On-orbit intelligent detection method and device for rotating and sweeping ultra-wideband satellites

Abstract

In the method, device and storage medium for on-orbit intelligent detection of targets for rotating and sweeping ultra-wide-band satellites proposed in this application, a lightweight target detection model is constructed, and the lightweight target detection model is tuned using a historical remote sensing image data set to obtain an optimal lightweight target detection model, and then the acquired remote sensing image to be detected is passed through the lightweight feature extraction backbone network in the optimal lightweight target detection model to obtain a backbone network output feature map, and then the backbone network output feature map is subjected to multiple different convolution operations to obtain feature maps of multiple scales, and the feature maps of multiple scales are respectively passed through the corresponding detector classifiers to obtain corresponding multiple preset frames, and the multiple preset frames are subjected to rapid maximum suppression processing to obtain the target detection result. The method proposed in this application does not need to rely on expert experience, reduces the amount of parameters and calculations, reduces the model size and computing power consumption, improves the efficiency of satellite on-orbit target detection, and realizes rapid detection of targets by satellites.

Description

Technical Field

[0001] The present application relates to the technical field of rotating swing-sweep ultra-wideband satellite target detection, and in particular to an on-orbit intelligent target detection method, device and storage medium for rotating swing-sweep ultra-wideband satellites. Background Art

[0002] With the rapid development of earth observation technology, satellites in orbit with different widths and resolutions in high, medium and low orbits together constitute an earth observation network. Among them, ultra-wide-width remote sensing satellites adopt a new ultra-wide coverage imaging system with vertical track scanning and along-track splicing. The satellite orbit altitude is 500km, the sub-satellite spatial resolution is 1 meter, and the imaging width is 3000km, which can achieve full coverage of the earth's surface in a single day. However, the amount of data per unit time of remote sensing images acquired by ultra-wide-width satellites is as high as 420Gbps, which far exceeds the data transmission bandwidth and storage space of ultra-wide-width satellites. Therefore, ultra-wide-width satellites cannot transmit remote sensing images to the ground for processing, so ultra-wide-width satellites are required to detect targets in remote sensing images.

[0003] In the related technologies, satellite target detection methods based on artificially designed features mainly select the area of ​​interest in the image based on sliding window operations, which lacks pertinence and easily causes window redundancy and high time complexity. In addition, feature extraction is subsequently performed based on artificially designed features, where artificially designed features rely on the experience of experts and have low robustness. Satellite target detection methods based on deep learning require a large amount of memory and computing resources, which cannot be supported by the onboard resources of ultra-wide satellites. Summary of the invention

[0004] The present application provides a method, device and storage medium for on-orbit intelligent detection of targets for rotating swing-scan ultra-wideband satellites, so as to at least solve the technical problems in related technologies that rely on expert experience and cannot be supported by on-board resources.

[0005] The first aspect of the present application provides an on-orbit intelligent detection method for a rotating swing-scan ultra-wideband satellite, the method comprising:

[0006] Constructing a lightweight target detection model, wherein the lightweight target detection model includes a lightweight feature extraction backbone network;

[0007] Use historical remote sensing image datasets to tune the lightweight target detection model and obtain the optimal lightweight target detection model;

[0008] Inputting the acquired remote sensing image to be detected into the optimal lightweight target detection model, and obtaining a backbone network output feature map through the lightweight feature extraction backbone network in the optimal lightweight target detection model;

[0009] The output feature map of the backbone network is subjected to multiple convolution operations to obtain feature maps of multiple scales;

[0010] The feature maps of multiple scales are respectively passed through the corresponding detector classifiers to obtain corresponding multiple preset frames, and the multiple preset frames are subjected to fast maximum value suppression processing to obtain the target detection result.

