Launch vehicle status identification method and system combined with shallow cloud removal from remote sensing images

Abstract

The present invention discloses a launch vehicle state identification method and system combined with shallow cloud removal from remote sensing images, comprising: obtaining original optical remote sensing image data, performing preprocessing, and inputting the data into an end-to-end integrated network; performing shallow cloud removal on the preprocessed image using a DehazeNet cloud removal network in the network; performing target feature extraction, feature integration, launch vehicle positioning detection, and state classification and identification on the image after shallow cloud removal using a cascaded SC-Yolo target detection network; performing supervised training by calculating the respective loss functions of the cloud removal network and the target detection network, feeding back the launch vehicle positioning detection result to the cloud removal network, and adjusting the cloud removal weight; inputting the optical remote sensing image data to be tested into the trained end-to-end integrated network, and outputting the launch vehicle positioning detection result and the state classification result. The present invention improves the launch vehicle state identification accuracy under shallow cloud interference and reduces the false alarm rate.

Description

Technical Field

[0001] The present invention relates to the technical field of computer image processing, and more particularly to a method and system for identifying the state of a launch vehicle in combination with shallow cloud and fog removal from remote sensing images. Background Art

[0002] Launch vehicle status identification is crucial in military applications. Quickly and accurately obtaining the location and status information of enemy launch vehicles plays a key role in strategic decision-making and tactical implementation. Remote sensing technology, due to its wide detection range and high efficiency and real-time characteristics, has become an important means of obtaining launch vehicle status information. Using remote sensing technology to obtain launch vehicle status can provide necessary information for strategic decision-making, effectively shorten the decision-making cycle, and has great research value.

[0003] At present, popular deep learning methods have been able to achieve good recognition of the status of remote sensing image launch vehicles. However, they also face increasingly severe challenges, including huge changes in the target's visual appearance due to viewpoint changes, occlusion, cloud and fog interference, etc. Especially in foggy conditions, the visual features of the launch vehicle are easily obscured, making it difficult for traditional algorithms to accurately identify the status of the launch vehicle. These complex background interference factors make it easy for targets to be missed or misdetected, resulting in difficulty in further improving the accuracy of launch vehicle status recognition.

[0004] Therefore, how to improve the accuracy of transmitting vehicle status identification of remote sensing images under shallow cloud and fog interference is a problem that technicians in this field urgently need to solve. Summary of the Invention

[0005] In view of this, the present invention provides a launch vehicle status identification method and system combined with shallow cloud and fog removal of remote sensing images to solve the technical problems mentioned in the background technology.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for identifying the state of a launch vehicle combined with shallow cloud and fog removal from remote sensing images comprises the following steps:

[0008] S101, obtaining original optical remote sensing image data;

[0009] S102, preprocessing the original optical remote sensing image data and inputting it into an end-to-end integrated network in which a cloud removal network and a target detection network are sequentially connected;

[0010] S103, removing shallow clouds and fog from the pre-processed optical remote sensing image data through the DehazeNet cloud removal network in the end-to-end integrated network;

[0011] S104, using the cascaded SC-Yolo target detection network in the end-to-end integrated network to perform target feature extraction, feature integration, launch vehicle positioning detection, and launch vehicle status classification and identification on the image data after the shallow cloud and fog are removed;

[0012] S105, performing supervised training by calculating the loss functions of the DehazeNet cloud removal network and the cascaded SC-Yolo target detection network, and feeding back the launch vehicle positioning detection results of the cascaded SC-Yolo target detection network to the DehazeNet cloud removal network to adjust the cloud removal weight during cloud removal;

[0013] S106: Obtain the optical remote sensing image data to be tested and input it into the trained end-to-end integrated network, and output the launch vehicle positioning detection results and the launch vehicle status classification and identification results.

[0014] Preferably, in step S102, the specific content of preprocessing the original optical remote sensing image data is: performing linear quantization and image block cropping on the obtained original remote sensing image, and marking the target true value frame and the target launch vehicle status.

[0015] Preferably, the DehazeNet cloud removal network includes a feature extraction block, a feature enhancement block, and an image reconstruction block;

[0016] The feature extraction block, feature enhancement block and image reconstruction block are connected in sequence, and the feature extraction block and image reconstruction block are skipped at the same feature map scale;

[0017] The feature extraction block takes the original image as input, which is processed through two convolutional layers and a downsampling layer to obtain a deep feature map of the image. The feature enhancement block uses the Swin-transformer encoder and decoder to enhance the deep feature map and restore the texture features of the transmitting vehicle. The image reconstruction block reconstructs the image after the feature enhancement through two convolutional layers and an upsampling layer, combining the multi-scale features obtained by the jump connection, to obtain a clear image without shallow clouds and fog.

