On-orbit real-time detection and identification method and device

Abstract

The present disclosure relates to a method and device for real-time detection and identification of targets on orbit. The method for real-time detection and identification of targets on orbit includes: based on the collected image containing the target, using an ultra-lightweight neural network to detect the target to extract potential target slices; splicing the extracted potential target slices into an image of a predetermined size by a quadtree method; using a deep neural network to identify the spliced ​​image and output the identification result. By adopting the present disclosure, it is possible to achieve real-time on-orbit target detection and identification of massive data when the computing power and storage resources of the satellite are limited, and use an ultra-lightweight neural network to perform target detection on low-resolution images. While improving the detection efficiency, most of the irrelevant data in the original image is filtered out to focus on potential targets; for high-resolution images, potential targets are centrally identified based on a deep learning network, which is equivalent to parallel identification of multiple slices, thereby further improving processing timeliness.

Description

Technical Field

[0001] The present disclosure relates to the field of on-orbit image processing and target detection and recognition, and more specifically, to a method and device for on-orbit real-time detection and recognition of targets. Background Art

[0002] Satellites will form a continuous large-scale imaging area during the swing or push-scan imaging process. With the continuous advancement of satellite-borne imaging technology, the spatial resolution of imaging is getting higher and higher, and the imaging width is getting larger and larger, which makes the amount of data generated by satellites in the continuous imaging process larger and larger. It is not realistic to fully store and download the massive image data continuously imaged and output by satellite-borne cameras. Especially in the case of sea imaging, due to the sparse distribution of targets (for example, ships), most of the imaging data is invalid data. In addition, due to the constraints of satellite size, power consumption, etc., the allocation of resources such as computing power and storage is relatively limited, resulting in a large gap in satellite-borne data processing capabilities compared to the ground. Summary of the invention

[0003] In order to solve the above problems, the present disclosure proposes a method and device for real-time on-orbit target detection and identification, a computing system and a computer-readable storage medium. With the help of on-orbit target detection and identification, a large amount of invalid data can be quickly filtered out, and the target of interest can be focused on, which can reduce the pressure of on-board storage and satellite-to-ground data transmission while improving the timeliness of information.

[0004] According to one aspect of the present disclosure, a method for on-orbit real-time detection and identification of targets is provided, the method comprising: based on a collected image containing the target, detecting the target using an ultra-lightweight neural network to extract potential target slices; splicing the extracted potential target slices into an image of a predetermined size using a quadtree method; and recognizing the spliced ​​image using a deep neural network and outputting a recognition result.

[0005] Optionally, based on the acquired image containing the target, the step of using an ultra-lightweight neural network to detect the target to extract potential target slices includes: downsampling the acquired image to a preset resolution; inputting the downsampled image into the ultra-lightweight neural network for detection to output the position and size of the potential target; and based on the position and size of the potential target, intercepting a potential target slice on the acquired image.

[0006] Optionally, the step of using a deep neural network to recognize the spliced ​​images and output the recognition results includes: inputting the spliced ​​images of a predetermined size into the deep neural network for recognition processing, and outputting the recognition results including at least one of a valid target identity, a confidence level and a slice.

[0007] Optionally, the potential target slice is a circumscribed square image including the complete target.

[0008] Optionally, the size of the circumscribed square image is determined based on a current image resolution and an actual target size.

[0009] Optionally, the acquired target-containing images, potential target slices, spliced ​​images, and recognition results are cached in different cache queues, respectively.

[0010] Optionally, the on-orbit real-time target detection and identification method further includes: acquiring an image containing the target by means of a satellite imaging payload.

[0011] According to another aspect of the present disclosure, a device for on-orbit real-time detection and identification of targets is provided, comprising: a potential target slicing unit, configured to detect the target based on a collected image containing the target using an ultra-lightweight neural network to extract potential target slices; an image stitching unit, configured to stitch the extracted potential target slices into an image of a predetermined size using a quadtree method; and an image recognition unit, configured to recognize the stitched image using a deep neural network and output a recognition result.

[0012] Optionally, the potential target slicing unit is configured to: downsample the captured image to a preset resolution; input the downsampled image into an ultra-lightweight neural network for detection to output the position and size of the potential target; and capture a potential target slice on the captured image based on the position and size of the potential target.

[0013] Optionally, the image recognition unit is configured to input the spliced ​​images of a predetermined size into a deep neural network for recognition processing, and output a recognition result including at least one of a valid target identity, a confidence level, and a slice.