[0011] The second aspect of the present application provides an on-orbit intelligent detection device for a rotating swing-scan ultra-wideband satellite, comprising:

[0012] A construction module, used to construct a lightweight target detection model, wherein the lightweight target detection model includes a lightweight feature extraction backbone network;

[0013] The tuning module is used to tune the lightweight target detection model using historical remote sensing image datasets to obtain the optimal lightweight target detection model;

[0014] A first processing module is used to input the acquired remote sensing image to be detected into an optimal lightweight target detection model, and obtain a backbone network output feature map through a lightweight feature extraction backbone network in the optimal lightweight target detection model;

[0015] A second processing module is used to obtain feature maps of multiple scales by subjecting the output feature map of the backbone network to multiple convolution operations;

[0016] The output module is used to pass the feature maps of multiple scales through the corresponding detector classifiers to obtain corresponding multiple preset boxes, and perform fast maximum suppression processing on the multiple preset boxes to obtain the target detection result.

[0017] The computer storage medium proposed in the third aspect of the present application, wherein the computer storage medium stores computer-schedulable instructions; after the computer-schedulable instructions are scheduled by the processor, the method described in the first aspect above can be implemented.

[0018] The computer device proposed in the fourth aspect of the present application includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the method described in the first aspect above can be implemented.

[0019] The mobile terminal device proposed in the fifth aspect of the present application is characterized in that it includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it can implement the method described in the first aspect above.

[0020] The technical solution provided by the embodiments of the present application brings at least the following beneficial effects:

[0021] In the method, device and storage medium for on-orbit intelligent detection of targets for rotating swing-sweep ultra-wideband satellites proposed in this application, a lightweight target detection model is constructed, and the lightweight target detection model is tuned using a historical remote sensing image data set to obtain an optimal lightweight target detection model, and then the acquired remote sensing image to be detected is input into the optimal lightweight target detection model, and the backbone network output feature map is obtained through the lightweight feature extraction backbone network in the optimal lightweight target detection model, and then the backbone network output feature map is subjected to multiple different convolution operations to obtain feature maps of multiple scales, and the feature maps of multiple scales are respectively passed through the corresponding detector classifiers to obtain corresponding multiple preset frames, and the multiple preset frames are subjected to fast maximum suppression processing to obtain the target detection result. It can be seen from this that the method proposed in this application obtains the optimal lightweight target detection model through training of the historical remote sensing image data set, so that there is no need to rely on expert experience, and target detection can also be performed on the remote sensing image to be detected, reducing the number of parameters and calculations, and reducing the model size and computing power consumption. At the same time, the lightweight target detection model adopts a single-stage detection framework, and the lightweight computing unit of the lightweight feature extraction backbone network adopts group depthwise separable convolution, thereby meeting the limited on-board resources, improving the efficiency of satellite on-orbit target detection, and thus realizing ultra-wide-band satellite rapid detection of targets.

[0022] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0024] Figure 1 A schematic diagram of a flow chart of an on-orbit intelligent detection method for a rotating swing-scan ultra-wideband satellite according to an embodiment of the present application;

[0025] Figure 2 A schematic diagram of a process of group depth-separable convolution according to an embodiment of the present application;

[0026] Figure 3 A schematic diagram of a flow chart of an on-orbit intelligent target detection method based on an optimal lightweight target detection model provided according to an embodiment of the present application;

[0027] Figure 4 The present invention is a schematic diagram of the structure of an on-orbit intelligent detection device for a rotating and sweeping ultra-wideband satellite according to an embodiment of the present application. DETAILED DESCRIPTION

[0028] Embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0029] The following describes the on-orbit intelligent detection method and device for rotating and sweeping ultra-wide-band satellites according to an embodiment of the present application with reference to the accompanying drawings.

[0030] Embodiment 1

[0031] Figure 1 The flowchart of the method for intelligent on-orbit detection of a target for a rotating swing-scan ultra-wideband satellite provided according to an embodiment of the present application is as follows: Figure 1 As shown, it may include:

[0032] Step 101: Build a lightweight target detection model.

[0033] Among them, in the present application, the lightweight target detection model may include a lightweight feature extraction backbone network.

