A method, system, device and medium for predicting multi-view stereo matching quality

Abstract

The invention discloses a multi-view stereo matching quality prediction method, system, device and medium, and relates to the technical field of multi-view stereo matching. The method comprises: obtaining left and right image film altitude angles, film azimuth angles, solar altitude angles, solar azimuth angles and left and right image film altitude angles, azimuth angles, solar altitude angles, azimuth angle difference values, left and right image intersection angles, and the number of left and right image SIFT same-name points; constructing a multi-view stereo matching quality prediction model; inputting imaging parameter data of the image to be predicted into the multi-view stereo matching quality prediction model, encoding the imaging parameters using an encoder, inputting the encoded features into a decoder for decoding, and obtaining an image pair reconstruction quality prediction value; the method solves the problem of low matching accuracy and efficiency due to uneven quality of multi-view images.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-view stereo matching, and in particular to a multi-view stereo matching quality prediction method, system, device and medium. Background Art

[0002] Image pair selection technology is an effective strategy in remote sensing image processing. It is used to select one or several pairs of images with better reconstruction effects from a large amount of image data for further analysis and processing, which is conducive to improving the efficiency and accuracy of dense matching. For example, a multi-constrained optimal stereo image pair automatic selection method is used to facilitate better dense matching; it automatically selects a set of stereo matching image pairs that meet the requirements through multiple constraints such as baseline constraints, image plane normal vector constraints, feature matching constraints, and intersection angle constraints. This method can improve the accuracy and efficiency of matching, while taking into account matching accuracy and intersection accuracy, and ensures a large overlap between images. However, this method requires each image to be constrained as a reference image, which will increase the amount of calculation when the amount of data is large. A multi-view image selection matching method based on feature point guidance first extracts and matches significant feature points from the reference image, and then evaluates and sorts the quality of the images participating in the matching according to the significance of the matching measure. However, this method mainly selects images through feature points, and cannot effectively select in areas where feature points are sparse. In addition, the amount of calculation is large when processing large-scale data sets, which will increase the running time of the algorithm and affect the matching efficiency. Facciolo G et al. selected image pairs by constructing a correlation matrix to measure the correlation between the parameters calculated from the rational polynomial coefficient (RPC) model of the image and the DSM completeness; through analysis, they determined the three parameters most related to completeness, namely the intersection angle, the maximum deviation from the bottom angle, and the time difference between the left and right images, and concluded that image pairs with intersection angles in the range of 5°-45° and maximum deviation from the bottom angle less than 40° should be preferred. Within this range, all image pairs are sorted in ascending order of the time difference value obtained. However, this method relies too much on the three selected parameters and ignores other parameters that may affect the reconstruction accuracy. It is not very applicable to areas with complex and steep terrain and has poor generalization.

[0003] In summary, the currently used image pair selection method does not fully consider the imaging parameters of the image during the image acquisition process, and ignores other parameters that may affect the reconstruction accuracy. In addition, due to the uneven quality of multi-view images, all images involved in matching cannot achieve the best accuracy. The redundant images increase the time required for matching, resulting in low matching accuracy and efficiency. Summary of the invention

[0004] In view of the shortcomings of the prior art, which has low matching accuracy and efficiency due to the uneven quality of multi-view images, the present invention proposes a multi-view stereo matching quality prediction method, system, device and medium, which performs deep context information extraction through layered input of image parameters, feature selection network and anchor-based regression network, and screens out the most useful features for the model, thereby solving the problems existing in the prior art.

