGA) via a bus interface.

[0054] The image sensor is driven by a logic control unit (FPGA) to write configuration information to control the sensor's operating status and acquisition parameters;

[0055] The pre-trained lightweight detection model weight parameters are stored in external memory.

[0056] In some exemplary embodiments, the image sensor registers are configured. Since the system uses a high-resolution, wide dynamic range sensor, the register configuration needs to be completed by the ARM, and the configuration parameters are transmitted to the FPGA via the AXI interface. The FPGA drives the image sensor and writes relevant configuration information through the control port, thereby controlling the sensor's operating state and acquisition parameters.

[0057] Before the system runs, the model data, such as the weight parameters, obtained from the training of the MobileNetV4-SSD spacecraft detection algorithm obtained from the host computer, need to be stored in the external memory of the ARM terminal.

[0058] In some embodiments, as an example, in step 2, the high-resolution image sensor receives the raw pixel data and sends the image data to the FPGA internal buffer through the image acquisition channel.

[0059] Furthermore, in one embodiment, the preprocessing in step 3 includes:

[0060] In the logic control unit (FPGA), the received raw target image data is processed by signal decoding, byte alignment, data width conversion, gamma correction, and noise reduction.

[0061] An IP core for a low-light enhancement algorithm for compressed images is generated using high-level synthesis HLS.

[0062] Preferably, in some embodiments, step 3 involves image preprocessing such as denoising and image enhancement, specifically including:

[0063] After the image sensor starts outputting images, the FPGA performs tasks such as receiving the received data, decoding the image signal, byte alignment, data width conversion, gamma correction, and noise reduction. In addition, HLS is used to convert the compressed image low-light enhancement model into VHDL or Verilog language to generate an IP core that can be synthesized in the FPGA, converting the raw data into a format that is easy for subsequent image processing. Then, the processed data is transmitted at high speed to the ARM end via the AXI bus, where the ARM is responsible for buffering the image frames and writing them into memory.

[0064] Furthermore, in one embodiment, the spacecraft detection algorithm described in step 4 uses MobileNetV4 as the backbone feature extraction network and integrates the SSD detection framework to obtain target detection results at different scales.

[0065] Preferably, in some embodiments, the MobileNetV4 backbone feature extraction network introduces a universal inverted bottleneck block (UIB) and a Mobile MQA attention block, and selects the output of a specific layer as the prediction feature layer to enhance the small target detection effect.

[0066] Preferably, in some embodiments, in the MobileNetV4 backbone feature extraction network, a CBAM module is connected after the feature extraction convolutional layer to enhance the feature representation capability. The CBAM module includes a channel attention module and a spatial attention module.

[0067] Preferably, in some embodiments, the spacecraft detection algorithm uses the K-means++ clustering algorithm to analyze the bounding boxes in the spacecraft dataset, generating multiple anchor box sizes and aspect ratios that match the dataset, to replace the default prior box settings (specifically, the K-means++ clustering algorithm is used to perform cluster analysis on the dimensions of the bounding boxes in the spacecraft dataset to generate multiple anchor box sizes that match the dataset, and then the anchor box aspect ratios are redesigned to replace the default anchor box settings).

[0068] Preferably, in some embodiments, the spacecraft detection algorithm performs model compression before inference, including pruning and quantizing the model, and performing point-to-point processing on the data.

[0069] Furthermore, in one embodiment, the target detection in step 4 is implemented on a (ZYNQ) heterogeneous platform, with the following specific architecture:

[0070] After the processing task scheduling and image frame buffering are completed on the processor (ARM) side;

[0071] An accelerated core (hardware accelerated IP core deployed on the FPGA) is deployed on the logic control unit to perform feature extraction, target regression classification prediction and image enhancement operations.

[0072] Preferably, in some embodiments, combined with Figure 2 and Figure 3 Step 4 involves detection using a spacecraft detection algorithm, specifically including:

[0073] Step 4-1: Create a spacecraft dataset;

[0074] Step 4-2: Replace the SSD detection framework backbone network with MobileNetV4 to extract image features and output feature maps at multiple scales.