[0034] Also, in this application, the lightweight target detection model uses a single-stage detection framework, so that the candidate box and target category can be calculated simultaneously, improving the detection efficiency.

[0035] Specifically, in the present application, the above-mentioned single-stage detection framework can be an SSD framework or a YOLO framework.

[0036] Step 102: Use the historical remote sensing image dataset to tune the lightweight target detection model to obtain the optimal lightweight target detection model.

[0037] Among them, in this application, tuning the lightweight target detection model through the historical remote sensing image dataset can include the following steps:

[0038] Step a: preprocess the historical remote sensing image dataset.

[0039] Among them, in the present application, the preprocessing of the historical remote sensing image data set may include data enhancement (for example, rotating the image angle, adding white noise, etc.) of the data in the historical remote sensing image data set.

[0040] Step b: divide the preprocessed remote sensing image dataset into a training set and a validation set.

[0041] Step c: input the training set into the lightweight target detection model and calculate the loss function of the lightweight target detection model.

[0042] Step d: According to the loss value of the loss function, the lightweight target detection model is tuned using the validation set to obtain the optimal lightweight target detection model.

[0043] In addition, it should be noted that this application uses historical remote sensing image data sets in the ground data center to tune the lightweight target detection model, so that the lightweight target detection model can be tuned using the resources on the ground to improve the efficiency of tuning the lightweight target detection model. And, after obtaining the optimal lightweight target detection model, the optimal lightweight target detection model is deployed on the ultra-wide satellite, so that the ultra-wide satellite can perform intelligent target detection on the acquired remote sensing images in orbit, solving the problem that "ultra-wide satellites cannot transmit remote sensing images to the ground for processing".

[0044] Among them, in this application, when the optimal lightweight target detection model is deployed on the ultra-wide satellite, Nvidia accelerated inference card hardware is used, and TensorRT is used for model pruning, quantization and compression, which reduces the space occupied on the ultra-wide satellite and improves computing power.

[0045] Step 103: input the acquired remote sensing image to be detected into the optimal lightweight target detection model, and obtain the backbone network output feature map through the lightweight feature extraction backbone network in the optimal lightweight target detection model.

[0046] Among them, in the present application, the lightweight feature extraction backbone network may include a multi-scale feature fusion module and an original feature enhancement module.

[0047] And, in the present application, the method of inputting the acquired remote sensing image to be detected into the optimal lightweight target detection model, and obtaining the backbone network output feature map through the lightweight feature extraction backbone network in the optimal lightweight target detection model may include the following steps:

[0048] Step 1: Input the acquired remote sensing image to be detected into the multi-scale feature fusion module and the original feature enhancement module in the optimal lightweight target detection model in parallel, and obtain the feature map of the multi-scale feature fusion module and the feature map of the original feature enhancement module respectively.

[0049] Step 2: Fuse the feature map of the multi-scale feature fusion module and the feature map of the original feature enhancement module to obtain the backbone network output feature map.

[0050] Furthermore, in the present application, the multi-scale feature fusion module and the original feature enhancement module respectively include N lightweight computing units connected in series, where N is a positive integer.

[0051] Among them, in the present application, the method of inputting the acquired remote sensing image to be detected into the multi-scale feature fusion module and the original feature enhancement module in the optimal lightweight target detection model in parallel to obtain the feature map of the multi-scale feature fusion module and the feature map of the original feature enhancement module respectively may include the following steps:

[0052] Step 1: Input the acquired remote sensing image to be detected into the multi-scale feature fusion module in the optimal lightweight target detection model. The input remote sensing image to be detected is the scale one feature. Then, the remote sensing image to be detected is input into the first lightweight computing unit to obtain the scale two feature. The obtained scale two feature is input into the second lightweight computing unit to obtain the scale three feature. And so on. The obtained scale features are sequentially passed through the remaining N-2 lightweight computing units to obtain the corresponding N-2 scale features. Then, the obtained N scale features are merged to obtain the feature map of the multi-scale feature fusion module.

