Three-dimensional geographic information acquisition system and acquisition method thereof

Abstract

The invention relates to the technical field of geographic information, and discloses a three-dimensional geographic information collection system and method, and the system comprises a data collection module which is used for collecting three-dimensional geographic image data; the data processing module is used for processing the three-dimensional geographic image data and repairing and reconstructing damaged image data, the data processing module comprises a generative adversarial network (GAN), the generative adversarial network comprises a generator and a discriminator, and the generator carries out image reconstruction based on an image damaged area; the data storage module is used for storing the repaired image data; and the output module is used for outputting the reconstructed image data. High-precision restoration of image data is achieved through the generative adversarial network, the precision and detail integrity of three-dimensional geographic information collection are improved, the system has the automatic processing capacity and is suitable for various geographic scenes, and the collection efficiency and the data quality are improved.

Description

Technical Field

[0001] The present invention relates to the field of geographic information technology, in particular to a three-dimensional geographic information acquisition system and an acquisition method thereof. Background Art

[0002] With the rapid development of 3D geographic information technology, the collection and processing of geographic information has been widely used in urban planning, environmental monitoring, disaster emergency response and other fields. Traditional 3D geographic information collection systems usually rely on drones, satellites or airborne sensors to obtain image data, and then construct 3D models of terrain or buildings through image stitching, point cloud data fusion and other methods. However, during the collection process, image data is often partially damaged or degraded due to external factors such as weather, lighting conditions, and obstructions, which seriously affects the accuracy and completeness of details of the 3D model.

[0003] In order to solve the problem of image data corruption, existing image restoration technologies mostly use interpolation, filtering or rule-based image reconstruction methods. Interpolation fills in missing areas with known pixel values, while filtering is used to remove noise or blur effects in images. In addition, some technologies also reduce occlusion problems by stitching multiple images from multiple angles, and subsequently use image enhancement algorithms to improve the overall image quality.

[0004] The existing image restoration technology lacks the ability to extract deep features, resulting in visually unnatural restoration results. In particular, in the case of large-area occlusion, strong noise or blurred images, traditional interpolation and filtering methods are difficult to restore details and structures in the image. In addition, the existing technology has limited applicability when dealing with a variety of complex geographical environments and cannot adapt to the needs of image restoration with different resolutions and types of damage. The restoration process often relies on manual adjustments and is inefficient. Summary of the invention

[0005] In view of the deficiencies of the prior art, the present invention provides a three-dimensional geographic information acquisition system and an acquisition method thereof, which solves the problems of insufficient image restoration accuracy, limited applicability, and inability to automatically and efficiently process complex geographic scenes in the prior art.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a three-dimensional geographic information acquisition system and an acquisition method thereof, comprising:

[0007] Data acquisition module, used to collect three-dimensional geographic image data;

[0008] A data processing module, used for processing the three-dimensional geographic image data, including repairing and reconstructing damaged image data, wherein the data processing module includes a generative adversarial network, wherein the generative adversarial network includes a generator and a discriminator, and the generator reconstructs the image based on the damaged area of ​​the image;

[0009] A data storage module, used for storing restored image data;

[0010] The output module is used to output the reconstructed image data.

[0011] The generative adversarial network trains the generator through adversarial loss and L1 loss, and the adversarial loss formula is:

[0012]

[0013] Among them, G(x,z) is the repaired image generated by the generator, D(x,y) is the discrimination result output by the discriminator, x is the damaged image, and y is the real image.

[0014] Preferably, the generator adopts a U-Net architecture, including multiple encoding layers and decoding layers. The encoding layer extracts image features through a downsampling convolution layer, and the decoding layer gradually restores the image through a deconvolution layer, and directly passes the corresponding encoding layer features to the decoding layer through a jump connection.

[0015] Preferably, the discriminator adopts the PatchGAN architecture to discriminate local areas of the image, performs true and false discrimination on each local area of ​​the image through several convolutional layers, and outputs a two-dimensional feature map.

[0016] Preferably, the data processing module further comprises:

[0017] A data preprocessing unit, used for normalizing, denoising and simulating damaged areas of the original image data;

[0018] The data enhancement unit is used to rotate, crop and add noise to the collected image data to expand the data set and improve the generalization ability of the model.