[0075] Step 4-3: In the feature extraction stage, the CBAM module is introduced to enhance the feature representation capability;

[0076] Step 4-4: Use K-means++ clustering to generate anchor box sizes that match the spacecraft dataset instead of the default settings;

[0077] Steps 4-5: Input the multi-scale feature map into the SSD detection head for regression and classification prediction;

[0078] Steps 4-6 involve training and testing the model, and adjusting hyperparameters to optimize performance.

[0079] Steps 4-7 involve pruning and quantizing the model, then deploying it on the ZYNQ platform. After completing the processing tasks and scheduling on the ARM processor, feature extraction and SSD computing modules are deployed on the FPGA to accelerate neural network computation.

[0080] Furthermore, the specific process includes:

[0081] Step 4-1: Collect the spacecraft dataset, standardize its size and naming, and store it in VOC dataset format. Establish training, validation, and test sets, and use data annotation software to extract and label the spacecraft targets. The annotation tool used is, but not limited to, LabelImg. A bounding box is annotated on each spacecraft in each image, in VOC format, and the label file is an XML file. The annotated dataset is then divided into training, validation, and test sets for use in subsequent deep convolutional neural network object detection training.

[0082] Step 4-2: The MobileNetV4 network is used as the backbone feature extraction network. Based on this, the SSD algorithm detection head is fused to obtain target detection results at different scales. A total of 6 feature maps containing location and category information at 6 scales are extracted and used for the detection of targets of different sizes.

[0083] MobileNetV4 introduces the Universal Inverted Bottleneck (UIB) and Mobile MQA attention blocks, enhancing feature extraction and information fusion capabilities. This achieves a new balance between accuracy and efficiency, further optimizing mobile inference performance. The outputs of the ExtraDW block in layer 3 and the Conv2D block in layer 5 of the MobileNetV4 network are selected as the prediction feature layer to improve the detection performance of small targets and obtain more target location information.

[0084] Step 4-3 involves using a lightweight CBAM (Convolutional Block Attention Module) after the predicted feature maps output from the six feature extraction convolutional layers to optimize the algorithm's feature extraction capabilities and improve the accuracy of small object detection. CBAM includes a channel attention module and a spatial attention module. The processed feature maps are then input into the SSD multi-scale detection head for object regression and classification prediction.

[0085] Step 4-4: To build a robust model, the data augmentation method of this spacecraft detection algorithm is basically the same as that of the classic SSD method. The training dataset is expanded by randomly cropping and flipping the original data. To detect various spacecraft with different aspect ratios, the K-means++ clustering algorithm is used to cluster the bounding boxes in the spacecraft training set, generating six anchor box sizes. The aspect ratio of the anchor boxes is redesigned. The ZYNQ-based MobileNetV4-SSD spacecraft detection algorithm of this embodiment sets three default anchor boxes with different aspect ratios: 1, 2, and 1 / 2, replacing the default SSD anchor boxes, significantly improving the target detection rate and further reducing the number of model parameters.

[0086] Steps 4-5: Select the feature maps output by the CBAM module from Conv4_3, Conv7, Conv8_2, Conv9_2, Conv10_2, and Conv11_2 in the MobileNetV4-SSD network structure for prediction, reducing the feature map scale and the number of anchor boxes. The feature map sizes are 19×19, 10×10, 5×5, 3×3, 2×2, and 1×1, respectively, and the number of output channels are 160, 960, 512, 256, 256, and 128, respectively, reducing the number of algorithm parameters.

[0087] It should be noted that this spacecraft detection algorithm uses the same loss function as SSD, and regresses both position and target category simultaneously during training. Its loss function is the sum of confidence loss and position loss.

[0088] Steps 4-6: Build a target detection network model based on a deep convolutional neural network. Set the spacecraft as the detection target. Input the training set and validation set from the constructed dataset into the established target detection network model for training to obtain the weight file and detection model. Based on the target detection network prediction with the spacecraft as the target, generate prediction boxes and output the prediction box parameter information. Perform loss calculation on the model, continuously modify the parameters, and improve the output results of the target detection network.

[0089] Furthermore, regarding evaluation metrics, typical object detection metrics include detection accuracy, detection efficiency, and localization accuracy. This invention, however, focuses on object detection accuracy, detection efficiency, and deployment feasibility. Detection accuracy is evaluated using mAP (mean Average precision); detection efficiency is evaluated using FPS (frames per second); and deployment feasibility is evaluated using model size.

