Unsupervised dense matching network training method, network training system and DSM generation method

Abstract

The invention relates to the technical field of computer vision and deep learning, in particular to an unsupervised dense matching network training method, a network training system and a DSM generation method, and the training method comprises the steps: obtaining a left image pair and a right image pair of a to-be-processed region; the left image pair and the right image pair are input into an unsupervised dense matching network with pixel-level structural similarity loss, namely PSSIM loss, as a loss function, the network is trained, and the unsupervised dense matching network is constructed based on a PSMNetb network structure; wherein the calculation of the PSSIM loss function comprises the following steps of: performing up-sampling on two images to be compared, calculating the similarity by using a structural similarity function, namely an SSIM function, and averaging calculation results. According to the method, the pixel-level structure similarity loss is introduced, so that the problem of pixel-level matching interference caused by local window calculation of the classical SSIM is effectively solved, and the precision and reliability of disparity map generation are remarkably improved under the condition that manual annotation is not needed.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and deep learning technology, and in particular to an unsupervised dense matching network training method, network training system, and DSM generation method. Background Technology

[0002] Digital Surface Models (DSMs), as fundamental data for modern geospatial analysis, play a crucial role in accurately describing the three-dimensional morphology of the Earth's surface and its objects. They comprehensively reflect the elevation information of natural terrain and man-made features, providing vital spatial data support for various fields such as urban planning, environmental monitoring, agricultural management, disaster assessment, wireless network planning, new energy development, and aviation navigation. DSM generation relies on distance information acquired from remote sensing platforms, combined with sensor position and attitude parameters. Currently, the main methods for acquiring distance information include airborne lidar and image-based dense matching. Compared to lidar, image-based dense matching methods offer advantages such as lower cost and easier data acquisition, thus enjoying wider application.

[0003] The core of dense matching is to construct a disparity map by finding corresponding points pixel by pixel in overlapping images, and then converting the disparity into distance information using a photogrammetric geometric model to finally generate a Disparity Model (DSM). To reduce computational complexity, epipolar constraints are usually introduced to simplify the two-dimensional matching problem into a one-dimensional search. The specific process includes epipolar geometry calculation, searching for corresponding points along the epipolar line, generating a disparity map, and converting the disparity map into a target three-dimensional product using sensor interior and exterior orientation elements. The accuracy of dense matching directly determines the quality of the final DSM and is a key step in the entire process. However, existing dense matching methods still have several shortcomings in practical applications, mainly in the following three aspects:

[0004] (1) The level of automation is limited, making it difficult to meet the surveying and mapping needs that emphasize both high precision and high efficiency. The existing workflow still relies heavily on manual intervention, especially in scenarios with high precision requirements, where a large amount of manual data collection and editing is still required, resulting in low efficiency and high cost;

[0005] (2) Traditional methods often use manually designed feature descriptors, which rely on professional knowledge and have weak generalization ability. Due to the significant differences in texture features between different image datasets, fixed descriptors are difficult to adapt to diverse scenes, resulting in insufficient matching robustness;

[0006] (3) In challenging areas such as repetitive textures, weak textures, occlusion, and lighting variations, the matching accuracy of traditional methods drops significantly, often requiring complex post-processing, which may introduce errors and affect the reliability of the results.

[0007] In recent years, with the practical application of deep learning technology, data-driven dense matching methods have shown significant advantages. Deep networks can automatically learn feature representations suitable for the task, reducing reliance on manually designed features and improving both automation and the robustness and accuracy of the algorithm. Currently, deep learning-based dense matching networks can be mainly divided into two categories: one is the general encoder-decoder structure represented by DispNet. This type of network structure borrows from the U-Net design, has a large parameter scale, simple structure, and fast inference speed, and is suitable for various pixel-level prediction tasks, but its specialization in dense matching is relatively weak; the other is the specialized network represented by GCNet, PSMNet, and GwcNet. Its structure borrows from the traditional stereo matching approach, introduces cost construction and three-dimensional convolution aggregation mechanisms, and outputs disparity maps through probabilistic disparity search and soft regression. Its accuracy is significantly higher than that of general networks, but its network structure is complex and it consumes a lot of computational resources.