[0019] The acquisition method of the three-dimensional geographic information acquisition system includes:

[0020] Step S1: Collect three-dimensional geographic image data to generate an initial data set;

[0021] Step S2: preprocessing the image data in the initial data set to generate image data with damaged areas;

[0022] Step S3: inputting the damaged image data into a generator of a generative adversarial network, and generating repaired image data through the generator;

[0023] Step S4: using a discriminator to distinguish true from false on the restored image data, and training the generator through adversarial loss and L1 loss until the quality of the generated image reaches a preset threshold;

[0024] Step S5: Outputting the restored image data.

[0025] Preferably, in step S2, the preprocessing step includes image normalization processing, which scales the pixel values ​​of the image data to the interval [0, 1] or [-1, 1] to meet the input requirements of the generative adversarial network.

[0026] Preferably, in step S4, the combined loss formula of the adversarial loss and the L1 loss is:

[0027] L G (G) = L GAN (G,D)+λL L1 (G)

[0028] Among them, λ is the weight coefficient used to balance the adversarial loss and L1 loss. The L1 loss is defined as:

[0029]

[0030] Preferably, the generator in step S3 adopts a U-Net architecture, including a downsampling convolution layer and an upsampling deconvolution layer, and directly passes the features of the encoder to the corresponding layer of the decoder through a skip connection to retain the high-resolution features of the image.

[0031] Preferably, the discriminator in step S4 adopts a PatchGAN structure, and the discriminator judges an image area of ​​70x70 pixels by a convolution operation, and outputs a two-dimensional true and false feature map.

[0032] Preferably, the method further includes increasing the diversity of the data set through data augmentation operations after step S2, wherein the data augmentation includes image rotation, scaling, random cropping and random noise addition to improve the generalization ability of the generative adversarial network.

[0033] The present invention provides a three-dimensional geographic information collection system and a collection method thereof. It has the following beneficial effects:

[0034] 1. The present invention combines a generative adversarial network with a U-Net architecture to perform high-precision restoration of three-dimensional geographic images with occlusion, blur or noise pollution. The collaborative work of the generator and the discriminator ensures that the generated restored image is close to the real image in both global and local details, thereby greatly improving the quality and resolution of the image. Compared with traditional interpolation and filtering methods, the restoration result of the present invention is more natural in visual effect and is more suitable for subsequent three-dimensional modeling.

[0035] 2. The three-dimensional geographic information acquisition system of the present invention can be applied to a variety of complex geographical environments, such as cities, forests, mountains, etc., and can process image data of different resolutions and various forms of damage. By generating the adaptive characteristics of adversarial networks, the present invention can flexibly respond to the image restoration needs of various scenes, greatly improving the applicability and generalization ability of the system in practical applications, so that it can be applied to geographic information acquisition and processing in multiple fields.

[0036] 3. The system architecture of the present invention is highly automated. The entire process of image acquisition, preprocessing, restoration and output can be completed without human intervention. The automatic optimization mechanism of the generative adversarial network makes the image restoration process more efficient and can process a large amount of geographic information data in a shorter time, thereby improving the overall efficiency of three-dimensional geographic information acquisition. This automation feature is particularly suitable for large-scale data acquisition tasks such as drones and satellites, and can significantly reduce labor costs and improve productivity. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a system architecture diagram of the present invention;

[0038] Figure 2 It is a structural schematic diagram of the data acquisition module of the present invention;

[0039] Figure 3 It is a structural schematic diagram of a data preprocessing module of the present invention;

[0040] Figure 4 It is a structural schematic diagram of a data processing module of the present invention;

[0041] Figure 5 It is a structural schematic diagram of the data storage module of the present invention;

[0042] Figure 6 It is a structural schematic diagram of the output module of the present invention. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the specification of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0044] Example:

[0045] Please see attached Figure 1 - Attachment Figure 6The present invention relates to a three-dimensional geographic information acquisition system and acquisition method based on deep learning, which are mainly used to deal with the image damage problem caused by external conditions (such as clouds, shadows, insufficient light, etc.) during the geographic information acquisition process. The damaged or low-quality images are repaired and reconstructed by using a generative adversarial network (GAN) and a convolutional neural network (CNN) to improve the acquisition quality and accuracy of three-dimensional geographic information.