[0053] Step 2: Input the acquired remote sensing image to be detected into the original feature enhancement module in the optimal lightweight target detection model, input the input remote sensing image to be detected into the first lightweight computing unit to obtain the first feature, input the remote sensing image to be detected and the first feature into the second lightweight computing unit to obtain the second feature, and so on, the remote sensing image to be detected and the obtained features are sequentially passed through the remaining N-2 series lightweight computing units to obtain the Nth feature, which is the feature map of the original feature enhancement module.

[0054] And, in this application, the lightweight computing unit is formed by group depthwise separable convolution, wherein the group depthwise separable convolution is to stack multiple layers of depthwise convolution and then uniformly perform pointwise convolution. Figure 2 A schematic diagram of a process flow of a group depth-separable convolution provided for this application.

[0055] Specifically, in this application, it is assumed that the number of group depth convolutions included in the depth convolution is: J, and the size of the input remote sensing image to be detected is: D F ×D F ×M, the convolution kernel size is: D K ×D K ×M, the output feature map size of the backbone network is: D F ×D F ×N.

[0056] The number of parameters of the depth convolution is: (D K ×D K ×1)×M,

[0057] The number of point-by-point convolution parameters is: (1×1×M)×N,

[0058] The number of parameters of depth-wise separable convolution is: D K ×D K×M+M×N, the number of parameters for J depth-wise separable convolutions is: J×(D K ×D K ×M+M×N),

[0059] The number of parameters for the group depth-wise separable convolution is: J×D K ×D K ×M+M×N, the ratio of the number of parameters of the depthwise separable convolution to the number of parameters of the group depthwise separable convolution is: Therefore, the group depth-wise separable convolution achieves the goal of being lightweight.

[0060] Step 104: The feature map output by the backbone network is subjected to multiple convolution operations to obtain feature maps of multiple scales.

[0061] Among them, in this application, different lightweight target detection models correspond to different numbers of convolution operations.

[0062] Step 105: pass the feature maps of multiple scales through the corresponding detector classifiers respectively to obtain corresponding multiple preset boxes, and perform fast maximum suppression processing on the multiple preset boxes to obtain the target detection result.

[0063] In this application, the target detection result may include at least one of target slice, target category, confidence, and coordinate information. Also, in this application, after the ultra-wide satellite obtains the target detection result, it returns the obtained target detection result to the ground. At this time, the data volume of the target detection result is less than the data transmission bandwidth of the ultra-wide satellite, so that the target detection result can be returned to the ground through the ultra-wide satellite.

[0064] Based on the above content, the present application provides examples for the above steps 103 to 105 based on the obtained optimal lightweight target detection model.

[0065] Figure 3 The following is a flow chart of an on-orbit intelligent detection method for a target based on an optimal lightweight target detection model according to an embodiment of the present application. In this application, the optimal lightweight target detection model uses the SSD framework. Figure 3 As shown in the figure, the acquired remote sensing image to be detected is input into the multi-scale feature fusion module and the original feature enhancement module in the optimal lightweight target detection model in parallel, and the feature map of the multi-scale feature fusion module and the feature map of the original feature enhancement module are fused to obtain the backbone network output feature map. Among them, the multi-scale feature fusion module and the original feature enhancement module each include 3 lightweight computing units connected in series.

[0066] Specifically, Figure 3As shown, in the present application, the acquired remote sensing image to be detected is input into the multi-scale feature fusion module in the optimal lightweight target detection model, the input remote sensing image to be detected is the scale one feature, and then the remote sensing image to be detected is input into the first lightweight computing unit to obtain the scale two feature, the obtained scale two feature is input into the second lightweight computing unit to obtain the scale three feature, the obtained scale one feature, scale two feature, and scale three feature are merged to obtain the feature map of the multi-scale feature fusion module.

