Satellite image parallax estimation method and system based on multi-scale geometric coding and texture decoding

Abstract

The invention discloses a satellite image parallax estimation method and system based on multi-scale geometric coding and texture decoding, and belongs to the field of satellite image parallax estimation, and the method comprises the steps: obtaining to-be-measured satellite image data; inputting obtained satellite image data to be measured into the trained satellite image parallax estimation model to obtain a parallax estimation map of the satellite image; wherein the training of the satellite image parallax estimation model comprises the following steps: constructing a satellite image data set; building a satellite image parallax estimation model; and training the constructed satellite image parallax estimation model by using the constructed satellite image data set to obtain a trained satellite image parallax estimation model. According to the method, the satellite image parallax estimation model is established to perform parallax estimation on the satellite image, so that texture inconsistency in a parallax updating process is relieved, the accuracy of parallax estimation is improved, geometric details of the image can be captured, and the parallax prediction capability of a network on complex terrain structures such as stacked and slender buildings is improved.

Description

Technical Field

[0001] The present invention belongs to the field of satellite image parallax estimation, and particularly relates to a satellite image parallax estimation method and system based on multi-scale geometric coding and texture decoding. Background Art

[0002] The parallax estimation of satellite stereo images is an important research topic in the fields of photogrammetry and remote sensing. This task aims to obtain accurate matching points for each pixel between a pair of epipolar rectified satellite stereo images and generate a parallax map. By utilizing the relationship between parallax and depth, the three-dimensional information of the scene can be easily reconstructed. Therefore, this technology has been widely applied in multiple fields, including 3D city reconstruction and digital earth platforms, etc.

[0003] After decades of development, the existing parallax estimation methods can be roughly divided into two categories: traditional manually designed methods and learning-based methods. Traditional manually designed methods usually include the following steps: matching cost calculation, cost aggregation, preliminary parallax calculation, and parallax refinement. Representative methods include semi-global matching, belief propagation, and graph cut. These algorithms are highly dependent on parameter selection, with poor robustness and low efficiency. In addition, their performance often significantly degrades in stacked and slender buildings. Learning-based methods mainly rely on the powerful feature extraction and representation capabilities of neural networks. Such methods learn the complex patterns and features between data through convolutional neural networks, thereby improving the parallax estimation ability. Generally, the process of such methods includes: constructing a cost volume using a convolutional neural network, then aggregating or filtering the cost volume using the network, and finally generating a parallax map. However, the texture of the parallax maps generated by such methods is often not clear enough and not smooth enough.

[0004] Generally speaking, there are still the following two difficult problems in the current parallax estimation of satellite stereo images: (1) Satellite stereo images are captured by optical sensors, while the ground truth parallax maps are obtained by LiDAR sensors. Due to the inherent differences between these two types of sensors, the inconsistency in texture levels between satellite stereo images and ground truth parallax maps is inevitable. Usually, satellite stereo images are dense and smooth, while ground truth parallax maps are sparse and rough. In addition, due to factors such as occlusion, the ground truth parallax of some regions is missing, further exacerbating the sparsity of the parallax map; (2) Complex terrain structures (such as stacked and slender buildings) often have indistinguishable geometric features. The existing stereo matching methods for satellite stereo images usually have difficulty distinguishing these geometric features, resulting in conflicts at the feature level and local matching errors in parallax estimation. Therefore, how to obtain a dense and smooth parallax map and be able to distinguish complex terrain structures is the key problem in satellite stereo image parallax estimation.

[0005] Therefore, it is necessary to design a satellite image parallax estimation method and system based on multi-scale geometric coding and texture decoding for the above problems. Summary of the Invention

[0006] The object of the present invention is to provide a satellite image parallax estimation method and system based on multi-scale geometric coding and texture decoding for the problems existing in the prior art. For the key problems of satellite image parallax estimation such as how to obtain a dense and smooth parallax map and distinguish complex terrain structures, by constructing a multi-scale geometric feature encoder to encode the geometric information and local features of satellite stereo images, the geometric feature information of complex terrain structures can be effectively captured. By constructing a texture decoding module to further enhance the multi-scale geometric coding body, a texture representation in the feature space is generated; by constructing a parallax iteration unit, the features are gradually decoded and a parallax map is generated. By integrating high-level texture information, the texture inconsistency in the parallax update process is effectively alleviated, and the geometric details of the image can also be captured, improving the parallax prediction ability of the network for complex terrain structures such as stacked and slender buildings; by constructing a parallax refinement module, taking the reconstruction error of the satellite image as the input, the generated parallax map is further refined, thereby improving the accuracy of parallax estimation.