[0046] System Architecture

[0047] The three-dimensional geographic information acquisition system of the present invention is composed of the following modules:

[0048] Data acquisition module: responsible for acquiring original three-dimensional geographic image data from drones, satellites, airborne sensors and other equipment. Data acquisition covers different terrains (such as urban buildings, mountains, forests, etc.), and ensures the integrity of the collected data through multi-angle and multi-level methods.

[0049] Data preprocessing module: used to perform preliminary processing on the collected image data, including data normalization, denoising, and simulation of damaged parts in the image (such as cloud occlusion, blur, and noise areas) to generate damaged image data to be input into the subsequent repair model. Through normalization, the image pixel values ​​are mapped to the [0,1] or [-1,1] interval to ensure data consistency.

[0050] Data processing module:

[0051] Generative Adversarial Network (GAN): The core of the data processing module is to use the generative adversarial network to repair and reconstruct the damaged image. The generative adversarial network consists of a generator and a discriminator. The generator is responsible for generating the repaired image based on the damaged image, while the discriminator is used to distinguish the repaired image from the real image.

[0052] Generator (U-Net architecture): The generator adopts the U-Net structure, extracts multi-scale features of the image through multi-layer convolution, and reconstructs the image layer by layer using deconvolution. At the same time, it passes the features in the encoder directly to the decoder through jump connections to retain the high-resolution details of the image.

[0053] Discriminator (PatchGAN architecture): The discriminator is based on the PatchGAN architecture and can judge whether a local area of ​​the image (such as a 70x70 pixel block) is true or false, which enables the model to detect local features in the generated image more finely.

[0054] Loss function design: The generator is trained by a combination of adversarial loss and L1 loss. Adversarial loss is used to guide the generator to generate realistic images, and L1 loss is used to ensure the similarity between the generated image and the target image at the pixel level. The loss formula is:

[0055]

[0056] Among them, x represents the input damaged image, y represents the real image, G(x) represents the repaired image output by the generator, and D(x,y) represents the output of the discriminator.

[0057] Data storage module: used to store restored image data and original image data. The restored image can be directly used for subsequent 3D modeling or further processing.

[0058] Output module: Output the restored image data to the 3D Geographic Information System (GIS) for subsequent 3D modeling and analysis. The output module also supports the visualization of the restoration results.

[0059] Method steps

[0060] Step S1: Collecting image data

[0061] The data acquisition module obtains 3D geographic image data from drones, satellites or other sensors. The collected image data can be different scenes such as urban environments, forests, mountains, etc. The data needs to have different resolutions to meet the needs of various application scenarios.

[0062] Step S2: Preprocessing of image data

[0063] Before data processing, the system will preprocess the raw image data. The steps include:

[0064] Data normalization: Scale the pixel values ​​of the image to the range of [0,1] or [-1,1] to unify the format of the input data and ensure that the model can process the image correctly.

[0065] Damage simulation: Generate images with damaged areas by randomly adding occlusion (such as clouds), noise, or blur. This step provides more samples with damage information for model training.

[0066] Step S3: The generator generates the repaired image

[0067] By inputting damaged image data into the generator of the generative adversarial network, the generator extracts image features through the U-Net architecture and repairs and completes the damaged areas. The generator can effectively reconstruct the lost information in the image and maintain the high-resolution details of the image.

[0068] Step S4: The discriminator evaluates the authenticity of the image

[0069] The discriminator uses the PatchGAN architecture to evaluate the authenticity of the generated repaired image and the real image. It makes judgments based on local areas of the image (such as 70x70 pixel blocks), ensuring that the repaired image is not only close to the real image in general, but also maintains a high degree of similarity in local details.

[0070] Step S5: Optimization and output

[0071] Through the combination of adversarial loss and L1 loss, the system continuously optimizes the generator until the quality of the restored image reaches the preset standard. Subsequently, the restored image data is stored and output to the 3D geographic information system for subsequent processing or modeling.

[0072] Principle Analysis

[0073] The basic principle of Generative Adversarial Network (GAN) is to use the game relationship between the generator and the discriminator to perform image restoration. The goal of the generator is to generate a restoration image that is as realistic as possible, while the goal of the discriminator is to distinguish which image is generated and which image is original. Through repeated iterative training, the generator gradually learns how to generate restoration images that are difficult to distinguish from real images.

