Method, apparatus, and equipment for generating navigation images for unmanned vehicles based on random weights
By using a style transfer network based on random weights and optimizing the generator with navigation image discriminators and remote sensing image discriminators, the generalization and generation quality issues of autonomous vehicle navigation image generation in off-road environments are solved, achieving efficient navigation map generation and meeting the mission requirements of autonomous vehicles in multiple environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NAT UNIV OF DEFENSE TECH
- Filing Date
- 2023-11-24
- Publication Date
- 2026-05-26
Smart Images
Figure CN117611430B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of unmanned vehicle navigation image generation technology, and in particular to an unmanned vehicle navigation image generation method, apparatus and device based on random weights. Background Technology
[0002] Obtaining a navigation map of the current environment, similar to the map style displayed by Gaode Maps navigation software, is very difficult. Obtaining a navigation map through point cloud data or directly through a navigation system is feasible, but it is time-consuming, has low accuracy, and is highly complex, resulting in significant technical costs. However, style transfer from remote sensing images (satellite imagery) can significantly reduce technical costs while maintaining generation quality. Currently, most style transfer techniques for multi-environment hybrid mapping in autonomous driving mainly improve the loss function of the internal network to enhance model adaptability and training performance, but lack substantial progress and do not improve optimization methods. Furthermore, current style transfer techniques are widely applied to various tasks of autonomous vehicles in urban environments but lack application in off-road environments, resulting in weak model generalization and failing to meet the needs of autonomous vehicles performing tasks in unknown and diverse environments.
[0003] Traditional style transfer models lack robustness and generalization, resulting in two major drawbacks: firstly, low training efficiency, leading to unsatisfactory quality and clarity in generated navigation maps; and secondly, insufficient training hindering their ability to handle data from off-road environments. While some patents have proposed improving style transfer models using the SWA method, this approach has not been applied to map transfer and still faces issues such as poor map quality and low training efficiency. Summary of the Invention
[0004] Therefore, it is necessary to provide a method, apparatus, and device for generating navigation images for unmanned vehicles based on random weights, which can reconstruct navigation images in unknown environments, in order to address the aforementioned technical problems.
[0005] A method for generating navigation images for unmanned vehicles based on random weights, the method comprising:
[0006] Construct a style transfer network for navigation images. The style transfer network consists of a first generator and a second generator.
[0007] The remote sensing images collected by the unmanned vehicle are used as the first training sample dataset. The first training sample dataset is input into the first generator. The first generator processes the first training sample dataset using the dense random weight averaging method to obtain the feature map and upsampled features of the remote sensing images, and outputs the navigation image.
[0008] The first generator is trained based on the pre-built navigation image discriminator and the navigation image to obtain a trained first generator.
[0009] The navigation image is obtained by style fusion of the navigation image using the trained first generator.
[0010] The predicted navigation image is used as the second training sample dataset. The second training sample dataset is input into the second generator. The second generator performs restoration processing on the real navigation image to obtain the feature map and upsampled features of the navigation image, and outputs the predicted remote sensing image.
[0011] The second generator is trained based on the predicted remote sensing images to obtain a trained second generator.
[0012] Based on the predicted remote sensing image and the optimized navigation image discriminator based on the remote sensing image, a real navigation image discriminator is obtained.
[0013] A cyclic loss function is constructed by using a real navigation image discriminator. The trained second generator then uses the cyclic loss function to correct the predicted navigation image, thus obtaining the real navigation image.
[0014] In one embodiment, the first generator and the second generator are inverse neural networks to each other.
[0015] In one embodiment, the method further includes: the first generator uses a dense random weight averaging method to calculate the contribution of each generator in the style transfer network based on the weights of the first generator.
[0016]
[0017] Among them, w c P represents the weight of the sampling dimension of the remote sensing image. c Let C represent the sampling contribution of each generator in the style transfer network to the remote sensing image, and C be the number of sampling layers for each generator in the style transfer network. After regularizing the contribution, the average network parameters of each generator are calculated to obtain the average weight:
[0018]
[0019] in, P represents the contribution after regularization. c Let C represent the sampling contribution of each generator in the style transfer network to the remote sensing image, and let C be the number of sampling layers for each generator in the style transfer network. The first training sample data is processed according to the average weight to obtain the feature map and upsampled features of the remote sensing image, and the navigation image is output to the pre-constructed navigation image discriminator and the second generator.