[0018] Preferably, the cascaded SC-Yolo object detection network includes a backbone network, a neck network, an object detection head, and a state classification head connected in sequence;

[0019] The backbone network is used to gradually extract feature maps at different levels from the image data after removing shallow clouds and fog. The neck network is used to perform multi-scale feature fusion. The target detection head is used to detect and locate the launch vehicle through multi-scale fusion feature maps. The state classification head is used to identify the specific state of the launch vehicle based on the output results of the target detection head.

[0020] Preferably, the target detection head takes the multi-scale fusion feature map as input and the appearance features of the launch vehicle as the true value. It is first processed by two 3×3 convolution layers, and then processed by the bounding box regression layer and the 1×1 convolution layer respectively, and then outputs the bounding box position of the launch vehicle in the original image.

[0021] Preferably, the state classification head performs further state identification based on the output of the detection head, taking the transmitting vehicle slice as input and the different state information of the transmitting vehicle as the true value. After passing through a 2×2 convolution layer, a flattening layer and two fully connected layers, the state identification result of the transmitting vehicle is output.

[0022] Preferably, in step S105, the feedback adjustment loss function for adjusting the cloud removal weight during cloud removal is specifically:

[0023]

[0024] Where i is the i-th pixel of the image, X is the image to be declouded, Y is the feedback image with the target detection frame, that is, the positioning detection result of the transmitting vehicle, and λ is the cloud removal weight coefficient of the area around the target detection frame.

[0025] A launch vehicle state identification system combined with shallow cloud removal from remote sensing images is based on the aforementioned launch vehicle state identification method combined with shallow cloud removal from remote sensing images, comprising: a remote sensing image acquisition module, a preprocessing module, an end-to-end integrated network, a supervised training module, and a cloud removal weight feedback adjustment module. The end-to-end integrated network comprises a sequentially connected DehazeNet cloud removal network and a cascaded SC-Yolo target detection network.

[0026] A launch vehicle status identification device combined with remote sensing image shallow cloud and fog removal includes an image acquisition device, a memory and a processor;

[0027] Image acquisition equipment, used for acquiring optical remote sensing images to be detected;

[0028] The memory stores a computer program executable on the processor;

[0029] The processor is connected to the image acquisition device and the memory data. When the processor executes the computer program, it implements the method for identifying the state of the launch vehicle combined with the removal of shallow clouds and fog from remote sensing images, and automatically identifies and classifies the state of the launch vehicle in the optical remote sensing image to be detected collected by the image acquisition device.

[0030] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements a method for identifying the state of a launch vehicle combined with shallow cloud and fog removal from remote sensing images.

[0031] It can be seen from the above technical solution that compared with the existing technology, the present invention discloses a method and system for identifying the status of a launch vehicle combined with the removal of shallow clouds and fog from remote sensing images, which optimizes the problem of difficulty in identifying the status of the launch vehicle caused by the weakening of the appearance characteristics of the launch vehicle under cloud and fog coverage conditions, improves the accuracy of the launch vehicle status identification under the interference of shallow clouds and fog, and at the same time reduces the false alarm rate to a certain extent, thereby obtaining a high-precision launch vehicle status identification result. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0033] Figure 1 A flow chart of the launch vehicle status identification method combined with shallow cloud and fog removal from remote sensing images provided by the present invention;

[0034] Figure 2 A schematic diagram of an end-to-end integrated network connection provided by the present invention;

[0035] Figure 3 Schematic diagram of the DehazeNet network model provided by the present invention;

[0036] Figure 4 Schematic diagram of the SC-Yolo network model provided by the present invention;

[0037] Figure 5 Schematic diagram of the comparison between the shallow cloud removal and target detection provided by the present invention and the original image;

[0038] Figure 6 Schematic diagram of the target detection head and state classification head provided by the present invention;

[0039] Figure 7 This is a schematic diagram of the launch vehicle status identification result provided by the present invention;

[0040] Figure 8 This is a schematic diagram of the launch vehicle status identification device provided by the present invention. DETAILED DESCRIPTION

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0042] The embodiment of the present invention discloses a method for identifying the state of a launch vehicle by combining remote sensing image shallow cloud removal. Figure 1 and Figure 2 As shown, the following steps are included:

[0043] S101, obtaining original optical remote sensing image data;

[0044] In this embodiment, the original optical remote sensing image data is an original 16-bit optical remote sensing image taken by a satellite;