[0008] Despite significant progress in dense matching using deep learning, existing methods still largely rely on supervised training with a large number of pixel-level ground truth labels. However, high-quality disparity annotation is costly, limiting its widespread application in practical engineering. Therefore, unsupervised or weakly supervised methods that do not rely on manual annotation have become an important research direction. Unsupervised dense matching, by designing reasonable loss functions (such as photometric consistency and smoothness constraints) to guide network learning, is more in line with practical application needs and represents an inevitable trend for future development.

[0009] Unsupervised loss functions mainly include reconstruction loss function, smoothing loss function, left-right consistency loss function, and maximum depth search loss function. Many unsupervised methods in academic papers use combinations and weights of these loss functions as constraints. Among these losses, the reconstruction loss function is the core constraint, while the others are auxiliary constraints. If the reconstruction loss function is removed, network training cannot converge. The reconstruction loss function consists of two parts: L1 and structural similarity method (SSIM). Preliminary experiments show that using only SSIM achieves higher accuracy. However, the classic (SSIM) requires consideration of local information, with a size of at least 3 pixels × 3 pixels, which does not meet the pixel-level requirement of dense matching. Excessive local information can cause surrounding information to interfere with matching pixels, thus reducing matching accuracy. To address this problem, this invention designs pixel-level structural similarity (PSSIM). Summary of the Invention

[0010] To address the shortcomings of existing dense matching methods, such as reliance on manual intervention and annotation, weak generalization ability, low accuracy in challenging regions with repetitive textures, and the inability of traditional SSIM loss to meet pixel-level matching requirements due to local information interference, this invention proposes an unsupervised dense matching network training method, network training system, and DSM generation method. These methods significantly improve the accuracy and reliability of disparity map generation without requiring manual annotation.

[0011] To achieve the above objectives, the technical solution adopted is:

[0012] This invention provides an unsupervised dense matching network training method, comprising the following steps:

[0013] Obtain the left and right image pairs of the region to be processed;

[0014] The left and right image pairs are input into an unsupervised dense matching network that uses pixel-level structural similarity loss, i.e., PSSIM loss, as the loss function, and the network is trained. The unsupervised dense matching network is constructed based on the PSMNetb network structure.

[0015] The calculation of the PSSIM loss function includes: upsampling the two images to be compared, calculating the similarity using the structural similarity function, i.e., the SSIM function, and averaging the results.

[0016] According to the unsupervised dense matching network training method of the present invention, the training process of the unsupervised dense matching network further includes:

[0017] Pair the left and right images with I L ,I R Input an unsupervised dense matching network to obtain the left disparity map D. L ;

[0018] Using a spatial transformation network, based on Figure I on the right R Compared with left disparity diagram D L The reconstructed image is obtained from the left.

[0019] Calculate the original left figure I L Reconstructed image from the left PSSIM loss between;

[0020] The horizontally flipped left and right images are input into an unsupervised dense matching network to obtain the right disparity map D. R ;

[0021] Using a spatial transformation network, based on the horizontally flipped left and right disparity maps D R Reconstruction yields the right reconstructed image. And perform a horizontal flip;

[0022] Calculate the original right figure I R Reconstructed right image after inversion PSSIM loss between;

[0023] Based on the two PSSIM loss values ​​mentioned above, the network parameters are backpropagated and updated.

[0024] According to the unsupervised dense matching network training method of the present invention, the calculation formula for the PSSIM loss is further as follows:

[0025]

[0026] Among them, F avg F represents the average function. SSIM F represents the structural similarity function. up Let I1 and I2 represent the upsampling function, and I1 and I2 represent the two images to be compared, here being the original left image I. L Reconstructed image from the left Or the original right figure I R Reconstructed image with right C piexl_ssim This represents the pixel-level structural similarity loss value.

[0027] According to the unsupervised dense matching network training method of the present invention, the structural similarity function uses a 3×3 convolution kernel to calculate local statistics, replacing the Gaussian convolution kernel in the traditional SSIM.