[0005] A multi-view stereo matching quality prediction method comprises the following steps:

[0006] Obtaining two image parameters when shooting the scene to be measured at different viewing angles, which include left image parameters and right image parameters; the left image parameters and the right image parameters both include a picture altitude angle, a picture azimuth angle, a sun altitude angle, and a sun azimuth angle;

[0007] Input two image parameters of the scene to be tested when photographed from different viewing angles into the multi-view stereo matching quality prediction model, encode the image parameters using an encoder, input the encoded features into a decoder for decoding, and obtain an image pair reconstruction quality prediction value; and predict the multi-view stereo matching quality according to the image pair reconstruction quality prediction value;

[0008] The step of encoding the image parameters by using an encoder specifically includes the following steps:

[0009] The image altitude angle and image azimuth angle of the left and right image parameters and the sun altitude angle and sun azimuth angle of the left and right image parameters are respectively used as independent features for splicing and fusion;

[0010] The difference between the left image parameter and the right image parameter is used as the basic feature, and the spliced ​​and fused features are used as the context feature to infer the relationship between the basic feature and the context feature. Based on the relationship between the basic feature and the context feature, the features in the context feature that are related to the predicted value of the image pair reconstruction quality are screened out.

[0011] The basic features are fused with the filtered features as anchor points in the regression network, and are fused with the intersection angle of the left image parameters and the right image parameters, and the number of SIFT same-name points between the left image parameters and the right image parameters to obtain the encoded features.

[0012] Furthermore, after obtaining the two image parameters when shooting at different viewing angles of the scene to be measured, the image parameters are preprocessed; the preprocessing includes performing satellite image plane correction on the image parameters, performing feature matching on the corrected imaging parameter data, and performing regional network adjustment processing on the matched imaging parameter data.

[0013] Furthermore, before the relationship between the basic features and the context features is inferred, the context features are preliminarily screened, and the preliminarily screening includes the following steps:

[0014] The basic feature parameters With context feature parameters Send them into the linear layer to get the basic features and contextual features :

[0015] ,

[0016] ,

[0017] in, and They are basic feature parameters and context feature parameters, is the weight matrix, is the offset vector, is the activation function;

[0018] Will and Perform feature fusion, and use the exponential linear unit as the activation function after the normalization layer to obtain the input for inferring the relationship between basic features and context features:

[0019] ,

[0020] in are the linear layer weights, is the bias term, is a function used to connect two or more arrays, ELU is an activation function, is the normalization function.

[0021] Furthermore, according to the relationship between the basic features and the context features, a gate selection mechanism is used to filter out features in the context features that are related to the predicted values ​​of the image pair reconstruction quality, which specifically includes the following steps:

[0022] The original features are obtained through the linear layer:

[0023] ,

[0024] in are the linear layer weights, is the bias term;

[0025] By adding the basic features and contextual features After being sent to the linear layer for Hadamard product calculation of each dimensional feature product, the feature is compressed through one-dimensional convolution to obtain feature correlation ; Use linear layer to map it to obtain parameter features :

[0026] ,

[0027] ,

[0028] in, and are the weight matrices of basic features and context features respectively, and are the offsets of the basic features and context features respectively, is the weight matrix of feature correlation, is the offset of feature correlation;

[0029] Then we get the initial features of the relationship between the basic features and the context features and ;

[0030] Further feature extraction and expression of the relationship between basic features and context features, inference to obtain deep relationship features :

[0031] ,

[0032] ,

[0033] in, for and The result after feature concatenation and mapping through a linear layer, is the weight matrix, is the offset, and are weight matrices and offsets respectively; ;

[0034] Based on deep relationship Use a gate selection mechanism to select features in the context that are relevant to the prediction of the reconstruction quality of the image pair :

[0035] ,

[0036] ,

[0037] in is the Sigmoid function, It is the feature selected by the gate selection mechanism;

[0038] The features related to the predicted value of image pair reconstruction quality Add to the basic features and normalize, we get:

[0039] ,

[0040] in is the addition of corresponding indexes, is the normalized feature, is the normalization function.

[0041] Furthermore, the cross-attention mechanism is used to fuse the basic features with the filtered features as anchor points in the regression network, and then fused with the intersection angle of the left image parameters and the right image parameters, and the number of SIFT same-name points between the left image parameters and the right image parameters.