[0090] Steps 4-7 involve pruning and quantizing the model, then deploying it on the ZYNQ platform. The processing tasks and scheduling are handled on the ARM processor. Feature extraction and SSD computation modules are deployed on the FPGA to accelerate neural network computation. The specific process includes:

[0091] Before deploying the neural network on the hardware platform, the data is processed to be localized. The detection model is pruned and quantized to compress the model size, which reduces the storage space and multiplier requirements of the ZYNQ hardware platform, making the hardware implementation of the network model more feasible while retaining sufficient detection accuracy.

[0092] The DPU IP is integrated as a configurable module into the FPGA of the ZYNQ device and connected to the ARM via the AXI bus to perform deep learning inference tasks for spacecraft detection.

[0093] The ARM processor reads preprocessed RGB image data stored in memory and transmits it to the low-light enhancement module of the FPGA via the AXI bus for processing. After the system powers on, the weight parameters trained by the MobileNetV4-SSD spacecraft detection algorithm in the external memory of the ARM are loaded and transmitted to the FPGA for spacecraft detection on the preprocessed image. The detection results are then sent back and cached on the ARM.

[0094] Preferably, in some embodiments, step 5 outputs information such as the location and type of the spacecraft target. The specific process includes:

[0095] The Single Multiple Frame Detector (SSD) predicts the positions and categories of multiple spacecraft targets simultaneously in a single forward computation, thus achieving high real-time target detection performance. Its basic idea is to predict candidate anchor boxes of different scales and aspect ratios across the entire image using multi-layer convolutional feature maps, and then combine this with non-maximum suppression (NMS) to obtain the final detection result.

[0096] Furthermore, in one embodiment, in step 6, the data storage adopts a ping-pong buffer structure for data transmission, and the detection results and image data are transmitted to the host computer for display in real time via Ethernet.

[0097] Preferably, in some embodiments, step 6, which involves storing the image data and displaying it on the host computer, specifically includes the following process:

[0098] The images processed are stored in DDR memory, frame buffered, and managed using FIFO. For data storage, a ping-pong buffer structure is used for gapless data transfer. Two buffers maintain continuous data transfer between input and output, optimizing memory read speed, efficiently utilizing the waiting time required to load experimental data from external sources, and improving data throughput.

[0099] The images and spacecraft detection results are cached in DDR and then output to the host computer via gigabit Ethernet for real-time display and result recording.

[0100] In one embodiment, combined Figure 4 and Figure 5 This invention provides a low-power integrated system for spatial target image acquisition and real-time detection, the system comprising an image acquisition module, an image storage module, a lightweight image enhancement and spatial target recognition module, and an image transmission and display module;

[0101] The image sensor acquires images of spacecraft targets, transmits them to the FPGA via the acquisition channel, and performs image preprocessing after decoding and buffering.

[0102] The enhanced image is fed into a lightweight spacecraft detection algorithm for feature extraction and detection output. The detection results are transmitted to a host computer via Ethernet to achieve real-time image acquisition, processing, and display control.

[0103] Specific limitations regarding the integrated system for low-power spatial target image acquisition and real-time detection can be found in the limitations of the integrated method for low-power spatial target image acquisition and real-time detection described above, and will not be repeated here. Each module in the aforementioned integrated system for low-power spatial target image acquisition and real-time detection can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0104] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements:

[0105] Step 1: Initialize the image sensor parameters and load the model parameters;

[0106] Step 2: Acquire spacecraft target image data using an image sensor;

[0107] Step 3: Preprocess the acquired target image data;

[0108] Step 4: Use spacecraft detection algorithms to perform target detection on the preprocessed image;

[0109] Step 5: Output the location and category information of the spacecraft target;

[0110] Step 6: Store and display the image data and detection results.

[0111] For specific limitations on each step, please refer to the limitations of the integrated method for low-power spatial target image acquisition and real-time detection mentioned above, which will not be repeated here.

[0112] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being implemented when executed by a processor:

[0113] Step 1: Initialize the image sensor parameters and load the model parameters;

[0114] Step 2: Acquire spacecraft target image data using an image sensor;

[0115] Step 3: Preprocess the acquired target image data;

[0116] Step 4: Use spacecraft detection algorithms to perform target detection on the preprocessed image;

[0117] Step 5: Output the location and category information of the spacecraft target;

[0118] Step 6: Store and display the image data and detection results.