[0074] The generator of the U-Net architecture uses an encoding-decoding method, which can effectively extract multi-scale features of the image and use jump connections to retain high-resolution features, thereby retaining more details during image reconstruction. The discriminator of the PatchGAN architecture judges the local area of ​​the image to ensure that the generated image is not only realistic in the global sense, but also consistent with the real image in local details.

[0075] Beneficial Effects

[0076] High quality of image restoration

[0077] By combining generative adversarial networks and convolutional neural networks, this system can effectively repair various damaged image data, including areas blocked by clouds, blurred images, or areas with noise interference. The generated repaired images are almost visually indistinguishable from the real images. The generator optimizes the adversarial loss and L1 loss, and the generated images are both realistic and have good structural consistency.

[0078] Applicable to various scenarios

[0079] The present invention is applicable to various geographical environments, such as urban buildings, forests, mountains and other complex scenes. Through the processing capabilities of images with different resolutions, the system can generate high-precision image data suitable for various 3D modeling needs.

[0080] System efficiency

[0081] The system of the present invention reduces manual intervention in image processing through automated data processing flow and optimized deep learning model. The automated image restoration of the generative adversarial network can significantly shorten the restoration time, making the processing of large-scale image data more efficient, and is suitable for application in practical geographic information collection.

[0082] Enhance the accuracy of 3D modeling

[0083] The restored image data is used for 3D geographic modeling, which can significantly improve the accuracy of the model. By restoring the obscured or lost image information, the system provides complete high-resolution data for 3D modeling, ensuring the authenticity and details of the model.

[0084] High degree of automation

[0085] The system automatically handles the entire process from data collection, preprocessing to image restoration, greatly reducing human involvement and is suitable for large-scale deployment on automated collection equipment such as drones and satellites.

[0086] Example 1: Urban geographic information image restoration based on generative adversarial network

[0087] Specific steps:

[0088] Data collection: Use drones to take aerial photos in urban environments to obtain high-resolution image data. The image resolution is 2560×1440 pixels.

[0089] Preprocessing: Normalize the image data and the pixel value range to the interval [0,1]. Randomly select some image areas to add occlusion to simulate cloud cover, and the occlusion area accounts for 20% of the total image area.

[0090] Model configuration:

[0091] Generator: It adopts U-Net structure. The encoder contains 4 convolutional layers, the decoder contains 4 deconvolutional layers, and jump connections are performed between corresponding layers.

[0092] Discriminator: PatchGAN architecture is used, with input of 70×70 image blocks and 5 layers of convolution operations.

[0093] Loss function: Combined use of adversarial loss L GAN (G,D) and L1 loss L L1 (G), weight coefficient λ=100.

[0094] Model training: Use Adam optimizer with learning rate of 0.0002, β1=0.5, β2=0.999. Train for 50 epochs.

[0095] Image restoration: Input the damaged image into the trained generator to generate the restored image.

[0096] Result output: The restored image is evaluated by PSNR and SSIM indicators.

[0097] Process parameters:

[0098] Resolution: 2560×1440

[0099] Blocked area: 20%

[0100] λ: 100

[0101] Learning rate: 0.0002

[0102] Epoch:50

[0103] Comparative Example 1: Restoration using traditional interpolation methods

[0104] Specific steps:

[0105] Data collection: Same as Example 1, collecting urban environment images.

[0106] Preprocessing: Same occlusion processing.

[0107] Interpolation repair: Use bilinear interpolation to fill the pixel values ​​of the occluded area.

[0108] Image output: The restored image is evaluated using the PSNR and SSIM metrics.

[0109] Process parameters:

[0110] Interpolation method: Bilinear interpolation

[0111] Example 2: Forest image restoration based on generative adversarial network

[0112] Specific steps:

[0113] Data collection: UAVs were used to collect images in a forest environment with an image resolution of 1920 × 1080 pixels.

[0114] Preprocessing: Normalize the image data and normalize the pixel value range to the interval [-1,1]. Randomly add 5% salt and pepper noise to the image to simulate noise pollution in bad weather.

[0115] Model configuration:

[0116] Generator: It adopts U-Net structure, which contains 5 convolution and deconvolution layers.

[0117] Discriminator: PatchGAN is used, the input is a 128×128 image patch, and the number of convolutional layers is 6.