[0042] The present invention also includes a multi-view stereo matching quality prediction system, comprising:

[0043] An acquisition module is used to acquire two image parameters when the scene to be measured is photographed at different viewing angles, including left image parameters and right image parameters; the left image parameters and the right image parameters both include a picture altitude angle, a picture azimuth angle, a sun altitude angle, and a sun azimuth angle;

[0044] A prediction module is used to input two image parameters of the scene to be tested when photographed from different viewing angles into a multi-view stereo matching quality prediction model, encode the image parameters using an encoder, input the encoded features into a decoder for decoding, and obtain a predicted value of image pair reconstruction quality; and predict the multi-view stereo matching quality according to the predicted value of image pair reconstruction quality;

[0045] The process of encoding the image parameters by using an encoder specifically includes:

[0046] A feature fusion unit is used to combine and fuse the image altitude angle and image azimuth angle of the left and right image parameters with the sun altitude angle and sun azimuth angle of the left and right image parameters as independent features;

[0047] A screening unit is used to use the difference between the left image parameter and the right image parameter as the basic feature and the spliced ​​and fused feature as the context feature to infer the relationship between the basic feature and the context feature; based on the relationship between the basic feature and the context feature, screen out the features in the context feature that are related to the predicted value of the image pair reconstruction quality;

[0048] The encoding unit is used to fuse the basic features with the filtered features as anchor points in the regression network, and fuse them with the intersection angle of the left image parameters and the right image parameters, and the number of SIFT same-name points between the left image parameters and the right image parameters to obtain the encoded features.

[0049] The present invention also includes a multi-view stereo matching quality prediction computer device, including: a memory, a processor and a computer program stored in the memory, and the processor implements the steps of the multi-view stereo matching quality prediction method when executing the computer program.

[0050] The present invention also includes a readable storage medium, wherein the readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the steps of the multi-view stereo matching quality prediction method are executed.

[0051] The present invention provides a multi-view stereo matching quality prediction method, system, device and medium, which have the following beneficial effects:

[0052] In order to effectively utilize the image parameter information of the left and right images, the present invention inputs the image parameters of the left and right images in layers, and takes the image altitude angle and azimuth angle of the left and right images and the solar altitude angle and azimuth angle as two independent features for splicing and fusion, so that the multi-view stereo matching quality prediction model can better utilize each feature and fully guarantee the relevance and independence of each feature; at the same time, relational reasoning is performed on basic features and context features, and features useful for reconstruction quality prediction values ​​are selected according to the conclusions of relational reasoning, and intersection angles and the number of SIFT same-name points are used as prior knowledge to be fused with anchor points in the regression network, so that the entire model can make full use of effective features, improve the expression ability of the model, and solve the problem of low matching accuracy and efficiency due to the uneven quality of multi-view images. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flowchart of a method for predicting quality of multi-view stereo matching in an embodiment of the present invention;

[0054] Figure 2 This is a schematic diagram of a feature selection module framework in an embodiment of the present invention;

[0055] Figure 3 Schematic diagram of the comparison of the number and accuracy of different fused image pairs in the embodiment of the present invention. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0057] The present invention proposes a multi-view stereo matching quality prediction method, wherein multi-view stereo matching refers to the process of obtaining two-dimensional image information of the same scene from different perspectives of a satellite or camera, and then restoring the three-dimensional structure of the scene through a stereo matching algorithm, thereby reconstructing the scene; the prediction method specifically includes the following steps:

[0058] S1. Obtain image parameters. The image parameters include 14 imaging parameters, namely, left and right image altitude angle, azimuth angle, left and right image solar altitude angle, azimuth angle, left and right image altitude angle, azimuth angle difference value, left and right image solar altitude angle, azimuth angle difference value, left and right image intersection angle, and number of left and right image SIFT same-name points. The difference value is the absolute value of each angle value of the left image minus each angle value corresponding to the right image. Among them, the left and right images refer to two images taken from different perspectives when a satellite shoots the same scene, one is defined as the left image and the other is defined as the right image. The two images are matched to reconstruct the ground object.

[0059] S2. Preprocess the image data. The preprocessing process sequentially carries out satellite image plane correction, feature matching, regional network adjustment and other steps to improve the reliability of the data.