[0014] Optionally, the potential target slice is a circumscribed square image including the complete target.

[0015] Optionally, the size of the circumscribed square image is determined based on a current image resolution and an actual target size.

[0016] Optionally, the acquired target-containing images, potential target slices, spliced ​​images, and recognition results are cached in different cache queues, respectively.

[0017] Optionally, images containing the target are acquired by a satellite imaging payload.

[0018] According to another aspect of the present disclosure, a computing system is provided that includes at least one computing device and at least one storage device storing instructions, wherein the instructions, when executed by the at least one computing device, prompt the at least one computing device to execute the on-orbit real-time detection and identification method for targets as described above.

[0019] According to another aspect of the present disclosure, a computer-readable storage medium storing instructions is provided, wherein when the instructions are executed by at least one computing device, the at least one computing device is prompted to execute the on-orbit real-time detection and identification method for targets as described above.

[0020] By adopting the present disclosure, it is possible to achieve on-orbit real-time target detection and recognition of massive data when the satellite's computing power and storage resources are limited, use an ultra-lightweight neural network to perform target detection on low-resolution images, and filter out most of the irrelevant data in the original image while improving the detection efficiency, so as to focus on potential targets; for high-resolution images, potential targets are centrally identified based on a deep learning network, which is equivalent to multi-slice parallel identification, thereby further improving processing timeliness. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The above and / or other objects and advantages of the present disclosure will become more apparent through the following description of embodiments in conjunction with the accompanying drawings, in which:

[0022] Figure 1 is a flow chart showing a method for real-time detection and identification of targets on track according to an exemplary embodiment of the present disclosure;

[0023] Figure 2 is a diagram showing an on-orbit real-time detection and identification process of sea surface targets based on massive data according to an exemplary embodiment of the present disclosure;

[0024] Figure 3 is a block diagram showing an on-orbit target real-time detection and identification device according to an exemplary embodiment of the present disclosure;

[0025] Figure 4 is a block diagram illustrating a computing system including at least one computing device and at least one storage device storing instructions according to an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0026] Below, in conjunction with the accompanying drawings, a description of a specific embodiment is provided to help the reader obtain a comprehensive understanding of the method, device and / or system described herein. However, after understanding the disclosure of the present application, various changes, modifications and equivalents of the method, device and / or system described herein will be clear. For example, the order of operations described herein is only an example and is not limited to those orders set forth herein, but can be changed as will be clear after understanding the disclosure of the present application, except for operations that must occur in a specific order. In addition, for greater clarity and simplicity, the description of features known in the art may be omitted.

[0027] Considering that the distribution of the targets of interest (e.g., ships) on remote sensing images is sparse and localized relative to the coverage of remote sensing images. In addition, target detection has low requirements for target feature extraction, while target identification requires rich feature information and a more complex network structure, and the processing speed is relatively slow. The present disclosure centrally stitches the target slices at the original resolution for identification, which is equivalent to identifying multiple targets in parallel, effectively improving the identification speed. In particular, for satellite remote sensing images with relatively sparse targets, the improvement effect is more obvious. Moreover, with the help of the cache mechanism, the situation where the targets are locally concentrated can be effectively handled.

[0028] Figure 1 Detailed description is a flow chart showing a method for detecting and identifying an on-orbit target in real time according to an exemplary embodiment of the present disclosure.

[0029] like Figure 1 As shown, in step S101, based on the collected image containing the target, an ultra-lightweight neural network (e.g., an ultra-lightweight deep convolutional network) is used to detect the target (e.g., a ship) to extract potential target slices. In the example, the image containing the target is collected by a satellite imaging payload. And in the example, the collected image is downsampled to a preset resolution; the downsampled image is input into the ultra-lightweight neural network for detection to output the position and size of the potential target; based on the position and size of the potential target, a potential target slice is intercepted on the collected image.

[0030] For example, the size of the original image block read at one time is 4096*4096, and the resolution is changed to 1024*1024 by downsampling 4 times, and the input size of the ultra-lightweight detection network is designed to be 1024*1024. In the example, the potential target slice is an circumscribed square image including the complete target. The ultra-lightweight neural network is used for target detection on low-resolution images to filter out most of the irrelevant data in the original image while improving the detection efficiency and focusing on the potential target.

[0031] In step S102, the extracted potential target slices are spliced ​​into an image of a predetermined size using a quadtree method. In step S103, a deep neural network is used to recognize the spliced ​​image and output a recognition result.