[0007] According to one aspect of the present specification, a satellite image parallax estimation method based on multi-scale geometric coding and texture decoding is provided, including:

[0008] Obtain satellite image data to be measured;

[0009] Input the obtained satellite image data to be measured into a trained satellite image parallax estimation model to obtain a parallax estimation map of the satellite image; wherein, the training of the satellite image parallax estimation model includes:

[0010] Obtain satellite image data and corresponding true parallax estimation maps, and construct a satellite image data set;

[0011] Build a satellite image parallax estimation model, including: a multi-scale geometric feature encoder (MSGFE) for encoding the geometric information and local features of satellite images; a texture decoding module (TDM) for further enhancing the multi-scale geometric coding body and generating a texture representation in the feature space; a parallax iteration unit jointly guided by the multi-scale geometric feature encoder and the texture decoding module for gradually decoding features and generating a parallax map; a parallax refinement module (DRM) for taking the reconstruction error of the satellite image as the input and further refining the generated parallax map;

[0012] Use the constructed satellite image data set to train the built satellite image parallax estimation model to obtain a trained satellite image parallax estimation model.

[0013] Further, encode the geometric information and local features of the satellite image, including:

[0014] Perform successive downsampling and geometric encoding on the input satellite image to obtain geometric features at multiple scales respectively;

[0015] Perform successive upsampling, geometric encoding, and channel dimension concatenation on the geometric features at multiple scales.

[0016] Further, further enhance the multi-scale geometric encoding body to generate a texture representation in the feature space, including:

[0017] Perform calculations on the multi-scale geometric encoding body to obtain , and the calculation formula is as follows:

[0018]

[0019] Then perform calculations on the obtained to obtain , and the calculation formula is as follows:

[0020]

[0021] Continue to perform calculations on the calculated and to generate a high-level texture representation , and the generation calculation formula is as follows:

[0022]

[0023] where represents the multi-scale geometric encoding body, represents the activation function, represents convolution, represents average pooling, represents the sigmoid function, represents element-wise multiplication.

[0024] Further, gradually decode the features and generate a disparity map, including:

[0025] a. Calculate the correlation body of the input satellite image, perform three times of downsampling on the correlation body, and construct a correlation body pyramid;

[0026] b. Initialize the hidden state of the update iteration unit , as follows:

[0027]

[0028] c. Initialize the disparity = 0, sample the correlation body pyramid to obtain a cost body , where the sampling radius is ;

[0029] d. Based on , , and input update iteration unit, update the hidden state to obtain , use the disparity head to modulate the 1 / 4 scale hidden state in to obtain the disparity residual and update the disparity to obtain , as shown in the following formula:

[0030]

[0031]

[0032] From step a to step d, complete one update of the disparity iteration unit and generate a disparity map.

[0033] Furthermore, further refine the generated disparity map, including:

[0034] Complete N updates of the disparity through the disparity iteration unit to obtain ;

[0035] Upsample to obtain , use to perform grid-based resampling on the right image of the satellite image, and subtract it from the left image of the satellite image to obtain the reconstruction error , and the calculation formula is as follows:

[0036]

[0037] where is the grid-based resampling operation;

[0038] Pass and the reconstruction error through the convolutional layer and splice them along the channel dimension, as shown in the following formula:

[0039]

[0040] Pass the obtained into the U-shaped network and convolutional layer to obtain the disparity residual, and correct to obtain the finally generated disparity map .