[0119] For specific limitations on each step, please refer to the limitations of the integrated method for low-power spatial target image acquisition and real-time detection mentioned above, which will not be repeated here.

[0120] The detection results of the method and system of this invention in low-light environment are shown in the figure below. Figure 6 As shown, without fully consuming the logical resources of each part of the system, the total power consumption of the system in completing tasks such as image acquisition, image enhancement, and spacecraft detection is only 7.302W, which is far lower than the power consumption of the GPU when performing only spacecraft detection tasks. This indicates that the system has excellent power consumption control and provides a longer continuous operating capability. The system can complete spacecraft detection of a single frame image within 9.80 milliseconds. In applications with strict power requirements, although the average accuracy of spacecraft detection is slightly lower than that of the GPU, its parallel processing and customization capabilities give it an advantage in specific tasks. In summary, this verifies that the system has the advantages of fast edge computing speed and low power consumption.

[0121] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention without departing from its spirit and scope should be included within the protection scope of the present invention.

Claims

1. A low-power integrated method for spatial target image acquisition and real-time detection, characterized in that, The method includes: Step 1: Initialize the image sensor parameters and load the model parameters; Step 2: Acquire spacecraft target image data using an image sensor; Step 3: Preprocess the acquired target image data; Step 4: Use spacecraft detection algorithms to perform target detection on the preprocessed image; Step 5: Output the location and category information of the spacecraft target; Step 6: Store and display the image data and detection results.

2. The integrated method for low-power spatial target image acquisition and real-time detection according to claim 1, characterized in that, Step 1 specifically includes: The processor configures the image sensor registers and transmits the configuration parameters to the logic control unit via the bus interface. The logic control unit drives the image sensor to write configuration information in order to control the sensor's working state and acquisition parameters; The pre-trained lightweight detection model weight parameters are stored in external memory.

3. The integrated method for low-power spatial target image acquisition and real-time detection according to claim 1, characterized in that, Step 3 preprocessing includes: In the logic control unit, the received raw target image data is processed by signal decoding, byte alignment, data width conversion, gamma correction, and noise reduction. An IP core for a low-light enhancement algorithm for compressed images is generated using high-level synthesis HLS.

4. The integrated method for low-power spatial target image acquisition and real-time detection according to claim 1, characterized in that, The spacecraft detection algorithm described in step 4 uses MobileNetV4 as the backbone feature extraction network and integrates the SSD detection framework to obtain target detection results at different scales.

5. The integrated method for low-power spatial target image acquisition and real-time detection according to claim 4, characterized in that, The MobileNetV4 backbone feature extraction network introduces a general inverted bottleneck block (UIB) and a Mobile MQA attention block, and selects the output of a specific layer as the prediction feature layer to enhance the small target detection effect.

6. The integrated method for low-power spatial target image acquisition and real-time detection according to claim 4, characterized in that, In the MobileNetV4 backbone feature extraction network, a CBAM module is connected after the feature extraction convolutional layer to enhance the feature representation capability. The CBAM module includes a channel attention module and a spatial attention module.

7. The integrated method for low-power spatial target image acquisition and real-time detection according to claim 4, characterized in that, The spacecraft detection algorithm uses the K-means++ clustering algorithm to analyze the bounding boxes in the spacecraft image dataset, generating multiple anchor box sizes and aspect ratios that match the dataset, to replace the default prior box settings.

8. The integrated method for low-power spatial target image acquisition and real-time detection according to claim 4, characterized in that, The spacecraft detection algorithm performs model compression before inference, including pruning and quantization of the model, and point-to-point processing of the data.

9. The integrated method for low-power spatial target image acquisition and real-time detection according to claim 2, characterized in that, In step 4, object detection is implemented on a heterogeneous platform. The specific architecture is as follows: On the processor side, complete the processing task scheduling and image frame buffering; An acceleration core is deployed on the logic control unit to perform feature extraction, target regression classification prediction, and image enhancement operations.

10. The integrated method for low-power spatial target image acquisition and real-time detection according to claim 1, characterized in that, In step 6, data storage uses a ping-pong buffer structure for data transmission, and the detection results and image data are transmitted to the host computer for display in real time via Ethernet.

Images