[0118] Loss function: A combination of adversarial loss and perceptual loss is used, and perceptual loss extracts features through the VGG-19 network.

[0119] Model training: Use Adam optimizer, learning rate is 0.0001, β1=0.5, and train for 30 epochs.

[0120] Image Inpainting: Input the noisy image into the generator for inpainting.

[0121] Result output: evaluated by PSNR, SSIM and perceptual loss indicators.

[0122] Process parameters:

[0123] Resolution: 1920×1080

[0124] Noise: 5% salt and pepper noise

[0125] λ: 100

[0126] Learning rate: 0.0001

[0127] Epoch:30

[0128] Comparative Example 2: Noise removal using only Gaussian filtering

[0129] Specific steps:

[0130] Data collection: Same as Example 2, forest images are collected.

[0131] Preprocessing: Same noise addition.

[0132] Gaussian filter repair: Use a 3×3 Gaussian filter to smooth the noise.

[0133] Image output: The restored images are evaluated by PSNR, SSIM and perceptual loss.

[0134] Process parameters:

[0135] Filter: 3×3 Gaussian filter

[0136] Example 3: Mountain image restoration based on generative adversarial network

[0137] Specific steps:

[0138] Data collection: Aerial photography was performed using a drone in mountainous areas, with an image resolution of 2048 × 2048 pixels.

[0139] Preprocessing: Normalize the image data to the interval [0,1]. Add Gaussian blur to the image with blur level σ=3 to simulate image blur caused by flight jitter.

[0140] Model configuration:

[0141] Generator: It adopts U-Net structure, with 4 convolutional and deconvolutional layers respectively, and skip connection between encoder and decoder.

[0142] Discriminator: PatchGAN architecture, input is 70×70 image patch, number of convolutional layers is 5.

[0143] Loss function: Use adversarial loss L GAN (G,D) and L1 loss L L1 (G), λ=200.

[0144] Model training: Use Adam optimizer, learning rate is 0.0003, β1=0.5, and train for 40 epochs.

[0145] Image restoration: The input is a blurred image, and the generator generates a clear image.

[0146] Result output: Evaluated by PSNR and SSIM.

[0147] Process parameters:

[0148] Resolution: 2048×2048

[0149] Fuzziness: σ = 3

[0150] λ: 200

[0151] Learning rate: 0.0003

[0152] Epoch:40

[0153] Comparative Example 3: Deblurring using only the deconvolution method

[0154] Specific steps:

[0155] Data collection: Same as Example 3, mountain images are collected.

[0156] Preprocessing: Same Gaussian blur.

[0157] Deconvolution Inpainting: Deblurring is performed using a deconvolutional neural network (without using an adversarial network).

[0158] Image output: The restored image is evaluated by PSNR and SSIM.

[0159] Process parameters:

[0160] Using Deconvolution Layer

[0161] Comparative experiment

[0162] In order to verify the superiority of the three-dimensional geographic information acquisition system and its acquisition method of the present invention, a comparative experiment was designed. The image quality was evaluated by comparing the generative adversarial network (GAN) restoration method of the present invention with the traditional image restoration method. The experimental data included indicators such as PSNR, SSIM and perceptual loss, and the restoration effects in three environments, namely urban, forest and mountainous, were evaluated respectively.

[0163] Experimental design

[0164] Experimental object: Compare the image restoration effects of generative adversarial network (GAN) and traditional methods (interpolation, Gaussian filtering, deconvolution).

[0165] Evaluation indicators: Peak signal-to-noise ratio (PSNR), structural similarity (SSIM) and perceptual loss.

[0166] Test scenarios: city, forest, mountain.

[0167] Experimental steps:

[0168] The generative adversarial network of the present invention is used for image restoration, and the PSNR, SSIM, and perceptual loss are recorded.

[0169] The image restoration effects were compared using traditional methods, and the same evaluation indicators were recorded.