[0067] Also, in the present application, the acquired remote sensing image to be detected is input into the original feature enhancement module in the optimal lightweight target detection model, the input remote sensing image to be detected is input into the first lightweight computing unit to obtain the first feature, the remote sensing image to be detected and the first feature are input into the second lightweight computing unit to obtain the second feature, and then the remote sensing image to be detected and the second feature are input into the third lightweight computing unit to obtain the third feature, and the third feature is the feature map of the original feature enhancement module.

[0068] Furthermore, in the present application, the obtained backbone network output feature map is input through 5 different output channel convolution operations to obtain corresponding 6-scale feature maps, and the 6-scale feature maps are respectively passed through the corresponding detector classifiers to obtain corresponding 6 preset boxes and classification results, and then the 6 preset boxes are fast maximum suppression processed to obtain the target detection result.

[0069] In the method, device and storage medium for on-orbit intelligent detection of targets for rotating swing-sweep ultra-wideband satellites proposed in this application, a lightweight target detection model is constructed, and the lightweight target detection model is tuned using a historical remote sensing image data set to obtain an optimal lightweight target detection model, and then the acquired remote sensing image to be detected is input into the optimal lightweight target detection model, and the backbone network output feature map is obtained through the lightweight feature extraction backbone network in the optimal lightweight target detection model, and then the backbone network output feature map is subjected to multiple different convolution operations to obtain feature maps of multiple scales, and the feature maps of multiple scales are respectively passed through the corresponding detector classifiers to obtain corresponding multiple preset frames, and the multiple preset frames are subjected to fast maximum suppression processing to obtain the target detection result. It can be seen from this that the method proposed in this application obtains the optimal lightweight target detection model through training of the historical remote sensing image data set, so that there is no need to rely on expert experience, and target detection can also be performed on the remote sensing image to be detected, reducing the number of parameters and calculations, and reducing the model size and computing power consumption. At the same time, the lightweight target detection model adopts a single-stage detection framework, and the lightweight computing unit of the lightweight feature extraction backbone network adopts group depthwise separable convolution, thereby meeting the limited on-board resources, improving the efficiency of satellite on-orbit target detection, and thus realizing ultra-wide-band satellite rapid detection of targets.

[0070] Embodiment 2

[0071] Further, Figure 4 FIG. 1 is a schematic diagram of the structure of an on-orbit intelligent detection device for a rotating swing-scan ultra-wideband satellite according to an embodiment of the present application. Figure 4 As shown, it may include:

[0072] A construction module 401 is used to construct a lightweight target detection model, wherein the lightweight target detection model includes a lightweight feature extraction backbone network;

[0073] A tuning module 402 is used to tune the lightweight target detection model using the historical remote sensing image data set to obtain an optimal lightweight target detection model;

[0074] The first processing module 403 is used to input the acquired remote sensing image to be detected into the optimal lightweight target detection model, and obtain a backbone network output feature map through the lightweight feature extraction backbone network in the optimal lightweight target detection model;

[0075] The second processing module 404 is used to obtain feature maps of multiple scales by performing multiple convolution operations on the feature map output by the backbone network;

[0076] The output module 405 is used to pass the feature maps of multiple scales through the corresponding detector classifiers respectively to obtain corresponding multiple preset boxes, and perform fast maximum suppression processing on the multiple preset boxes to obtain the target detection result.

[0077] In order to implement the above embodiments, the present disclosure also proposes a computer storage medium.

[0078] The computer storage medium provided in the embodiment of the present application stores an executable program; after the executable program is executed by the processor, the following can be achieved: Figure 1 The method shown.

[0079] In order to implement the above embodiments, the present disclosure also provides a computer device.

[0080] The computer device provided in the embodiment of the present application includes a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the program, it can achieve the following Figure 1 The method shown.

[0081] In order to implement the above embodiments, the present disclosure also proposes a mobile terminal device.

[0082] The mobile terminal device provided in the embodiment of the present application includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the following can be achieved: Figure 1 The method shown in 1.