[0045] S102, preprocessing the original optical remote sensing image data and inputting it into an end-to-end integrated network in which a cloud removal network and a target detection network are sequentially connected;

[0046] S103, removing shallow clouds and fog from the pre-processed optical remote sensing image data through the DehazeNet cloud removal network in the end-to-end integrated network;

[0047] S104, using the cascaded SC-Yolo target detection network in the end-to-end integrated network to perform target feature extraction, feature integration, launch vehicle positioning detection, and launch vehicle status classification and identification on the image data after the shallow cloud and fog are removed;

[0048] S105, performing supervised training by calculating the loss functions of the DehazeNet cloud removal network and the cascaded SC-Yolo target detection network, and feeding back the launch vehicle positioning detection results of the cascaded SC-Yolo target detection network to the DehazeNet cloud removal network to adjust the cloud removal weight during cloud removal;

[0049] S106: Obtain the optical remote sensing image data to be tested and input it into the trained end-to-end integrated network, and output the launch vehicle positioning detection results and the launch vehicle status classification and identification results.

[0050] In order to further implement the above technical solution, in step S102, the specific content of preprocessing the original optical remote sensing image data is: linear quantization and image block cropping of the obtained original remote sensing image, marking the target true value frame and the target launch vehicle status;

[0051] Specifically: First, the original 16-bit remote sensing image is linearly quantized at 2% to obtain 8-bit visual image data; second, the obtained image data is cropped into 512×512 image blocks for making model training samples; then, the target ground truth box and the target launch vehicle status are annotated based on the processed clear image data.

[0052] In order to further implement the above technical solutions, Figure 3As shown in the figure, the DehazeNet network is improved based on U-net and its internal architecture is optimized to achieve high-fidelity shallow cloud removal of small target features in remote sensing images. Specifically, the DehazeNet cloud removal network includes feature extraction block, feature enhancement block and image reconstruction block.

[0053] The feature extraction block, feature enhancement block and image reconstruction block are connected in sequence, and the feature extraction block and image reconstruction block are skipped at the same feature map scale;

[0054] The feature extraction block takes the original image as input and passes it through two convolutional layers and a 2×2 downsampling layer to obtain a deep feature map of the image. The feature enhancement block uses a Swin-transformer encoder and decoder to enhance the deep feature map and restore the texture characteristics of the transmitting vehicle. The image reconstruction block reconstructs the image after the feature enhancement through two convolutional layers and a 2×2 upsampling layer, combining the multi-scale features obtained by skip connections, to obtain a clear image without shallow fog.

[0055] In this example, since the launch vehicle occupies fewer pixels in the remote sensing image, the features are weaker after two downsamplings in the feature extraction block, and it is difficult to retain the texture features of the launch vehicle using convolution. Therefore, the feature enhancement module uses a swin-transformer encoder instead of the convolution layer to better restore the texture features of the launch vehicle after feature enhancement; at the same time, a jump connection is made between the feature extraction block and the image reconstruction block at the same feature map scale, so that the image reconstruction block can more effectively obtain multi-scale features.

[0056] In order to further implement the above technical solutions, Figure 4 ,The cascaded SC-Yolo object detection network includes a backbone network, a neck network, an object detection head, and a state classification head, which are connected in sequence;

[0057] The backbone network is used to gradually extract feature maps at different levels from the image data after removing shallow clouds and fog. The neck network is used to perform multi-scale feature fusion. The target detection head is used to detect and locate the launch vehicle through multi-scale fusion feature maps. The state classification head is used to identify the specific state of the launch vehicle based on the output results of the target detection head.

[0058] In this embodiment, the input image size of the backbone network is a 512×512 image. Through a series of convolutional layers, pooling layers, and residual blocks, feature maps of different levels are gradually extracted. In this process, three feature maps are extracted from different levels of the backbone network, namely P2, P3, and P4. Among them, P2 is a high-resolution feature map extracted from the Conv2_x layer, which mainly captures the detail information of the image. P3 is a medium-to-high-resolution feature map extracted from the Conv3_x layer, which combines detail and semantic information. P4 is a low-resolution feature map extracted from the Conv5_x layer, which mainly captures deep semantic information. Next, the feature maps of different levels are input into the neck network (feature pyramid network) for multi-scale feature fusion. The multi-scale fused feature maps will be used for subsequent detection and classification tasks.