[0170] Experimental data table:

[0171]

[0172]

[0173] According to the table, the generative adversarial network (GAN) method of the present invention significantly improves the PSNR value in all scenes, especially in urban and mountain scenes, where the PSNR values ​​are increased by 6.67dB and 4.21dB respectively. This shows that the image quality after GAN restoration is closer to the real image than the traditional method; the structural similarity (SSIM) also shows a great advantage. In the images restored by GAN, the SSIM values ​​are all over 0.9, while the SSIM of the traditional method is around 0.8. This shows that after GAN restoration, the image is not only more similar globally, but also retains more complete local details; the GAN restoration method is also significantly better than the traditional method in terms of perceptual loss. Experimental data show that the perceptual loss of GAN is controlled in the range of 0.36-0.42, while the traditional method is between 2.54-3.12, indicating that the GAN method has a better effect in perceptual quality.

[0174] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A three-dimensional geographic information acquisition system, characterized in that: include: Data acquisition module, used to collect three-dimensional geographic image data; A data processing module, used for processing the three-dimensional geographic image data, including repairing and reconstructing damaged image data, wherein the data processing module includes a generative adversarial network, wherein the generative adversarial network includes a generator and a discriminator, and the generator reconstructs the image based on the damaged area of ​​the image; A data storage module, used for storing restored image data; The output module is used to output the reconstructed image data. The generative adversarial network trains the generator through adversarial loss and L1 loss, and the adversarial loss formula is: Among them, G(x,z) is the repaired image generated by the generator, D(x,y) is the discrimination result output by the discriminator, x is the damaged image, and y is the real image.

2. The three-dimensional geographic information acquisition system according to claim 1, characterized in that: The generator adopts a U-Net architecture, including multiple encoding layers and decoding layers. The encoding layer extracts image features through a downsampling convolution layer, and the decoding layer gradually restores the image through a deconvolution layer, and directly passes the corresponding encoding layer features to the decoding layer through a jump connection.

3. The three-dimensional geographic information acquisition system according to claim 1, characterized in that: The discriminator adopts the PatchGAN architecture to discriminate local areas of the image, distinguish true from false for each local area of ​​the image through several convolutional layers, and outputs a two-dimensional feature map.

4. The three-dimensional geographic information acquisition system according to claim 1, characterized in that: The data processing module further comprises: A data preprocessing unit, used for normalizing, denoising and simulating damaged areas of the original image data; The data enhancement unit is used to rotate, crop and add noise to the collected image data to expand the data set and improve the generalization ability of the model.

5. A collection method of a three-dimensional geographic information collection system, according to any one of claims 1 to 4, characterized in that: include: Step S1: Collect three-dimensional geographic image data to generate an initial data set; Step S2: preprocessing the image data in the initial data set to generate image data with damaged areas; Step S3: inputting the damaged image data into a generator of a generative adversarial network, and generating repaired image data through the generator; Step S4: using a discriminator to distinguish true from false on the restored image data, and training the generator through adversarial loss and L1 loss until the quality of the generated image reaches a preset threshold; Step S5: Outputting the restored image data.

6. The acquisition method of the three-dimensional geographic information acquisition system according to claim 5, characterized in that: In step S2, the preprocessing step includes image normalization processing, which scales the pixel values ​​of the image data to the interval [0, 1] or [-1, 1] to meet the input requirements of the generative adversarial network.

7. The acquisition method of the three-dimensional geographic information acquisition system according to claim 5, characterized in that: In step S4, the combined loss formula of adversarial loss and L1 loss is: L G (G)=L GAN (G,D)+λ L1 (G) Among them, λ is the weight coefficient used to balance the adversarial loss and L1 loss. The L1 loss is defined as:

8. The acquisition method of the three-dimensional geographic information acquisition system according to claim 5, characterized in that: The generator in step S3 adopts a U-Net architecture, including a downsampling convolution layer and an upsampling deconvolution layer, and directly passes the features of the encoder to the corresponding layer of the decoder through a skip connection to retain the high-resolution features of the image.

9. The acquisition method of the three-dimensional geographic information acquisition system according to claim 5, characterized in that: The discriminator in step S4 adopts the PatchGAN structure. The discriminator judges the image area of ​​70x70 pixels through a convolution operation and outputs a two-dimensional true and false feature map.

10. The acquisition method of the three-dimensional geographic information acquisition system according to claim 5, characterized in that: The method further includes increasing the diversity of the data set through data augmentation operations after step S2, wherein the data augmentation includes image rotation, scaling, random cropping and random noise addition to improve the generalization ability of the generative adversarial network.

Images