[0032] For example, potential target slices are spliced ​​into an image of a predetermined size; the image of the predetermined size is input into a deep neural network for recognition processing, and a recognition result including at least one of the valid target identity, confidence, and slice is output. For example, after receiving a potential target slice on an image block, multiple target slices are spliced ​​into an image of size H×H according to the quadtree method, and the spliced ​​image is input into the precise recognition network for recognition to detect the identity, confidence, slice, and other information of the recognized target. Quadtree slice splicing is used for high-resolution images, and potential targets are centrally identified based on a deep learning network, which is equivalent to multi-slice parallel recognition, further improving processing time efficiency. In addition, in the example, the collected image containing the target, the potential target slice, the spliced ​​image, and the recognition result are cached in different cache queues respectively.

[0033] According to the on-orbit real-time target detection and identification method of the exemplary embodiment of the present disclosure, starting from the sparsity and locality of the distribution of the target of interest in the remote sensing image, a two-stage parallel processing flow is designed to detect on a low-resolution image and then identify on a high-resolution image. In the first stage, a large amount of non-target data can be quickly filtered out, and in the second stage, batch fine identification is performed based on a small number of potential target slices, thereby realizing on-orbit real-time extraction and fine identification of targets based on massive data. In addition, the two processing flows can be coordinated with the designed data caching mechanism for parallel processing, providing a feasible way for on-orbit real-time target detection and identification of massive data.

[0034] Figure 2 is a diagram illustrating an on-orbit real-time detection and identification process of sea surface targets based on massive data according to an exemplary embodiment of the present disclosure.

[0035] The target detection and recognition method is divided into two steps: fast pre-check and precise identification for pipeline processing. Fast pre-check uses an ultra-lightweight detection network, the purpose of which is to quickly extract potential targets from massive data. In this step, a certain amount of false alarms can be included to avoid missed detections as much as possible, but accurate identity recognition is not performed. Accurate identification is performed on the basis of fast pre-check, using the pre-checked target slices to accurately identify the target type and eliminate false alarms. In order to ensure the real-time performance of the entire processing process, multi-threaded processing can be performed on one processor, or multiple processors can be used for parallel processing. Here, the method is described in a multi-threaded manner, and from the perspective of the integrity of the on-orbit processing flow and the operability of pipeline processing, in the example, potential targets of interest are quickly extracted from massive image data through pre-processing, and most of the invalid data is filtered out. Then, the quadtree splicing method is used to centrally identify potential targets, and based on multi-threaded parallelism, the pipeline processing of massive data is realized.

[0036] like Figure 2As shown in the figure, the entire on-orbit processing flow constructs four threads (i.e., image reading thread, data packaging and sending thread, fast pre-check thread, and precise identification thread) and three cache queues (i.e., image cache queue, slice cache queue, and result data cache queue). The image reading thread reads the camera continuous imaging data in real time and transmits the read data to the image cache queue; the fast pre-check thread reads the image from the image cache queue for fast target detection to extract potential target slices, and puts the extracted potential target slices into the target slice cache queue; the precise identification thread obtains slice data from the target slice cache queue for target identity recognition, and puts the recognition results and slice data into the result data queue; the data packaging and sending thread packages the data in the result data queue and sends it to the router or saves it to the memory.

[0037] For example, four threads can run synchronously, and the complete process includes: step one, constructing the above four threads and three cache queues; step two, if there is an image passed in, the image reading thread reads the image block of a predetermined size in real time, and transfers the read image block to the image cache queue; step three, if the image cache queue is not empty, the fast pre-check thread reads the head image from the image cache queue, and the head image is dequeued; step four, downsampling the acquired image to a preset resolution (mainly for preprocessing of the current high-resolution image. For target detection of ships, etc., a high resolution is not required. By detecting targets on lower resolution images that ensure that the target is detectable, the amount of processed data can be reduced, thereby improving the detection speed); step five, inputting the downsampled image into the trained fast pre-check network for target detection, outputting the potential target position and size, and according to the potential target position and size Small, intercept the target slice on the original image, store all the intercepted potential target slices in the target slice queue, and then the fast pre-check thread returns to step three; step six, if the target slice queue is not empty, the precise identification thread reads the target slices from the target slice queue in sequence, and the slices are dequeued at the same time; step seven, the precise identification thread splices the read slices into an image of a predetermined size using a quadtree method, and inputs the spliced ​​image into the precise identification network for identification processing, outputs identification information, and stores the valid target identity, confidence, slice and other information in the result data cache queue, and then the precise identification thread returns to step five; step eight, if the result data cache queue is not empty, the data packaging and sending thread reads the identification result information from the result data cache queue in sequence, and packages the result information in a predetermined format and sends it to the routing or storage unit, and the read result information is dequeued.