[0028] According to the unsupervised dense matching network training method of the present invention, the structural similarity function F is further... SSIM Defined as:

[0029]

[0030] Where I1 and I2 are the two input images, c1 and c2 are constants, and F m It is the mean function, F c It is the covariance function, F s It is a variance function.

[0031] According to the unsupervised dense matching network training method of the present invention, the PSMNetb network further comprises a feature pyramid module, a cost body construction module, and a three-dimensional convolutional aggregation module in sequence. The feature pyramid module is used to extract multi-scale feature maps from the input left and right images. The cost body construction module is used to construct a multi-scale three-dimensional cost body based on the multi-scale feature maps by matching the correspondence between the left and right feature maps under multiple preset disparities. The three-dimensional convolutional aggregation module is used to perform regularization processing on the three-dimensional cost body through stacked three-dimensional convolutional layers, and output the final disparity map through probabilistic processing, disparity regression, and upsampling.

[0032] Furthermore, the present invention also provides a method for generating a DSM, comprising the following steps:

[0033] Using the above-described unsupervised dense matching network training method, a trained unsupervised dense matching network model is obtained.

[0034] The left and right images of the region to be tested are input into the trained network model to generate a disparity map.

[0035] The parallax map is converted into distance information using photogrammetric geometry and the interior and exterior orientation elements of the sensor;

[0036] By utilizing distance information and combining sensor position and orientation, a digital surface model (DSM) is generated.

[0037] Furthermore, the present invention also provides an unsupervised dense matching network training system for implementing the above-described unsupervised dense matching network training method, the system comprising:

[0038] The image acquisition module is used to acquire left and right image pairs of the area to be processed;

[0039] The network training module is used to input the left and right image pairs into an unsupervised dense matching network that uses pixel-level structural similarity loss as the loss function, and train the network. The unsupervised dense matching network is constructed based on the PSMNetb network structure. The calculation of the PSSIM loss function includes: upsampling the two images to be compared, calculating the similarity using the structural similarity function, and averaging the calculation results.

[0040] The beneficial effects achieved by adopting the above technical solution are:

[0041] 1. Eliminate the dependence on dataset labels and expand application boundaries: This invention adopts an unsupervised training framework and uses a self-designed pixel-level structural similarity (PSSIM) loss as the core training constraint. This loss function evaluates the pixel-level consistency between the input image and the reconstructed image without relying on manually labeled high-quality disparity ground truth labels, thus completely breaking through the limitation of traditional supervised dense matching on dataset labels.

[0042] 2. Accurately improve disparity map precision and optimize matching effect: In view of the problem that traditional structural similarity (SSIM) loss needs to consider local information of ≥3×3 pixels and is easily interfered by surrounding pixels, the PSSIM loss of this invention focuses the similarity evaluation on the pixel level range, reducing the interference of irrelevant information on matching pixels; at the same time, combined with the image reconstruction verification mechanism of spatial transformation network (STN), it effectively improves the matching stability of challenging regions such as repetitive textures, weak textures, and occlusions, thereby significantly improving the precision of the disparity map output by the dense matching network.

[0043] 3. Strong network compatibility and support for performance expansion: The core improvements of this invention focus on the design of the loss function and unsupervised training logic. It does not rigidly bind the main structure of the dense matching network, and can be directly compatible with mainstream dedicated dense matching networks such as PSMNet and GwcNet. In the future, the overall dense matching accuracy can be further improved by upgrading the feature extraction capability or cost aggregation module performance of the adapted network, and it has flexible performance expansion space. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. The drawings are merely illustrative of some embodiments of the present invention and are not intended to limit the scope of the present invention to all embodiments.

[0045] Figure 1 This is a flowchart illustrating the unsupervised dense matching network training method according to an embodiment of the present invention;

[0046] Figure 2 This is a diagram of the PSMNetb network structure according to an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of pixel-level structural similarity loss according to an embodiment of the present invention;

[0048] Figure 4 This is a flowchart illustrating the DSM generation method according to an embodiment of the present invention. Detailed Implementation

[0049] The exemplary solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art.