[0020] In one embodiment, the method further includes: using a trained first generator to perform style fusion of the remote sensing image and a preset Gaussian model to obtain a predicted navigation image.
[0021] In one embodiment, the method further includes: constructing a cyclic loss function using a real navigation image discriminator, backpropagating the cyclic loss function to update the network parameters of each generator in the style transfer network, and calibrating it with the predicted navigation image to obtain the real navigation image.
[0022] An image generation device for autonomous vehicle navigation based on random weights, the device comprising:
[0023] The Style Transfer Network Builder module is used to construct a style transfer network for navigation images. The style transfer network includes a first generator and a second generator.
[0024] The remote sensing image conversion module is used to acquire remote sensing images collected by the unmanned vehicle as the first training sample dataset. The first training sample dataset is input into the first generator. The first generator processes the first training sample dataset using the dense random weight averaging method to obtain the feature map and upsampled features of the remote sensing image, and outputs the navigation image.
[0025] The first generator training module is used to train the first generator based on a pre-built navigation image discriminator and a navigation image to obtain the trained first generator.
[0026] The predicted navigation image acquisition module is used to perform style fusion of the navigation image through a trained first generator to obtain a predicted navigation image.
[0027] The navigation image conversion module is used to predict navigation images as the second training sample dataset. The second training sample dataset is input into the second generator, which performs restoration processing on the real navigation images to obtain the feature maps and upsampled features of the navigation images, and outputs the predicted remote sensing images.
[0028] The second generator training module is used to train the second generator based on the predicted remote sensing image to obtain the trained second generator.
[0029] The discriminator update module is used to optimize the navigation image discriminator based on the predicted remote sensing image and the remote sensing image to obtain the real navigation image discriminator.
[0030] The real navigation image generation module is used to construct a cyclic loss function through a real navigation image discriminator. The trained second generator then corrects the predicted navigation image based on the cyclic loss function to obtain the real navigation image.
[0031] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:
[0032] Construct a style transfer network for navigation images. The style transfer network consists of a first generator and a second generator.
[0033] The remote sensing images collected by the unmanned vehicle are used as the first training sample dataset. The first training sample dataset is input into the first generator. The first generator processes the first training sample dataset using the dense random weight averaging method to obtain the feature map and upsampled features of the remote sensing images, and outputs the navigation image.
[0034] The first generator is trained based on a pre-built navigation image discriminator and a navigation image to obtain a trained first generator.
[0035] The navigation image is obtained by style fusion of the navigation image using the trained first generator.
[0036] The predicted navigation image is used as the second training sample dataset. The second training sample dataset is input into the second generator. The second generator performs restoration processing on the real navigation image to obtain the feature map and upsampled features of the navigation image, and outputs the predicted remote sensing image.
[0037] The second generator is trained based on the predicted remote sensing images to obtain a trained second generator.
[0038] Based on the predicted remote sensing image and the optimized navigation image discriminator based on the remote sensing image, a real navigation image discriminator is obtained.
[0039] A cyclic loss function is constructed by using a real navigation image discriminator. The trained second generator then uses the cyclic loss function to correct the predicted navigation image, thus obtaining the real navigation image.
[0040] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:
[0041] Construct a style transfer network for navigation images. The style transfer network consists of a first generator and a second generator.
[0042] The remote sensing images collected by the unmanned vehicle are used as the first training sample dataset. The first training sample dataset is input into the first generator. The first generator processes the first training sample dataset using the dense random weight averaging method to obtain the feature map and upsampled features of the remote sensing images, and outputs the navigation image.
[0043] The first generator is trained based on a pre-built navigation image discriminator and a navigation image to obtain a trained first generator.
[0044] The navigation image is obtained by style fusion of the navigation image using the trained first generator.
[0045] The predicted navigation image is used as the second training sample dataset. The second training sample dataset is input into the second generator. The second generator performs restoration processing on the real navigation image to obtain the feature map and upsampled features of the navigation image, and outputs the predicted remote sensing image.
[0046] The second generator is trained based on the predicted remote sensing images to obtain a trained second generator.
[0047] Based on the predicted remote sensing image and the optimized navigation image discriminator based on the remote sensing image, a real navigation image discriminator is obtained.