[0038] The above step 2 belongs to the main operation of the image reading thread, step 8 belongs to the main operation of the data sending thread, steps 3 to 5 belong to the main operation of the fast pre-check thread, and steps 6 and 7 belong to the main operation of the accurate recognition thread. Since the functions of the image reading thread and the data sending thread are relatively simple, they will not be described in detail here. The following focuses on the fast pre-check thread and the accurate recognition thread.

[0039] In step five, the input image size received by the fast pre-check network is determined according to the resolution of the onboard camera and the onboard processor capability. For example, the input image size received by the fast pre-check network can be designed to be the size of the image block after downsampling. The size of the original image block read at one time is 4096*4096. By downsampling by 4 times, the resolution becomes 1024*1024, and the input size of the ultra-lightweight detection network can be designed to be 1024*1024. Using the trained ultra-lightweight detection network, the potential target position and size are output, and according to the output information (i.e., the potential target position and size), the target slice is intercepted on the original resolution image, and all detected potential target slices are sent to the precise identification thread, and then the fast pre-check thread returns to step two. For example, the target slice is an circumscribed square image containing the complete target, and its size is , , where s1 and s2 are determined according to the current image resolution and the actual target size, and target slices of different sizes are allowed to be included to meet the needs of targets of different sizes, for example, n=6, 7, 8, 9, 10. In addition, the minimum circumscribed square of the target can be expanded outward to a certain size to include a part of the background, and target slices of different sizes are allowed to be included. In addition, for ultra-lightweight detection networks, the model size can be reduced to the mega or K level based on the yolo-tiny series model through pruning, quantization, distillation and other technologies according to actual processing needs.

[0040] In step 7, the precise recognition network is a pre-trained detection recognition network, and the required input image size is H×H according to the target slice size. = ,s , for example, s=s2=10. The input image size of the precise recognition network can also be determined according to the design of the onboard processor capability. After the fine recognition thread receives the potential target slices on an image block, it splices multiple target slices into an image of size H×H according to the quadtree method, and inputs the spliced ​​image into the precise recognition network for recognition to detect the identity, confidence, slice and other information of the recognized target.

[0041] In the example, the specific splicing process is as follows:

[0042] 1) Establish a Empty image of , and based on Construct the quadtree T as follows: As the root node, the image is evenly divided into four Non-intersecting areas , , as . And so on, the node The four child nodes are denoted as , The leaf nodes are , and mark the extra nodes as unfilled.

[0043] 2) If the tile cache queue is not empty, read the image tiles in the tile cache queue in order , ;

[0044] 3) For slicing , set its width to , traverse the sdth layer nodes of T, if there is a node that is not filled, then Fill the area corresponding to the node, mark the node and all its ancestor nodes as filled, and return to step 2), otherwise Not fillable , proceed to step 4);

[0045] 4) Input to the precise identification network for identity recognition, store the identification results and slice information output by the network into the result data cache queue, and return to step 1).

[0046] It should be noted that in step 4) It does not need to be filled, but when the current slice cannot be filled, The input is sent to the precise identification network for identification, which ensures uninterrupted pipeline processing during the identification process. In this example, considering that the ship features are single and the types are limited, a network model of the yolo4-tiny level can be used to meet the needs of on-orbit real-time processing.

[0047] The present disclosure does not impose specific restrictions on the fast pre-check network and fine recognition network in step 5 and step 7, as long as they meet the input image size and recognition type. For example, a suitable network structure can be adopted or constructed according to the data volume, chip processing capability and real-time requirements, such as a yolo series network, an SSD (single-shot multi-box detector) network, etc. However, it should be clear that the fast pre-check network must be an ultra-lightweight detection network that can meet the real-time pipeline processing requirements of the acquisition frame; the fine recognition network can be larger than the fast pre-check network to extract richer target features for identity recognition. In addition, in order to avoid missed detection as much as possible in the first stage, a certain false alarm is allowed, and the detection threshold can be set relatively low. In addition, the on-orbit real-time detection and recognition process of sea surface targets based on massive data according to the exemplary embodiment of the present disclosure is not only suitable for single-processor multi-threaded processing, but also suitable for multi-processor parallel processing. For example, the SoC (ARM+FPGA)+GPU dual-processor mode can be used, and the image reading thread and the fast pre-check thread are suitable for processing on the SoC, especially the fast pre-check network is implemented by FPGA, and the accurate recognition thread and the package sending thread can be processed on the GPU.