[0041] According to one aspect of the present specification, there is provided a satellite image parallax estimation system based on multi-scale geometric encoding and texture decoding, including:

[0042] A data acquisition module for acquiring satellite image data to be measured;

[0043] A parallax estimation module for inputting the acquired satellite image data to be measured into a trained satellite image parallax estimation model to obtain a parallax estimation map of the satellite image; wherein, the training of the satellite image parallax estimation model includes:

[0044] Acquiring satellite image data and corresponding true parallax estimation maps to construct a satellite image data set;

[0045] Building a satellite image parallax estimation model, including: a multi-scale geometric feature encoder for encoding the geometric information and local features of the satellite image; a texture decoding module for further enhancing the multi-scale geometric encoding body to generate a texture representation in the feature space; a parallax iteration unit jointly guided by the multi-scale geometric feature encoder and the texture decoding module for gradually decoding features and generating a parallax map; a parallax refinement module for using the reconstruction error of the satellite image as an input to further refine the generated parallax map;

[0046] Training the built satellite image parallax estimation model using the constructed satellite image data set to obtain a trained satellite image parallax estimation model.

[0047] According to one aspect of the present specification, there is provided an electronic device including a memory and a processor, the memory storing a computer program, wherein the processor, when executing the computer program, implements the steps of the satellite image parallax estimation method based on multi-scale geometric encoding and texture decoding.

[0048] According to one aspect of the present specification, there is provided a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the satellite image parallax estimation method based on multi-scale geometric encoding and texture decoding.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] 1. Aiming at the sparsity problem of the parallax map, the present invention encodes the geometric information and local features of the satellite stereo image by constructing a multi-scale geometric feature encoder, can effectively capture the geometric feature information of complex terrain structures, and further enhances the multi-scale geometric encoding body by constructing a texture decoding module to generate a texture representation in the feature space.

[0051] 2. In view of the texture-level inconsistency problem between satellite images and disparity maps, the present invention constructs a disparity iteration unit to gradually decode features and generate a disparity map. By integrating high-level texture information, it effectively alleviates the texture inconsistency during the disparity update process and can also capture the geometric details of the image, improving the network's disparity prediction ability for complex terrain structures such as stacked and slender buildings.

[0052] 3. In view of the conflict at the feature level and the local matching error problem in disparity estimation, the present invention constructs a disparity refinement module, which takes the reconstruction error of the satellite image as input to further refine the generated disparity map, thereby improving the accuracy of disparity estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or in the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0054] Figure 1 It is the overall framework diagram of the embodiment of the present invention;

[0055] Figure 2 It is the schematic diagram of the sampling process of the embodiment of the present invention;

[0056] Figure 3 (a) It is the schematic diagram of the MSGFE structure of the embodiment of the present invention;

[0057] Figure 3 (b) It is the schematic diagram of the GEB structure of the embodiment of the present invention;

[0058] Figure 4 (a) It is the schematic diagram of the Multi-scale SRU structure of the embodiment of the present invention;

[0059] Figure 4 (b) It is the schematic diagram of the DRM structure of the embodiment of the present invention;

[0060] Figure 5 and Figure 6 It represents the comparison diagram of disparity estimation of different methods in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0061] It should be noted that:

[0062] In the present invention, the superscripts with special meanings are the numbers in brackets (if they are not numbers in brackets, they have no special meanings and are only for conveniently representing variables). If such superscripts act on variables, they represent the XX scale of the original image resolution. For example: if the size of the original image is H×W, then it means the size of the variable is H / 16×W / 16. If such superscripts act on operators, they represent multiples. For example:

[0063] In the present invention, the subscripts with special meanings are the numbers in brackets (if they are not numbers in brackets, they have no special meanings and are only for conveniently representing variables), and they represent the repetition times of a certain module. For example: It means to reuse GEB three times.

[0064] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present invention.