[0048] A cyclic loss function is constructed by using a real navigation image discriminator. The trained second generator then uses the cyclic loss function to correct the predicted navigation image, thus obtaining the real navigation image.
[0049] The aforementioned method, apparatus, and device for generating navigation images for unmanned vehicles based on random weights, through a first generator, can convert remote sensing images into navigation images, achieving style conversion. By training the generator using a pre-built navigation image discriminator, the generator's ability to generate realistic navigation images is improved. The first and second generators can achieve style fusion of navigation images and restoration of remote sensing images, thus completing the conversion between the two types of images. Optimizing the navigation image discriminator improves its ability to distinguish between real and predicted navigation images. Applying a recurrent loss function helps the generator learn better, improving the accuracy and realism of predicted navigation images. This enables efficient mapping in urban environments, allowing unmanned vehicles to meet the needs of performing tasks in unknown and multi-environment environments, and also improves the accuracy of navigation image generation. Attached Figure Description
[0050] Figure 1 This is a flowchart illustrating an unmanned vehicle navigation image generation method based on random weights in one embodiment.
[0051] Figure 2 This is a flowchart illustrating the steps of a style transfer network in a style transfer task in one embodiment.
[0052] Figure 3 Here is a block diagram of the style transfer network in one embodiment;
[0053] Figure 4 This is a structural block diagram of an unmanned vehicle navigation image generation device based on random weights in one embodiment;
[0054] Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation
[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0056] In one embodiment, such as Figure 1 As shown, a method for generating navigation images for autonomous vehicles based on random weights is provided, including the following steps:
[0057] Step 102: Construct a style transfer network for navigation images.
[0058] The style transfer network consists of a first generator and a second generator.
[0059] Step 104: Obtain the remote sensing images collected by the unmanned vehicle as the first training sample dataset, input the first training sample dataset into the first generator, and process the first training sample dataset using the dense random weight averaging method to obtain the feature map and upsampled features of the remote sensing images, and output the navigation image.
[0060] Step 106: Train the first generator based on the pre-built navigation image discriminator and the navigation image to obtain the trained first generator.
[0061] Step 108: Style fusion of the navigation image is performed using the trained first generator to obtain the predicted navigation image.
[0062] Step 110: The predicted navigation image is used as the second training sample dataset. The second training sample dataset is input into the second generator. The second generator performs restoration processing on the real navigation image to obtain the feature map and upsampled features of the navigation image, and outputs the predicted remote sensing image.
[0063] Step 112: Train the second generator based on the predicted remote sensing image to obtain the trained second generator.
[0064] Step 114: Optimize the navigation image discriminator based on the predicted remote sensing image and the remote sensing image to obtain the real navigation image discriminator.
[0065] Step 116: Construct a cyclic loss function using a real navigation image discriminator. The trained second generator then uses the cyclic loss function to correct the predicted navigation image, thus obtaining the real navigation image.
[0066] In the aforementioned method for generating navigation images for autonomous vehicles based on random weights, a first generator converts remote sensing images into navigation images, achieving style transfer. The generator is trained using a pre-built navigation image discriminator, improving its ability to generate realistic navigation images. The first and second generators enable style fusion of navigation images and restoration of remote sensing images, thus completing the conversion between the two types of images. Optimizing the navigation image discriminator enhances its ability to distinguish between real and predicted navigation images. Applying a recurrent loss function helps the generator learn better, improving the accuracy and realism of predicted navigation images. This allows for efficient mapping in urban environments, enabling autonomous vehicles to perform tasks in unknown and multi-environment conditions, including off-road environments, and improving the accuracy of navigation image generation.
[0067] In one embodiment, the first generator and the second generator are inverse neural networks to each other.
[0068] In one embodiment, the first generator uses a dense random weighted averaging method to calculate the contribution of each generator in the style transfer network based on the weights of the first generator:
[0069]
[0070] Among them, w c P represents the weight of the sampling dimension of the remote sensing image. c Let C represent the sampling contribution of each generator in the style transfer network to the remote sensing image, and C be the number of sampling layers for each generator in the style transfer network. After regularizing the contribution, the average network parameters of each generator are calculated to obtain the average weight:
[0071]
[0072] in, P represents the contribution after regularization. c Let C represent the sampling contribution of each generator in the style transfer network to the remote sensing image, and let C be the number of sampling layers for each generator in the style transfer network. The first training sample data is processed according to the average weight to obtain the feature map and upsampled features of the remote sensing image, and the navigation image is output to the pre-constructed navigation image discriminator and the second generator.