[0083] In the method, device and storage medium for on-orbit intelligent detection of targets for rotating swing-sweep ultra-wideband satellites proposed in this application, a lightweight target detection model is constructed, and the lightweight target detection model is tuned using a historical remote sensing image data set to obtain an optimal lightweight target detection model, and then the acquired remote sensing image to be detected is input into the optimal lightweight target detection model, and the backbone network output feature map is obtained through the lightweight feature extraction backbone network in the optimal lightweight target detection model, and then the backbone network output feature map is subjected to multiple different convolution operations to obtain feature maps of multiple scales, and the feature maps of multiple scales are respectively passed through the corresponding detector classifiers to obtain corresponding multiple preset frames, and the multiple preset frames are subjected to fast maximum suppression processing to obtain the target detection result. It can be seen from this that the method proposed in this application obtains the optimal lightweight target detection model through training of the historical remote sensing image data set, so that there is no need to rely on expert experience, and target detection can also be performed on the remote sensing image to be detected, reducing the number of parameters and calculations, and reducing the model size and computing power consumption. At the same time, the lightweight target detection model adopts a single-stage detection framework, and the lightweight computing unit of the lightweight feature extraction backbone network adopts group depthwise separable convolution, thereby meeting the limited on-board resources, improving the efficiency of satellite on-orbit target detection, and thus realizing ultra-wide-band satellite rapid detection of targets.

[0084] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0085] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code that includes one or more schedulable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be scheduled in the order shown or discussed, including scheduling functions in a substantially simultaneous manner or in reverse order based on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.

[0086] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. An on-orbit intelligent detection method for rotating and sweeping ultra-wideband satellites, characterized in that: The method comprises: Constructing a lightweight target detection model, wherein the lightweight target detection model includes a lightweight feature extraction backbone network; Use historical remote sensing image datasets to tune the lightweight target detection model and obtain the optimal lightweight target detection model; Inputting the acquired remote sensing image to be detected into the optimal lightweight target detection model, and obtaining a backbone network output feature map through the lightweight feature extraction backbone network in the optimal lightweight target detection model; The output feature map of the backbone network is subjected to multiple convolution operations to obtain feature maps of multiple scales; Passing feature maps of multiple scales through corresponding detector classifiers respectively to obtain corresponding multiple preset frames, and performing fast maximum suppression processing on the multiple preset frames to obtain target detection results; The lightweight feature extraction backbone network includes a multi-scale feature fusion module and an original feature enhancement module. The step of inputting the acquired remote sensing image to be detected into the optimal lightweight target detection model, and obtaining a backbone network output feature map through a lightweight feature extraction backbone network in the optimal lightweight target detection model, comprises: The acquired remote sensing image to be detected is inputted into the multi-scale feature fusion module and the original feature enhancement module in the optimal lightweight target detection model in parallel, and the feature map of the multi-scale feature fusion module and the feature map of the original feature enhancement module are obtained respectively; The feature map of the multi-scale feature fusion module and the feature map of the original feature enhancement module are fused to obtain the output feature map of the backbone network; The multi-scale feature fusion module and the original feature enhancement module respectively include N lightweight computing units connected in series, where N is a positive integer. The method of inputting the acquired remote sensing image to be detected into the multi-scale feature fusion module and the original feature enhancement module in the optimal lightweight target detection model in parallel, and obtaining the feature map of the multi-scale feature fusion module and the feature map of the original feature enhancement module respectively, comprises: The acquired remote sensing image to be detected is input into the multi-scale feature fusion module in the optimal lightweight target detection model. The input remote sensing image to be detected is the scale one feature. Then the remote sensing image to be detected is input into the first lightweight computing unit to obtain the scale two feature. The obtained scale two feature is input into the second lightweight computing unit to obtain the scale three feature. And so on. The obtained scale features are sequentially passed through the remaining N-2 lightweight computing units to obtain the corresponding N-2 scale features. Then the obtained N scale features are merged to obtain the feature map of the multi-scale feature fusion module. The acquired remote sensing image to be detected is input into the original feature enhancement module in the optimal lightweight target detection model, the input remote sensing image to be detected is input into the first lightweight computing unit to obtain the first feature, the remote sensing image to be detected and the first feature are input into the second lightweight computing unit to obtain the second feature, and so on, the remote sensing image to be detected and the obtained features are sequentially passed through the remaining N-2 series lightweight computing units to obtain the Nth feature, and the Nth feature is the feature map of the original feature enhancement module.