[0065] As Figure 1 shown, the embodiments of the present invention also provide a process for satellite image parallax estimation based on multi-scale geometric coding and texture decoding, including: obtaining satellite image data to be measured; inputting the obtained satellite image data to be measured into a trained satellite image parallax estimation model to obtain a parallax estimation map of the satellite image; wherein, the training of the satellite image parallax estimation model includes: obtaining satellite image data and corresponding true parallax estimation maps, and constructing a satellite image data set; building a satellite image parallax estimation model, including: a multi-scale geometric feature encoder for encoding the geometric information and local features of the satellite image; a texture decoding module for further enhancing the multi-scale geometric coding body and generating a texture representation in the feature space; a parallax iteration unit for gradually decoding features and generating a parallax map, and at the same time integrating high-level texture information and capturing the geometric details of the satellite image; a parallax refinement module for using the reconstruction error of the satellite image as an input to further refine the generated parallax map; training the built satellite image parallax estimation model with the constructed satellite image data set to obtain a trained satellite image parallax estimation model.

[0066] Specifically, the embodiments of the present invention also provide a method for extracting the features of the left image and the right image using the MSGFE with shared parameters ( ) and the multi-scale geometric coding body (including three scales of 1 / 4, 1 / 8, and 1 / 16 ), including the following sub-steps, where Figure 3 (a) is the structural schematic diagram of MSGFE, Figure 3 (b) is the structural schematic diagram of GEB:

[0067] 1. Downsample the left and right images by a factor of 4, and then input them into 2 cascaded geometric encoding modules GEB to obtain geometric features at a scale of 1 / 4. :

[0068] (1)

[0069] Where, represents the i-th cascaded GEB, represents downsampling by a factor of j.

[0070] 2. Downsample by a factor of 2, and then input it into 3 cascaded geometric encoding modules GEB to obtain geometric features at a scale of 1 / 8. :

[0071] (2)

[0072] 3. Downsample by a factor of 2, and then input it into 3 cascaded geometric encoding modules GEB to obtain geometric features at a scale of 1 / 16. :

[0073] (3)

[0074] 4. Downsample by a factor of 2, and then input it into 4 cascaded geometric encoding modules GEB. Then, upsample it by a factor of 2 and perform channel dimension concatenation with to obtain :

[0075] (4)

[0076] (5)

[0077] Where, represents channel dimension concatenation, represents upsampling by a factor of i.

[0078] 5. Input into 3 cascaded geometric encoding modules GEB. Then, upsample it by a factor of 2 and perform channel dimension concatenation with to obtain :

[0079] (6)

[0080] 6. For input three cascaded geometric encoding modules GEB, then perform 2x upsampling, and concatenate with along the channel dimension to obtain :

[0081] (7)

[0082] Input 2 and the convolutional layer to obtain left and right features :

[0083] (8)

[0084] (9)

[0085] Among them, represents a convolutional layer with a size of 3×3. Unless otherwise specified below, the size is 3×3.

[0086] Specifically, the embodiment of the present invention also provides the specific calculation process of GEB as follows:

[0087] 1. Assume the input of GEB is , first use three convolutional layers to obtain the key Q, query K, and value V, then calculate the self-attention, and obtain the global feature of one branch through pooling :

[0088] (10)

[0089] (11)

[0090] (12)

[0091] 2. Then obtain the weight map through the Conv, GELU, Conv, and sigmoid functions :

[0092] (13)

[0093] Among them, represents the sigmoid function.

[0094] 3. On the other branch, use DWConv for V to obtain the local feature :

[0095] (14)

[0096] Subsequently, generate the weight map :

[0097] (15)

[0098] Then, weighted to obtain features :[[]]

[0099] (16)

[0100] Among them, represents element-wise multiplication.

[0101] 4. Pass through the feed-forward network to obtain the output of GEB ,

[0102] (17)

[0103] (18)

[0104] Among them, represents the LayerNorm layer, represents an intermediate variable.

[0105] Specifically, the embodiment of the present invention also provides a method for generating a high-level texture representation using the designed TDM (including three scales of 1 / 4, 1 / 8, and 1 / 16), and the calculation process is as follows:

[0106] (19)

[0107] (20)

[0108] (21)

[0109] Among them, represents average pooling, represents the sigmoid function.

[0110] Specifically, the embodiment of the present invention also provides a method for calculating the 1-level correlation volume between the features of the left and right images:

[0111] (22)

[0112] Among them, the letters i, j, k, h represent element indices, , are both three-dimensional matrices.

[0113] Specifically, the embodiment of the present invention also provides a method for performing three downsamplings on the last dimension of the 1-level correlation volume to construct a correlation volume pyramid .