[0073] It is worth noting that the contribution regularization calculation method is as follows:
[0074]
[0075] Among them, P c Let C represent the sampling contribution of each generator in the style transfer network to the remote sensing image, and let C be the number of sampling layers for each generator in the style transfer network.
[0076] In one embodiment, a pre-trained first generator performs style fusion of the remote sensing image with a preset Gaussian model to obtain a predicted navigation image.
[0077] In one embodiment, a cyclic loss function is constructed using a real navigation image discriminator. The cyclic loss function is then backpropagated to update the network parameters of each generator in the style transfer network. The network parameters are then corrected along with the predicted navigation image to obtain the real navigation image.
[0078] In one embodiment, such as Figure 2 As shown, a style transfer network is provided for the steps in style transfer tasks, applied to, for example... Figure 3 The style transfer network shown in the diagram includes the following steps:
[0079] S10: Input remote sensing images into G A2B network.
[0080] G A2B (Generator A to B) is a generator that converts a type A image into a type B image. Type A images are remote sensing style map images, while type B images are navigation style map images.
[0081] Specifically, by pre-training the network to extract features of two styles, and then only needing to input a string of Gaussian noise to the network, a clear image of the corresponding style can be generated based on the features. This patent focuses on style transfer tasks for maps.
[0082] S20: Trained G A2B The network converts remote sensing images into corresponding navigation maps.
[0083] Specifically, after the input data preprocessing is completed, it is necessary to use G... A2B Network-based style transfer, G A2B The network employs a dense weighted average (SWAD) method.
[0084] Furthermore, the specific steps of the SWAD method's average weight strategy are as follows:
[0085] 1. The model weights are stored as W = [w1, ..., w2]. c The form is ], where each weight w c The dimension is K;
[0086] 2. Using formulas To calculate the contribution of each model;
[0087] 3. Using formulas Calculate the parameters of the regularized model;
[0088] 4. Using formulas Calculate the average value of multiple models.
[0089] It is worth noting that traditional networks are simply generators, with low efficiency in extracting data features and training, and cannot meet the needs.
[0090] S30: Input the obtained navigation map into D B The network is a discriminator for B-class images.
[0091] Specifically, after the conversion, theoretically, a navigation map with the same content as the remote sensing image in the source domain but a significantly different style should be obtained—that is, a Class B image. However, this training process is not instantaneous. A loss function is established to guide the model in updating its parameters, making the generated navigation map closer to the true navigation style. Therefore, a discriminator D is needed. B By measuring the difference between the two, and then narrowing the difference through an optimizer, the generated navigation map can be made more realistic and accurate.
[0092] S40: Trained G B2A The network then reconstructs the navigation map obtained in the second step into a remote sensing image of the source domain.
[0093] G B2A (Generator B to A: A generator that converts a B-class image into a A-class image).
[0094] Specifically, similar to G in S20 A2B Network, G here B2A The network is its reverse operation, specifically G B2A Its function is to convert Class B navigation maps into Class A remote sensing style maps. It is also based on the weighted averaging principle in S20 for training and prediction.
[0095] S50: Establish a “cyclic loss” between the source domain remote sensing image and the reconstructed remote sensing image.
[0096] It is worth noting that a remote sensing style map from the source domain was processed by G A2B and G B2A After two style transfers in opposite directions, there should be no change. However, it's possible that the restored image only has a style similar to the original. Figure 1 However, the content has changed, so a loss function needs to be established between the two and the network parameters need to be updated through backpropagation. This is also known as ensuring "cyclic consistency". The loss function established in this process is called the "cyclic loss function".
[0097] S60: The obtained loss function is backpropagated to update the network parameters.
[0098] It is worth noting that optimizing and backpropagating all the loss functions established during network training to update network parameters and reduce loss results in better generation performance.
[0099] S70: Outputs a navigation map after style transfer.
[0100] Specifically, after the model is trained, it can take a remote sensing style map as input and output a navigation style map.
[0101] It should be understood that, although Figure 1-2 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1-2 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.