[0050] This invention discloses an unsupervised dense matching network training method, the process of which is as follows: Figure 1 As shown, it includes the following steps:

[0051] Step S101: Obtain the left and right image pairs of the area to be processed.

[0052] For each set of left and right images, scale normalization, distortion correction, and grayscale normalization are performed to obtain standardized left and right image pairs.

[0053] Step S102: Input the left and right image pairs into an unsupervised dense matching network that uses pixel-level structural similarity loss (PSSIM loss) as the loss function, and train the network. The unsupervised dense matching network is constructed based on the PSMNetb network structure. The calculation of the PSSIM loss function includes: upsampling the two images to be compared, calculating the similarity using the structural similarity function (SSIM), and averaging the results. The training process of the unsupervised dense matching network specifically includes:

[0054] ① Pair the left and right images with I L ,I R Input an unsupervised dense matching network to obtain the left disparity map D. L .

[0055] This unsupervised dense matching network scheme selects the PSMNetB network structure, which offers a suitable balance between experimental matching accuracy and training speed. The network structure is as follows: Figure 2 As shown, the PSMNetb network sequentially includes a feature pyramid module, a cost volume construction module, and a 3D convolutional aggregation module. The feature pyramid module is used to extract multi-scale feature maps from the input left and right images. The cost volume construction module is used to construct a multi-scale 3D cost volume based on the multi-scale feature maps by matching the correspondence between the left and right feature maps under multiple preset disparities. The 3D convolutional aggregation module is used to regularize the 3D cost volume through stacked 3D convolutional layers, calculate the disparity probability using soft argmin and perform weighted summation regression, and then restore the size through bilinear interpolation upsampling to output the final sub-pixel level disparity map.

[0056] Left disparity diagram D L and input image I L ,I R The relationship between them can be represented as:

[0057] DL =F PSMNetb (θ;I L ,I R (1)

[0058] Among them, F PSMNetb θ represents the main structure of the PSMNetB network, and θ represents the parameters of the PSMNetB network.

[0059] ② Using Spatial Transformation Network (STN), based on Figure I on the right R Compared with left disparity diagram D L The reconstructed image is obtained from the left. STN's function is to reconstruct one image from another based on a displacement field (parallax map), providing geometrically accurate image deformation capabilities.

[0060] Using Figure I on the right R And left disparity diagram D L The left reconstructed image is obtained through an STN network. The reconstruction process can be described as follows:

[0061]

[0062] Among them, F grid F represents the function that generates the sampled grid from the disparity map. STN Represents a spatial transformation network.

[0063] ③ Calculate the original left figure I L Reconstructed image from the left PSSIM loss between.

[0064] like Figure 3 As shown, the core idea of ​​Pixel-Level Structural Similarity Loss Model (PSSIM) is to compare the similarity of two input images at a larger size, thereby relatively reducing the size of the SSIM and keeping the similarity evaluation within the pixel range to improve matching accuracy. Simply put, it can be understood as follows: by upsampling the image, the pixels are "enlarged," and then a small window is used for structural comparison. This allows the similarity evaluation of each pixel to rely almost entirely on its own information, avoiding interference from surrounding pixels, and providing the network with a more accurate and direct pixel-by-pixel matching signal, thus improving the accuracy of disparity map edges and details.

[0065] The formula for calculating PSSIM loss is:

[0066]

[0067] Among them, F avg This represents the averaging function, which averages the loss values ​​over the entire image; F SSIM F represents the structural similarity function. upI1 represents the upsampling function, which enlarges the image (e.g., using bilinear interpolation); I2 represents the two images to be compared, here the original left image I1. L Reconstructed image from the left Or the original right figure I R Reconstructed image with right C piexl_ssim This represents the pixel-level structural similarity loss value.