[0048] According to the exemplary embodiment of the present disclosure, the on-orbit real-time detection and identification process of sea surface targets based on massive data utilizes two-stage parallel processing for detection and identification, which can realize real-time target detection and identification of massive wide-width image sequences on-orbit.

[0049] Figure 3 is a block diagram showing an on-orbit target real-time detection and identification device according to an exemplary embodiment of the present disclosure.

[0050] like Figure 3 As shown, the on-orbit real-time detection and identification device for targets according to an exemplary embodiment of the present disclosure includes: a potential target slicing unit 301, configured to detect the target based on the acquired image containing the target by using an ultra-lightweight neural network to extract potential target slices; an image stitching unit 302, configured to stitch the extracted potential target slices into an image of a predetermined size by using a quadtree method; an image recognition unit 303, configured to recognize the stitched image by using a deep neural network and output a recognition result.

[0051] In the example, the potential target slicing unit 301 is configured to: downsample the captured image to a preset resolution; input the downsampled image into an ultra-lightweight neural network for detection to output the position and size of the potential target; and based on the position and size of the potential target, capture a potential target slice on the captured image.

[0052] In the example, the image recognition unit 303 is configured to input the spliced ​​images of a predetermined size into a deep neural network for recognition processing, and output a recognition result including at least one of a valid target identity, a confidence level, and a slice.

[0053] In an example, a potential target slice is a bounding square image that includes the entire target.

[0054] In an example, the size of the circumscribed square image is determined based on the current image resolution and the actual target size.

[0055] In the example, the acquired target-containing images, potential target slices, stitched images, and recognition results are cached in different cache queues, respectively.

[0056] In an example, images containing a target are acquired by a satellite imaging payload.

[0057] Combination of the above Figure 1 and Figure 2 The specific operations shown are respectively Figure 3 The operation is performed by corresponding units in the target on-orbit real-time detection and identification device 300 shown in the figure, and the specific operation details will not be repeated here.

[0058] Figure 4 is a block diagram illustrating a computing system including at least one computing device and at least one storage device storing instructions according to an exemplary embodiment of the present disclosure.

[0059] like Figure 4 As shown, a computing system 400 provided according to an exemplary embodiment of the present disclosure includes a computing device 401 and a storage device 402, wherein the storage device 402 stores computer executable instructions. When the computer executable instructions are executed by the computing device 401, the on-orbit real-time detection and identification method for targets described in any of the aforementioned embodiments is executed.

[0060] The computing device 401 is deployed in a server or client, and can also be deployed on a node device in a distributed network environment. In addition, the computing device 401 can be a PC computer, a tablet device, a personal digital assistant, a smart phone, a web application or other device capable of executing the above-mentioned instruction set. Here, the computing device is not necessarily a single computing device, but can also be any device or circuit that can execute the above-mentioned instructions (or instruction sets) individually or jointly. The computing device can also be part of an integrated control system or a system manager, or can be configured as a portable electronic device that is interconnected with an interface locally or remotely (for example, via wireless transmission). In the computing device, the processor includes a central processing unit (CPU), a graphics processing unit (GPU), a programmable logic device, a dedicated processor system, a microcontroller or a microprocessor. As an example and not a limitation, the processor also includes an analog processor, a digital processor, a microprocessor, a multi-core processor, a processor array, a network processor, etc.

[0061] According to another aspect of the present disclosure, a computer-readable storage medium storing instructions is provided, and when the instructions are executed by at least one computing device, the at least one computing device is prompted to execute the on-orbit real-time detection and identification method of the target described in any of the aforementioned embodiments. Computer-readable storage media include magnetic media such as floppy disks and tapes, optical media (including compact disk (CD) ROM and DVD ROM), magneto-optical media such as floppy disks, hardware devices such as ROM, RAM, and flash memory designed to store and execute program commands. The instructions may include language codes executable by a computer using an interpreter and machine language codes generated by a compiler.

[0062] By adopting the present disclosure, it is possible to achieve on-orbit real-time target detection and recognition of massive data when the satellite's computing power and storage resources are limited, use an ultra-lightweight neural network to perform target detection on low-resolution images, and filter out most of the irrelevant data in the original image while improving the detection efficiency, so as to focus on potential targets; for high-resolution images, potential targets are centrally identified based on a deep learning network, which is equivalent to multi-slice parallel identification, thereby further improving processing timeliness.