[0102] In one embodiment, such as Figure 4 As shown, an unmanned vehicle navigation image generation device based on random weights is provided, including: a style transfer network construction module 402, a remote sensing image conversion module 404, a first generator training module 406, a predicted navigation image acquisition module 408, a navigation image conversion module 410, a second generator training module 412, a discriminator update module 414, and a real navigation image generation module 416, wherein:
[0103] Style transfer network building module 402 is used to build a style transfer network for navigation images. The style transfer network includes a first generator and a second generator.
[0104] The remote sensing image conversion module 404 is used to acquire remote sensing images collected by the unmanned vehicle as the first training sample dataset, input the first training sample dataset into the first generator, and process the first training sample dataset using the dense random weight averaging method to obtain the feature map and upsampling features of the remote sensing image, and output the navigation image.
[0105] The first generator training module 406 is used to train the first generator based on a pre-built navigation image discriminator and a navigation image to obtain the trained first generator.
[0106] The predicted navigation image acquisition module 408 is used to perform style fusion of the navigation image through a trained first generator to obtain a predicted navigation image.
[0107] The navigation image conversion module 410 is used to predict navigation images as a second training sample dataset. The second training sample dataset is input into the second generator, and the second generator performs restoration processing on the real navigation images to obtain the feature map and upsampled features of the navigation images, and outputs the predicted remote sensing image.
[0108] The second generator training module 412 is used to train the second generator based on the predicted remote sensing image to obtain the trained second generator.
[0109] The discriminator update module 414 is used to optimize the navigation image discriminator based on the predicted remote sensing image and the remote sensing image to obtain the real navigation image discriminator.
[0110] The real navigation image generation module 416 is used to construct a cyclic loss function through a real navigation image discriminator. The trained second generator corrects the predicted navigation image according to the cyclic loss function to obtain the real navigation image.
[0111] Specific limitations regarding the random weight-based autonomous vehicle navigation image generation device can be found in the limitations of the random weight-based autonomous vehicle navigation image generation method described above, and will not be repeated here. Each module in the aforementioned random weight-based autonomous vehicle navigation image generation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.
[0112] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 5As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for generating navigation images for unmanned vehicles based on random weights. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device's casing, or an external keyboard, touchpad, or mouse.
[0113] Those skilled in the art will understand that Figure 3-5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0114] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to perform the following steps:
[0115] Construct a style transfer network for navigation images. The style transfer network consists of a first generator and a second generator.
[0116] The remote sensing images collected by the unmanned vehicle are used as the first training sample dataset. The first training sample dataset is input into the first generator. The first generator processes the first training sample dataset using the dense random weight averaging method to obtain the feature map and upsampled features of the remote sensing images, and outputs the navigation image.
[0117] The first generator is trained based on a pre-built navigation image discriminator and a navigation image to obtain a trained first generator.
[0118] The navigation image is obtained by style fusion of the navigation image using the trained first generator.
[0119] The predicted navigation image is used as the second training sample dataset. The second training sample dataset is input into the second generator. The second generator performs restoration processing on the real navigation image to obtain the feature map and upsampled features of the navigation image, and outputs the predicted remote sensing image.
[0120] The second generator is trained based on the predicted remote sensing images to obtain a trained second generator.
[0121] Based on the predicted remote sensing image and the optimized navigation image discriminator based on the remote sensing image, a real navigation image discriminator is obtained.
[0122] A cyclic loss function is constructed by using a real navigation image discriminator. The trained second generator then uses the cyclic loss function to correct the predicted navigation image, thus obtaining the real navigation image.
[0123] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:
[0124] Construct a style transfer network for navigation images. The style transfer network consists of a first generator and a second generator.
[0125] The remote sensing images collected by the unmanned vehicle are used as the first training sample dataset. The first training sample dataset is input into the first generator. The first generator processes the first training sample dataset using the dense random weight averaging method to obtain the feature map and upsampled features of the remote sensing images, and outputs the navigation image.
[0126] The first generator is trained based on a pre-built navigation image discriminator and a navigation image to obtain a trained first generator.
[0127] The navigation image is obtained by style fusion of the navigation image using the trained first generator.
[0128] The predicted navigation image is used as the second training sample dataset. The second training sample dataset is input into the second generator. The second generator performs restoration processing on the real navigation image to obtain the feature map and upsampled features of the navigation image, and outputs the predicted remote sensing image.