[0068] The SSIM function uses a 3×3 convolution kernel to calculate local statistics, replacing the Gaussian convolution kernel in traditional SSIM. The SSIM function F SSIM Defined as:

[0069]

[0070] Where I1 and I2 are the two input images, c1 and c2 are constants, preferably c1 = 0.01 and c2 = 0.03, F m It is the mean function, F c It is the covariance function, F s It is a variance function.

[0071] F m (·) is defined as follows:

[0072]

[0073] In the formula, k is the similarity radius. In this scheme, the value is 1, which means that a 3×3 pixel window is used to calculate the local statistics.

[0074] F c (·) is defined as follows:

[0075]

[0076] F s (·) is defined as follows:

[0077] F s (I)=F m (I 2 )-F m (I) 2 (7)

[0078] ④ Input the horizontally flipped left and right images into an unsupervised dense matching network to obtain the right disparity map D. R .

[0079] ⑤ Using a spatial transformation network, based on the horizontally flipped left and right disparity maps D R Reconstruction yields the right reconstructed image. The reconstruction process is described in Formula 2, and the reconstructed right image is then horizontally flipped in reverse.

[0080] ⑥ Calculate the original right figure I R Between and the right reconstructed image after inversion The PSSIM loss between them is calculated using Formula 3.

[0081] ⑦ Based on the two PSSIM loss values ​​mentioned above, backpropagate and update the network parameters.

[0082] This invention also discloses a method for generating a DSM, such as... Figure 4 As shown, it includes the following steps:

[0083] Step S201: Using the above-mentioned unsupervised dense matching network training method, the trained unsupervised dense matching network model is obtained;

[0084] Step S202: Input the left and right images of the region to be tested into the trained network model to generate a disparity map;

[0085] Step S203: Convert the parallax map into distance information using the photogrammetric geometric model and the internal and external orientation elements of the sensor (i.e., the camera's internal parameters and external attitude);

[0086] Step S204: Using distance information, combined with sensor position and orientation (i.e., exterior orientation elements), generate a digital surface model (DSM).

[0087] Corresponding to the above method, embodiments of the present invention also disclose an unsupervised dense matching network training system, the system comprising:

[0088] The image acquisition module is used to acquire the left and right image pairs captured by the binocular vision system;

[0089] The network training module is used to input the left and right image pairs into an unsupervised dense matching network that uses pixel-level structural similarity loss PSSIM as the loss function to train the network. The unsupervised dense matching network adopts the PSMNetb network. The calculation of the PSSIM loss function includes: upsampling the two images to be compared, calculating the similarity using the structural similarity SSIM function, and averaging the calculation results.

[0090] The following is a specific example to better understand this solution.

[0091] (I) Model Training Process

[0092] (1) Parameter settings

[0093] Since unsupervised methods require computation of the network twice per iteration, resulting in significant memory requirements for training, the batchsize parameter is simulated using gradient accumulation and set to 8. The optimizer used is Adam, with β1 and β2 parameters set to 0.9 and 0.999, respectively. Furthermore, to reduce memory overhead during training, the input images are randomly cropped during training, with a crop size of 512 pixels × 256 pixels. This has two advantages: ① It standardizes memory consumption, preventing memory overflow and training termination due to changes in image size, or excessive memory requirements that prevent training from continuing; ② It enhances the randomness of the dense matching dataset, achieving a degree of dataset augmentation. The maximum number of training epochs is set to 200. Considering the differences in network convergence speed during actual training, network convergence is determined by the point at which neither the loss function value nor the test accuracy significantly decreases, at which point training ends.

[0094] (2) Dataset Requirements

[0095] The two images in the dataset are rearranged using epipolar lines, eliminating vertical parallax and requiring no cropping.

[0096] (II) Model Reasoning Process

[0097] The input dataset is cropped according to the computer's video memory size. For a graphics card with 16GB of video memory, the input image size can generally be set to 1000×800. The data is then input into an unsupervised dense matching network to form a disparity map.

[0098] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0099] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0100] The units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations are not considered to be beyond the scope of this invention.

[0101] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Optionally, all or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiments can be implemented in hardware or as a software functional module. This invention is not limited to any particular combination of hardware and software.