2. The method according to claim 1, characterized in that The lightweight computing unit is constructed by a group of depthwise separable convolutions, wherein the group of depthwise separable convolutions is to stack multiple layers of depthwise convolutions and then uniformly perform pointwise convolutions.

3. The method according to claim 1, characterized in that The use of historical remote sensing image datasets to tune the lightweight target detection model includes: Preprocess historical remote sensing image datasets; Divide the preprocessed remote sensing image dataset into a training set and a validation set; Inputting the training set into a lightweight object detection model, and calculating a loss function of the lightweight object detection model; The lightweight target detection model is tuned using the validation set according to the loss value of the loss function to obtain the optimal lightweight target detection model.

4. The method according to claim 1, characterized in that: The lightweight target detection model adopts a single-stage detection framework.

5. An on-orbit intelligent detection device for rotating and sweeping ultra-wideband satellites, characterized in that: The device comprises: A construction module, used to construct a lightweight target detection model, wherein the lightweight target detection model includes a lightweight feature extraction backbone network; The tuning module is used to tune the lightweight target detection model using historical remote sensing image datasets to obtain the optimal lightweight target detection model; A first processing module is used to input the acquired remote sensing image to be detected into an optimal lightweight target detection model, and obtain a backbone network output feature map through a lightweight feature extraction backbone network in the optimal lightweight target detection model; A second processing module is used to obtain feature maps of multiple scales by subjecting the output feature map of the backbone network to multiple convolution operations; An output module is used to pass the feature maps of multiple scales through corresponding detector classifiers respectively to obtain corresponding multiple preset boxes, and perform fast maximum suppression processing on the multiple preset boxes to obtain target detection results; The lightweight feature extraction backbone network includes a multi-scale feature fusion module and an original feature enhancement module, and the first processing module is further used for: The acquired remote sensing image to be detected is inputted into the multi-scale feature fusion module and the original feature enhancement module in the optimal lightweight target detection model in parallel, and the feature map of the multi-scale feature fusion module and the feature map of the original feature enhancement module are obtained respectively; The feature map of the multi-scale feature fusion module and the feature map of the original feature enhancement module are fused to obtain the output feature map of the backbone network; The multi-scale feature fusion module and the original feature enhancement module respectively include N lightweight computing units connected in series, where N is a positive integer, and the first processing module is further used for: The acquired remote sensing image to be detected is input into the multi-scale feature fusion module in the optimal lightweight target detection model. The input remote sensing image to be detected is the scale one feature. Then the remote sensing image to be detected is input into the first lightweight computing unit to obtain the scale two feature. The obtained scale two feature is input into the second lightweight computing unit to obtain the scale three feature. And so on. The obtained scale features are sequentially passed through the remaining N-2 lightweight computing units to obtain the corresponding N-2 scale features. Then the obtained N scale features are merged to obtain the feature map of the multi-scale feature fusion module. The acquired remote sensing image to be detected is input into the original feature enhancement module in the optimal lightweight target detection model, the input remote sensing image to be detected is input into the first lightweight computing unit to obtain the first feature, the remote sensing image to be detected and the first feature are input into the second lightweight computing unit to obtain the second feature, and so on, the remote sensing image to be detected and the obtained features are sequentially passed through the remaining N-2 series lightweight computing units to obtain the Nth feature, and the Nth feature is the feature map of the original feature enhancement module.

6. A computer storage medium, wherein: The computer storage medium stores computer executable instructions; after the computer executable instructions are executed by the processor, any method described in claims 1-4 can be implemented.

7. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method according to any one of claims 1 to 4 is implemented.

8. A mobile terminal device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method according to any one of claims 1 to 4 is implemented.

Images