[0129] The second generator is trained based on the predicted remote sensing images to obtain a trained second generator.
[0130] Based on the predicted remote sensing image and the optimized navigation image discriminator based on the remote sensing image, a real navigation image discriminator is obtained.
[0131] A cyclic loss function is constructed by using a real navigation image discriminator. The trained second generator then uses the cyclic loss function to correct the predicted navigation image, thus obtaining the real navigation image.
[0132] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0133] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0134] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.
Claims
1. A random weight based unmanned vehicle navigation image generation method, characterized in that, The method includes: A style transfer network for navigation images is constructed; the style transfer network includes a first generator and a second generator; The remote sensing images collected by the unmanned vehicle are used as the first training sample dataset. The first training sample dataset is input into the first generator. The first generator processes the first training sample dataset using the dense random weight averaging method to obtain the feature map and upsampled features of the remote sensing images, and outputs the navigation image. The first generator is trained based on the pre-built navigation image discriminator and the navigation image to obtain the trained first generator. The navigation image is style-fused using the trained first generator to obtain a predicted navigation image; The predicted navigation image is used as the second training sample dataset. The second training sample dataset is input into the second generator. The second generator performs restoration processing on the real navigation image to obtain the feature map and upsampled features of the navigation image, and outputs the predicted remote sensing image. The second generator is trained based on the predicted remote sensing image to obtain the trained second generator; The navigation image discriminator is optimized based on the predicted remote sensing image and the remote sensing image to obtain the real navigation image discriminator; A cyclic loss function is constructed using the real navigation image discriminator. The trained second generator then uses the cyclic loss function to correct the predicted navigation image, resulting in a corrected real navigation image.
2. The method according to claim 1, characterized in that, The first generator and the second generator are inverse neural networks to each other.
3. The method according to claim 1, characterized in that, The first generator processes the first training sample dataset using a dense random weighted averaging method to obtain the feature map and upsampled features of the remote sensing image, and outputs a navigation image. The first generator processes the first training sample data using a dense random weighted averaging method to obtain the feature map and upsampled features of the remote sensing image, and outputs the navigation image to the pre-constructed navigation image discriminator and the second generator.
4. The method according to claim 3, characterized in that, The trained first generator is used to perform style fusion on the navigation image to obtain a predicted navigation image, including: The trained first generator performs style fusion of the remote sensing image and the preset Gaussian model to obtain a predicted navigation image.
5. The method according to claim 4, characterized in that, A cyclic loss function is constructed using the real navigation image discriminator. The trained second generator then uses the cyclic loss function to correct the predicted navigation image, resulting in a corrected real navigation image, including: A cyclic loss function is constructed using the real navigation image discriminator. The cyclic loss function is then backpropagated to update the network parameters of each generator in the style transfer network. The results are then corrected using the predicted navigation image to obtain the corrected real navigation image.
6. An unmanned vehicle navigation image generation device based on random weights, characterized in that, The device includes: A style transfer network construction module is used to construct a style transfer network for navigation images; the style transfer network includes a first generator and a second generator; The remote sensing image conversion module is used to acquire remote sensing images collected by the unmanned vehicle as the first training sample dataset, input the first training sample dataset into the first generator, and the first generator processes the first training sample dataset using the dense random weight averaging method to obtain the feature map and upsampled features of the remote sensing image, and outputs the navigation image. The first generator training module is used to train the first generator based on a pre-built navigation image discriminator and the navigation image to obtain the trained first generator. The predicted navigation image acquisition module is used to perform style fusion of the navigation image through the trained first generator to obtain a predicted navigation image; The navigation image conversion module is used to take the predicted navigation image as the second training sample dataset, input the second training sample dataset into the second generator, perform restoration processing on the real navigation image through the second generator, obtain the feature map and upsampled features of the navigation image, and output the predicted remote sensing image. The second generator training module is used to train the second generator based on the predicted remote sensing image to obtain the trained second generator. The discriminator update module is used to optimize the navigation image discriminator based on the predicted remote sensing image and the remote sensing image to obtain the real navigation image discriminator; The real navigation image generation module is used to construct a cyclic loss function through the real navigation image discriminator, and the trained second generator corrects the predicted navigation image according to the cyclic loss function to obtain the corrected real navigation image.
7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.