[0059] In order to further implement the above technical solutions, Figure 5 As shown in the figure, the target detection head takes the multi-scale fusion feature map as input and the appearance features of the launch vehicle as the true value. It is first processed by two 3×3 convolutional layers, and then processed by the bounding box regression layer and the 1×1 convolutional layer respectively, and outputs the bounding box position of the launch vehicle in the original image.

[0060] In order to further implement the above technical solution, the SC-Yolo network was improved on the basis of the original Yolo network. A state classification head for state identification was added on the basis of the detection head to better realize the state classification between targets with similar features. Specifically: the state classification head performs further state identification based on the output of the detection head, with the launch vehicle slice as input and the different state information of the launch vehicle as the true value. After a 2×2 convolution layer, a flattening layer and two fully connected layers, the state identification result of the launch vehicle is output. The launch vehicle state includes parking state, dynamic state and combat state.

[0061] like Figure 6 and Figure 7 As shown, Figure 6 (a) is the image block after the original 16-bit image is quantized and cropped. Figure 6 (b) is the image after shallow cloud removal and target detection; Figure 7 (a) in the figure indicates that the launch vehicle is in combat state. Figure 7 (b) in the figure shows that the launch vehicle is identified as parked.

[0062] In this embodiment, the DehazeNet cloud removal network and the cascaded SC-Yolo target detection network calculate their respective loss functions and perform gradient backpropagation. Specifically, the DehazeNet cloud removal network uses the declouded image as the true value to calculate the loss, and the SC-Yolo target detection network uses the launch vehicle state result as the true value to calculate the loss. Among them, the loss function of the target detection head is Yolo's regression loss, and the loss function of the state classification head is Yolo's classification loss. At the same time, the loss function of DehazeNet will also adjust the weights based on the image with target position information output after the SC-Yolo network completes target detection.

[0063] In order to further implement the above technical solution, in step S105, the feedback adjustment loss function for adjusting the cloud removal weight during cloud removal is specifically:

[0064]

[0065] Wherein, i is the i-th pixel point of the image, X is the image to be declouded, Y is the feedback image with the target detection frame, that is, the positioning detection result of the transmitting vehicle, and λ is the cloud removal weight coefficient of the area around the target detection frame. In this embodiment, with the assistance of the feedback image with the detection result, the cloud removal network can be provided with the target position information, so that the weight coefficient is used to enhance the cloud removal strength around the target, effectively enhance the target features, and thus improve the performance of state recognition.

[0066] A launch vehicle state identification system combined with shallow cloud removal from remote sensing images is based on a launch vehicle state identification method combined with shallow cloud removal from remote sensing images. The system includes: a remote sensing image acquisition module, a preprocessing module, an end-to-end integrated network, a supervised training module, and a cloud removal weight feedback adjustment module. The end-to-end integrated network includes a sequentially connected DehazeNet cloud removal network and a cascaded SC-Yolo target detection network.

[0067] A launch vehicle status identification device 303 combined with remote sensing image shallow cloud removal, such as Figure 8 As shown, it includes an image acquisition device 301, a memory 302 and a processor 303;

[0068] Image acquisition device 301, used to collect optical remote sensing images to be detected;

[0069] The memory 302 stores a computer program that can be run on the processor 303;

[0070] The processor 303 is data-connected to the image acquisition device 301 and the memory 302. When the processor 303 executes the computer program, it implements a method for identifying the status of a launch vehicle combined with the removal of shallow clouds and fog from remote sensing images, and automatically identifies and classifies the status of the launch vehicle in the optical remote sensing image to be detected collected by the image acquisition device 301.

[0071] In actual applications, the memory 302, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer executable programs and modules; the memory 302 may include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required by at least one function, and the data storage area can store data created by the processor 302, etc.; in addition, the memory 302 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device; in some aspects, the memory 302 may optionally include a memory remotely located relative to the processor 303, and these remote memories may be connected to the processor 303 via a network, and the network includes but is not limited to the Internet, an intranet, a local area network, a mobile communication network and a combination thereof.

[0072] The processor 303 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or a combination of the above chips.

[0073] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements a method for identifying the state of a launch vehicle combined with shallow cloud and fog removal from remote sensing images.