[0114] Specifically, the embodiment of the present invention further provides to initialize and update the hidden state of the Multi-scale SRU (including three scales of 1 / 4, 1 / 8, and 1 / 16):

[0115] (23)

[0116] Among them, the hidden state ; initialize the disparity = 0, and sample the correlation volume pyramid according to the initialized disparity to obtain the cost volume , where the sampling radius is , and the sampling process is as Figure 2 shown.

[0117] Specifically, the embodiment of the present invention will obtain the , , , input to the update unit Multi-scale SRU to update the hidden state, and obtain (including three scales of 1 / 4, 1 / 8, and 1 / 16). Use the disparity head to modulate the 1 / 4-scale hidden state in to obtain the disparity residual and update the disparity to obtain , where Figure 4 (a) is the structural schematic diagram of the Multi-scale SRU:

[0118] (24)

[0119] (25)

[0120] So far, a single update process is completed.

[0121] Specifically, the embodiment of the present invention further provides the specific steps of the Multi-scale SRU:

[0122] 1. At the 1 / 16 scale, input , and the 2-fold downsampled into the SRU to obtain the updated .

[0123] 2. At the 1 / 8 scale, input , , the 2-fold downsampled and the 2-fold upsampled into the SRU to obtain the updated .

[0124] 3. At the 1 / 4 scale, , , and the 2x upsampled input is fed into the SRU to obtain the updated . Thus, the update process of the hidden state is completed.

[0125] The specific formula of the SRU is:

[0126] (26)

[0127] Where, represents the output of the GRU with a 3x3 convolutional layer, represents the output of the GRU with a 7x7 convolutional layer. The calculation process of the GRU is:

[0128] (27)

[0129] (28)

[0130] (29)

[0131] (30)

[0132] Where, , represents the input hidden state of the GRU, represents the variable obtained by concatenating other inputs except the hidden state along the channel dimension, is the output hidden state of the GRU.

[0133] Then, the cost volume is repeatedly sampled from the correlation volume pyramid according to the initialized disparity, and the obtained , , , is input into the update unit Multi-scale SRU to update the hidden state. These two steps are repeated N times for the disparity to obtain .

[0134] Specifically, the embodiment of the present invention also provides to perform 4x upsampling to obtain , and use to perform grid-based resampling on the right image , and subtract it from the left image to obtain the reconstruction error :

[0135] (31)

[0136] Among them, is a grid-based resampling operation.

[0137] Specifically, the embodiments of the present invention also provide and reconstruction error input into the DRM to obtain the final disparity map , which specifically includes the following sub-steps, where Figure 4 (b) is the structural schematic diagram of the DRM:

[0138] 1. Pass the disparity and the reconstruction error through the convolutional layer and splice them together through the channel dimension:

[0139] (32)

[0140] 2. Input the obtained into the Unet and convolutional layer to obtain the disparity residual, and correct to obtain the finally generated disparity map :

[0141] (33)

[0142] Specifically, the embodiments of the present invention also provide calculating the loss function and optimizing the parameters of the overall network composed of all the above steps with this:

[0143] (34)

[0144] Among them, represents the true disparity value, represents the weight adjustment factor, specifically 0.9, is the smooth L1 function, and its expression is:

[0145] (35)

[0146] Specifically, the embodiments of the present invention also provide inputting the test image into the trained overall network model, obtaining the predicted disparity map, and using the PSMNet and HMSMNet methods for comparison. The satellite image inspection estimation results corresponding to different methods are as Figure 5 、 6 shown.

[0147] Specifically, in order to quantitatively evaluate the disparity estimation results, the embodiments of the present invention select the US3D dataset for comparison and introduce EPE (average endpoint error, the lower the better) and D1 (the percentage of endpoint errors greater than 3 pixels, the lower the better) as evaluation indicators. The quantitative comparison results are shown in Table 1.

[0148] Table 1 Quantitative Analysis of Different Parallax Estimation Methods

[0149]

[0150] As can be seen from Table 1, the method proposed by the present invention uses a multi-scale geometric coding body and high-level texture representation as a guide to obtain a continuous and smooth parallax map, and has the strongest parallax prediction ability for complex ground object structures.