[0060] S3. Construct a multi-view stereo matching quality prediction model R-Transformer, extract and encode the imaging parameters of the input image, use the attention mechanism to fuse multi-source features, and query and refine each feature through the feature selection network and the anchor-based regression network. Finally, the image pair reconstruction quality prediction value is obtained through the decoder. The model is as follows Figure 1 As shown, specifically including:

[0061] (1) Input module: The input part mainly inputs various parameters of the image into the model so that the network can better perform relationship reasoning and feature screening. The present invention mainly uses 14 imaging parameters such as the left image photo altitude angle, photo azimuth angle; the left image solar altitude angle, solar azimuth angle; the right image photo altitude angle, photo azimuth angle; the right image solar altitude angle, solar azimuth angle; intersection angle; the number of SIFT same-name points; the left and right image photo altitude angle, azimuth angle difference value and the left and right image solar altitude angle, azimuth angle difference value to conduct experiments. The time difference between the left and right images was also considered during the experiment, but it was finally found that adding the time difference value would reduce the model performance, so it was finally removed.

[0062] (2) Encoder: It encodes each imaging parameter and outputs the encoded features for the feature selection network to perform relationship reasoning and feature screening. The various imaging parameters of high-resolution images are multi-source and have different impacts on the reconstruction quality. Therefore, the present invention inputs each parameter in layers to allow the network to learn more precisely, thereby learning more effective context information and obtaining the best prediction results.

[0063] ① Hierarchical feature extraction and fusion module: The attention mechanism proposed in Transformer is used in the feature extraction and fusion module. In order to effectively utilize the imaging parameter information of the left and right images, the parameters are input in layers. First, the altitude angle and azimuth angle of the left and right images and the altitude angle and azimuth angle of the sun are taken as two independent features. After linear changes, the features are dimensionally upgraded. Next, the image and sun features are spliced ​​together for the self-attention mechanism, so that they can be considered as the features of the image itself. It is used for both left and right images. Based on the above results, the quality of the left and right images is paired, mainly using the cross-attention mechanism, which allows the information of the two images to interact.

[0064] ② Feature selection network module: After the feature extraction and fusion module, the altitude and azimuth of the left and right images and the solar altitude and azimuth features are encoded. When performing feature fusion, some features that do not need to be paid attention to will make network learning difficult and reduce the expressive power of the model. At the same time, there are a large number of redundant features in high-dimensional features, which also need to be identified and removed by the network. If the above problems are not considered, it will cause network learning difficulties and parameter waste. In order to enable the network to purposefully select and focus on scene elements and give greater weight to elements and features that need to be focused on, a feature selection network module is proposed, which is mainly implemented by a sparse attention mechanism.

[0065] The input of sparse attention is basic features and context features, that is, based on prior knowledge and relational reasoning, basic features aggregate useful context features. First, linear transformations are performed on basic feature parameters and context feature parameters respectively, and recorded as and :

[0066] ,

[0067] ,

[0068] in, and They are basic feature parameters and context feature parameters, is the weight matrix, is the offset vector, is the activation function; then the basic feature parameters With context feature parameters Send them into the linear layer to get the basic features and contextual features . Then and Perform feature fusion, and use the exponential linear unit as the activation function after the normalization layer to obtain the input of the feature selection module based on relational reasoning:

[0069] .

[0070] in are the linear layer weights, is the bias term.

[0071] The gate selection mechanism is further used to filter out the features in the context features that have a greater impact on the prediction results, which specifically includes the following steps:

[0072] The original features are obtained through the linear layer:

[0073] ;

[0074] in are the linear layer weights, is the bias term;

[0075] By adding the basic features and contextual features After being sent to the linear layer for Hadamard product calculation of each dimensional feature product, the feature is compressed through one-dimensional convolution to obtain feature correlation ; Use linear layer to map it to obtain parameter features :

[0076] ,

[0077] ,

[0078] in, and are the weight matrices of basic features and context features respectively, and are the offsets of the basic features and context features respectively, is the weight matrix of feature correlation, is the offset of feature correlation.