[0074] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0075] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying the state of a launch vehicle by combining remote sensing imagery with shallow cloud removal, characterized in that: The following steps are involved: S101, obtaining original optical remote sensing image data; S102, preprocessing the original optical remote sensing image data and inputting it into an end-to-end integrated network in which a cloud removal network and a target detection network are sequentially connected; S103, removing shallow clouds and fog from the pre-processed optical remote sensing image data through the DehazeNet cloud removal network in the end-to-end integrated network; S104, using the cascaded SC-Yolo target detection network in the end-to-end integrated network to perform target feature extraction, feature integration, launch vehicle positioning detection, and launch vehicle status classification and identification on the image data after the shallow cloud and fog are removed; S105, performing supervised training by calculating the loss functions of the DehazeNet cloud removal network and the cascaded SC-Yolo target detection network, and feeding back the launch vehicle positioning detection results of the cascaded SC-Yolo target detection network to the DehazeNet cloud removal network to adjust the cloud removal weight during cloud removal; S106: Obtain the optical remote sensing image data to be tested and input it into the trained end-to-end integrated network, and output the launch vehicle positioning detection result and the launch vehicle status classification and identification result; The DehazeNet cloud removal network includes feature extraction block, feature enhancement block and image reconstruction block; The feature extraction block, feature enhancement block and image reconstruction block are connected in sequence, and the feature extraction block and image reconstruction block are skipped at the same feature map scale; The feature extraction block takes the original image as input and passes it through two convolutional layers and a downsampling layer to obtain a deep feature map of the image. The feature enhancement block uses a Swin-transformer encoder and decoder to enhance the deep feature map and restore the texture characteristics of the transmitting vehicle. The image reconstruction block reconstructs the image after the feature enhancement through two convolutional layers and an upsampling layer, combining the multi-scale features obtained by skip connections, to obtain a clear image without shallow fog. The cascaded SC-Yolo object detection network consists of a backbone network, a neck network, an object detection head, and a state classification head connected in sequence; The backbone network is used to gradually extract feature maps at different levels from the image data after removing shallow clouds and fog. The neck network is used to perform multi-scale feature fusion. The target detection head is used to detect and locate the launch vehicle through multi-scale fusion feature maps. The state classification head is used to identify the specific state of the launch vehicle based on the output results of the target detection head.

2. The method for identifying the state of a launch vehicle combined with shallow cloud removal from remote sensing images according to claim 1 is characterized in that: In step S102, the specific content of preprocessing the original optical remote sensing image data is: linear quantization and image block cropping of the obtained original remote sensing image, and marking the target true value frame and the target launch vehicle status.

3. The method for identifying the state of a launch vehicle combined with shallow cloud removal from remote sensing images according to claim 1, characterized in that: The target detection head takes the multi-scale fusion feature map as input and the appearance features of the launch vehicle as the true value. It is first processed by two 3×3 convolutional layers, and then processed by a bounding box regression layer and a 1×1 convolutional layer respectively, and outputs the bounding box position of the launch vehicle in the original image.

4. The method for identifying the state of a launch vehicle combined with shallow cloud removal from remote sensing images according to claim 1, characterized in that: The state classification head performs further state identification based on the output of the detection head. It takes the launch vehicle slice as input and the different state information of the launch vehicle as the true value. After a 2×2 convolutional layer, a flattening layer and two fully connected layers, it outputs the state identification result of the launch vehicle.

5. The method for identifying the state of a launch vehicle in combination with shallow cloud removal from remote sensing images according to claim 1, characterized in that: In step S105, the feedback adjustment loss function for adjusting the cloud removal weight during cloud removal is specifically: ； Where i is the i-th pixel of the image, X is the image to be declouded, Y is the feedback image with the target detection frame, that is, the positioning detection result of the transmitting vehicle, and λ is the cloud removal weight coefficient of the area around the target detection frame.

6. A launch vehicle status identification system combined with remote sensing image shallow cloud removal, characterized in that: A launch vehicle status identification method combined with shallow cloud removal from remote sensing images based on any one of claims 1-5 includes: a remote sensing image acquisition module, a preprocessing module, an end-to-end integrated network, a supervised training module, and a cloud removal weight feedback adjustment module. The end-to-end integrated network includes a sequentially connected DehazeNet cloud removal network and a cascaded SC-Yolo target detection network.

7. A launch vehicle status identification device combined with remote sensing image shallow cloud removal, characterized in that: including an image acquisition device, a memory, and a processor; Image acquisition equipment, used for acquiring optical remote sensing images to be detected; The memory stores a computer program executable on the processor; The processor is connected to the image acquisition device and the memory data. When the processor executes the computer program, it implements the launch vehicle status identification method combined with the remote sensing image shallow cloud removal as described in any one of claims 1 to 5, and automatically identifies and classifies the launch vehicle status in the optical remote sensing image to be detected collected by the image acquisition device.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for identifying the state of a launch vehicle combined with shallow cloud and fog removal from remote sensing images as described in any one of claims 1 to 5 is implemented.

Images