[0151] The implementation basis of each embodiment of the present invention is achieved through programmed processing by a device with a processor function. Therefore, in engineering practice, the technical solutions and functions of each embodiment of the present invention are encapsulated into various modules. Based on this actual situation, on the basis of the above embodiments, an embodiment of the present invention provides a satellite image parallax estimation system based on multi-scale geometric coding and texture decoding, which is used to execute a satellite image parallax estimation method based on multi-scale geometric coding and texture decoding in the above method embodiments.

[0152] The system includes: a data acquisition module for acquiring satellite image data to be measured; a parallax estimation module for inputting the acquired satellite image data to be measured into a trained satellite image parallax estimation model to obtain a parallax estimation map of the satellite image; wherein, the training of the satellite image parallax estimation model includes: acquiring satellite image data and corresponding true parallax estimation maps to construct a satellite image data set; building a satellite image parallax estimation model, including: a multi-scale geometric feature encoder for encoding the geometric information and local features of the satellite image; a texture decoding module for further enhancing the multi-scale geometric coding body to generate a texture representation in the feature space; a parallax iteration unit for gradually decoding features and generating a parallax map, while integrating high-level texture information and capturing the geometric details of the satellite image; a parallax refinement module for using the reconstruction error of the satellite image as an input to further refine the generated parallax map; and training the built satellite image parallax estimation model using the constructed satellite image data set to obtain a trained satellite image parallax estimation model.

[0153] The satellite image parallax estimation system based on multi-scale geometric coding and texture decoding provided by the embodiments of the present invention addresses the sparsity problem of the parallax map and the texture-level inconsistency problem between the satellite image and the parallax map. By using several modules, it further enhances the multi-scale geometric coding body by constructing a texture decoding module to generate a texture representation in the feature space. By constructing a parallax iteration unit, it gradually decodes the features and generates a parallax map. By integrating high-level texture information, it effectively alleviates the texture inconsistency during the parallax update process and can also capture the geometric details of the image, improving the network's parallax prediction ability for complex terrain structures such as stacked and slender buildings. By constructing a parallax refinement module, taking the reconstruction error of the satellite image as the input, it further refines the generated parallax map, thereby improving the accuracy of parallax estimation.

[0154] Based on the same inventive concept as the foregoing embodiments, the embodiments of the present invention also provide an electronic device, including a memory and a processor. The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement a satellite image parallax estimation method based on multi-scale geometric coding and texture decoding as proposed in the above embodiments.

[0155] The embodiments of the present invention also provide a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it overcomes the problems that the parallax map is sparse and rough, there is a texture-level inconsistency between the satellite image and the parallax map, and complex terrain structures have indistinguishable geometric features. It effectively alleviates the texture inconsistency during the parallax update process, can capture the geometric details of the image, and improves the network's parallax prediction ability for complex terrain structures such as stacked and slender buildings.

[0156] This storage medium can be any non-volatile storage device such as a hard disk, a solid-state drive, a flash drive, an optical disc, etc., for storing computer program code and necessary data files. The stored computer program includes: a data acquisition module and a parallax estimation module.

[0157] Finally, it should be noted that the above specific embodiments are only relatively representative examples of the present invention. Obviously, the present invention is not limited to the above specific embodiments and can have many variations. Any simple modification, equivalent change, and modification made to the above specific embodiments based on the technical essence of the present invention shall be considered to fall within the protection scope of the present invention.

Claims

1. A satellite image parallax estimation method based on multi-scale geometric coding and texture decoding, characterized in that Including: Obtain satellite image data to be measured; Input the obtained satellite image data to be measured into the trained satellite image disparity estimation model to obtain a disparity estimation map of the satellite image; wherein, the training of the satellite image disparity estimation model includes: Obtain satellite image data and corresponding true disparity estimation maps, and construct a satellite image data set; Build a satellite image disparity estimation model, including: a multi-scale geometric feature encoder for encoding the geometric information and local features of the satellite image; a texture decoding module for further enhancing the multi-scale geometric encoding body to generate a texture representation in the feature space; a disparity iteration unit jointly guided by the multi-scale geometric feature encoder and the texture decoding module for gradually decoding features and generating a disparity map; a disparity refinement module for using the reconstruction error of the satellite image as input to further refine the generated disparity map; Use the constructed satellite image data set to train the built satellite image disparity estimation model to obtain a trained satellite image disparity estimation model.