[0102] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An unsupervised dense matching network training method, characterized in that, Includes the following steps: Obtain the left and right image pairs of the region to be processed; The left and right image pairs are input into an unsupervised dense matching network that uses pixel-level structural similarity loss, i.e., PSSIM loss, as the loss function, and the network is trained. The unsupervised dense matching network is constructed based on the PSMNetb network structure. The calculation of the PSSIM loss function includes: upsampling the two images to be compared, calculating the similarity using the structural similarity function, i.e., the SSIM function, and averaging the results.

2. The unsupervised dense matching network training method according to claim 1, characterized in that, The training process of the unsupervised dense matching network specifically includes: Pair the left and right images with I L ,I R Input an unsupervised dense matching network to obtain the left disparity map D. L ; Using a spatial transformation network, based on Figure I on the right R Compared with left disparity diagram D L The reconstructed image is obtained from the left. Calculate the original left figure I L Reconstructed image from the left PSSIM loss between; The horizontally flipped left and right images are input into an unsupervised dense matching network to obtain the right disparity map D. R ; Using a spatial transformation network, based on the horizontally flipped left and right disparity maps D R Reconstruction yields the right reconstructed image. And perform a horizontal flip; Calculate the original right figure I R Reconstructed right image after inversion PSSIM loss between; Based on the two PSSIM loss values ​​mentioned above, the network parameters are backpropagated and updated.

3. The unsupervised dense matching network training method according to claim 2, characterized in that, The formula for calculating the PSSIM loss is as follows: Among them, F avg F represents the average function. SSIM F represents the structural similarity function. up Let I1 and I2 represent the upsampling function, and let I1 and I2 represent the two images to be compared. Here, I1 is the original left image. L Reconstructed image from the left Or the original right figure I R Reconstructed image with right C piexl_ssim This represents the pixel-level structural similarity loss value.

4. The unsupervised dense matching network training method according to claim 3, characterized in that, The structural similarity function uses a 3×3 convolution kernel to calculate local statistics, replacing the Gaussian convolution kernel in traditional SSIM.

5. The unsupervised dense matching network training method according to claim 4, characterized in that, The structural similarity function F SSIM Defined as: Where I1 and I2 are the two input images, c1 and c2 are constants, and F m It is the mean function, F c It is the covariance function, F s It is a variance function.

6. The unsupervised dense matching network training method according to claim 1, characterized in that, The PSMNetb network sequentially includes a feature pyramid module, a cost volume construction module, and a 3D convolutional aggregation module. The feature pyramid module is used to extract multi-scale feature maps from the input left and right images. The cost volume construction module is used to construct a multi-scale 3D cost volume based on the multi-scale feature maps by matching the correspondence between the left and right feature maps under multiple preset disparities. The 3D convolutional aggregation module is used to perform regularization processing on the 3D cost volume through stacked 3D convolutional layers, and output the final disparity map through probabilistic processing, disparity regression, and upsampling.

7. A method for generating a DSM, characterized in that, Includes the following steps: The unsupervised dense matching network training method according to any one of claims 1-6 is used to obtain a trained unsupervised dense matching network model; The left and right images of the region to be tested are input into the trained network model to generate a disparity map. The parallax map is converted into distance information using photogrammetric geometry and the interior and exterior orientation elements of the sensor; By utilizing distance information and combining sensor position and orientation, a digital surface model (DSM) is generated.

8. An unsupervised dense matching network training system, characterized in that, For implementing the unsupervised dense matching network training method as described in any one of claims 1-6, the system comprises: The image acquisition module is used to acquire left and right image pairs of the area to be processed; The network training module is used to input the left and right image pairs into an unsupervised dense matching network that uses pixel-level structural similarity loss as the loss function, and train the network. The unsupervised dense matching network is constructed based on the PSMNetb network structure. The calculation of the PSSIM loss function includes: upsampling the two images to be compared, calculating the similarity using the structural similarity function, and averaging the calculation results.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the training method as described in any one of claims 1 to 6.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the training method as described in any one of claims 1 to 6.

Images