[0063] The processes, methods or algorithms disclosed herein may be transmitted to or implemented by a processing device, a controller or a computer, which may include any existing programmable electronic control unit or a dedicated electronic control unit. Similarly, the processes, methods or algorithms may be stored in a variety of forms as data and instructions that can be executed by a controller or a computer, including but not limited to information being permanently stored on a non-writable storage medium (such as a ROM device) and information being variably stored on a writable storage medium (such as a floppy disk, a tape, a CD, a RAM device, and other magnetic and optical media). The processes, methods or algorithms may also be implemented in a software executable object. Optionally, the processes, methods or algorithms may be implemented in whole or in part using suitable hardware components (such as an ASIC, an FPGA, a state machine, a controller or other hardware components or devices) or a combination of hardware components, software components and firmware components.

[0064] Although the present disclosure includes specific examples, it will be apparent to those of ordinary skill in the art that various changes in form and detail may be made in these examples without departing from the spirit and scope of the claims and their equivalents. The examples described herein will be considered to be descriptive only and not for limiting purposes. The description of the features or aspects in each example will be considered to be applicable to similar features or aspects in other examples. Suitable results may be obtained if the described techniques are performed in a different order, and / or if the components in the described systems, architectures, devices, or circuits are combined in different ways and / or replaced or supplemented with other components or their equivalents in the described systems, architectures, devices, or circuits. Therefore, the scope of the present disclosure is not limited by specific embodiments, but by the claims and their equivalents, and all variations within the scope of the claims and their equivalents will be interpreted as included in the present disclosure.

Claims

1. A method for real-time detection and identification of targets on orbit, characterized in that: The on-orbit real-time detection and identification method of the target comprises: Based on the acquired image containing the target, the target is detected by using an ultra-lightweight neural network to extract potential target slices, wherein the potential target slice is a circumscribed square image including the complete target, and the size of the circumscribed square image is determined based on the current image resolution and the actual target size to allow potential target slices of different sizes to be included; The extracted potential target slices are stitched into an image of a predetermined size by a quadtree method; Use deep neural networks to recognize the spliced ​​images and output the recognition results. The steps of detecting the target using an ultra-lightweight neural network based on the acquired image containing the target to extract potential target slices include: Downsample the acquired image to a preset resolution; The downsampled image is input into the ultra-lightweight neural network for detection to output the location and size of potential targets; Based on the position and size of the potential target, a potential target slice is captured on the acquired image.

2. The method for real-time detection and identification of targets on track according to claim 1, characterized in that: The steps of using a deep neural network to recognize the spliced ​​images and output the recognition results include: The spliced ​​images of a predetermined size are input into a deep neural network for recognition processing, and a recognition result including at least one of a valid target identity, a confidence level, and a slice is output.

3. The method for real-time detection and identification of targets on track according to claim 1, characterized in that: The acquired images containing the target, potential target slices, spliced ​​images and recognition results are cached in different cache queues respectively.

4. The method for real-time detection and identification of targets on track according to claim 1, characterized in that: The on-orbit real-time detection and identification method for a target also includes: collecting an image containing the target by using a satellite imaging payload.

5. A device for real-time detection and identification of targets on track, characterized in that: The target on-orbit real-time detection and identification device comprises: A potential target slicing unit is configured to detect the target based on the acquired image containing the target by using an ultra-lightweight neural network to extract potential target slices, wherein the potential target slice is a circumscribed square image including the complete target, and the size of the circumscribed square image is determined based on the current image resolution and the actual target size to allow potential target slices of different sizes to be included; an image stitching unit configured to stitch the extracted potential target slices into an image of a predetermined size by a quadtree method; The image recognition unit is configured to recognize the spliced ​​images using a deep neural network and output a recognition result. Among them, the potential target slicing unit is further configured to: downsample the captured image to a preset resolution; input the downsampled image into an ultra-lightweight neural network for detection to output the position and size of the potential target; and based on the position and size of the potential target, capture a potential target slice on the captured image.

6. A computing system comprising at least one computing device and at least one storage device storing instructions, characterized in that: When the instructions are executed by the at least one computing device, the at least one computing device is prompted to execute the on-orbit real-time detection and identification method for targets according to any one of claims 1-4.

7. A computer-readable storage medium storing instructions, characterized in that: When the instructions are executed by at least one computing device, the at least one computing device is prompted to execute the on-orbit real-time detection and identification method for targets according to any one of claims 1-4.

Images