2. The satellite image parallax estimation method based on multi-scale geometric coding and texture decoding according to claim 1, characterized in that, Encoding the geometric information and local features of the satellite image includes: Perform successive downsampling and geometric encoding on the input satellite image to obtain geometric features at multiple scales respectively; Perform successive upsampling, geometric encoding, and channel dimension splicing on the geometric features at multiple scales.

3. A satellite image parallax estimation method based on multi-scale geometric coding and texture decoding according to claim 1, characterized in that Further enhancing the multi-scale geometric encoding body to generate a texture representation in the feature space, including: Perform calculations on the multi-scale geometric encoded volume to obtain , and the calculation formula is as follows: , Then calculate the obtained to obtain , and the calculation formula is as follows: , The calculated and continue to calculate the high-level texture representation , and the generation calculation formula is as follows: , Among them represents a multi-scale geometric coding body represents an activation function represents convolution represents average pooling represents the sigmoid function represents element-wise multiplication 4. A method for satellite image parallax estimation based on multi-scale geometric coding and texture decoding according to claim 1, characterized in that, Gradually decoding features and generating a disparity map, including: a. Calculate the correlation volume of the input satellite image, perform three times of downsampling on the correlation volume, and construct a correlation volume pyramid; b. Initialize the hidden state of the update iteration unit , as shown in the following formula: , c. Initialize the disparity = 0, and sample the correlation volume pyramid to obtain a cost volume , where the sampling radius is ; d. Based on , , and input update and iteration unit to update the hidden state to obtain , and use the disparity head to modulate the 1 / 4 scale hidden state in to obtain the disparity residual and update the disparity to obtain , as shown in the following formula: , , From step a to step d, complete one update of the disparity iteration unit and generate a disparity map.

5. A satellite image parallax estimation method based on multi-scale geometric coding and texture decoding according to claim 1, characterized in that Further refining the generated disparity map, including: Complete the disparity through the disparity iteration unit for N updates to obtain ; Up-sample to obtain . Using to perform grid-based resampling on the right image of the satellite image , and subtract it from the left image of the satellite image to obtain the reconstruction error . The calculation formula is as follows: , Among them, is a grid-based resampling operation; Combine and the reconstruction error through a convolutional layer and concatenate them along the channel dimension as follows: , The obtained input U-shaped network and convolutional layer to obtain the disparity residual, and perform correction to obtain the finally generated disparity map , completing the construction of the disparity refinement module.

6. A satellite image parallax estimation system based on multi-scale geometric coding and texture decoding, characterized in that Including: An acquisition data module for obtaining satellite image data to be measured; A disparity estimation module for inputting the obtained satellite image data to be measured into the trained satellite image disparity estimation model to obtain a disparity estimation map of the satellite image; wherein, the training of the satellite image disparity estimation model includes: Obtain satellite image data and corresponding true disparity estimation maps, and construct a satellite image data set; Build a satellite image disparity estimation model, including: a multi-scale geometric feature encoder for encoding the geometric information and local features of the satellite image; a texture decoding module for further enhancing the multi-scale geometric encoding body to generate a texture representation in the feature space; a disparity iteration unit jointly guided by the multi-scale geometric feature encoder and the texture decoding module for gradually decoding features and generating a disparity map; a disparity refinement module for using the reconstruction error of the satellite image as input to further refine the generated disparity map; Use the constructed satellite image data set to train the built satellite image disparity estimation model to obtain a trained satellite image disparity estimation model.

7. An electronic device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the satellite image disparity estimation method based on multi-scale geometric encoding and texture decoding 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 the processor, it implements the steps of the satellite image disparity estimation method based on multi-scale geometric encoding and texture decoding according to any one of claims 1 to 5.

Images