[0079] Then we get the initial features of the relationship between the basic features and the context features and .

[0080] Further feature extraction and expression of the relationship between basic features and context features, inference to obtain deep relationship features :

[0081] ;

[0082] ;

[0083] in, for and The result after feature concatenation and mapping through a linear layer, is the weight matrix, is the offset, and are weight matrices and offsets respectively; .

[0084] Based on deep relationship Use the gate selection mechanism to select features in the context that have a greater impact on the prediction results :

[0085] ;

[0086] ;

[0087] in is the Sigmoid function, It is the feature selected by the gate selection mechanism.

[0088] Select useful features Add to the base features and normalize:

[0089] ;

[0090] in is the addition of corresponding indexes, is the normalized feature, is the normalization function.

[0091] Through the feature selection network based on relational reasoning, the basic features are fused with the useful features obtained by screening, which reduces the learning difficulty and achieves efficient fusion of features. Specifically, the basic features are the difference values ​​of the altitude angle and azimuth angle of the left and right images and the difference values ​​of the solar altitude angle and azimuth angle of the left and right images as the basic features, and the context features are the fusion features of the left and right images obtained in the first step, with the purpose of screening out the features that contribute more to the network. Figure 2 The feature selection module is demonstrated.

[0092] ③ Anchor-based regression network: The intersection angle and the number of SIFT homonymous points are used as prior knowledge for network learning. At the same time, the features obtained by the feature selection network are used as anchor points to fuse the intersection angle and the number of SIFT homonymous points. The intersection angle and the number of SIFT homonymous points are linearly transformed and then concatenated, and the cross-attention mechanism is used to obtain the final features.

[0093] (3) Decoder: It consists of two branches: convolutional layer and pooling layer. It mainly performs decoding operations through convolution and pooling to compress features so that the network can better utilize each feature.

[0094] (4) Output the final prediction result.

[0095] S4. Quality assessment: Obtain all possible combinations of stereo pairs, use the Semi-Global Matching (SGM) algorithm as the core dense matching algorithm, and use the Census similarity measure (a local non-parametric similarity measure) as the matching cost. Two DSMs are generated for each stereo pair, one is the pre-interpolation DSM that retains the original DSM, and the other is the post-interpolation DSM obtained using the Delaunay triangulation interpolation method. They are compared with the true value DSM obtained by the LiDAR (Light Detection and Ranging, LiDAR) for accuracy verification. The pre-interpolation DSM is used to evaluate the accuracy of the original 3D points, while the post-interpolation DSM represents the accuracy evaluation of the final product.

[0096] Experimental proof and evaluation:

[0097] (1) Dataset and selected imaging parameters. The present invention uses the multi-view image part of the 2019 Data Fusion Contest (DFC 2019) data set for experiments. The source data includes 42 areas acquired by the Worldview3 satellite in a certain area. These images were taken under different conditions with a spatial resolution of about 0.35 meters, which can provide 8-band multispectral information. Each image is accompanied by its imaging parameters, including the altitude and azimuth of the image, the altitude and azimuth of the sun, the acquisition time, etc. The terrain in the area is mainly plains, mainly composed of settlements, buildings, urban vegetation, roads, and small water bodies. The present invention is not limited to the 14 imaging parameters used. In addition to the known image parameters, the overlap between the left and right images and the deviation from the bottom angle can also be added.

[0098] First, the data needs to be preprocessed, including image plane correction, feature matching, regional block adjustment and other steps to process the experimental data in order to improve the reliability of the data.

[0099] Next, the preprocessed image data are paired two by two to obtain all possible stereo image pair combinations.

[0100] Next, the semi-global matching (SGM) algorithm is used as the core dense matching algorithm, and Census is used as the matching cost. Two DSMs are generated for each stereo pair. One is the DSM before interpolation that retains the original DSM, and the other is the DSM after interpolation obtained using the Delaunay triangulation method. They are compared with the true DSM obtained by LiDAR for accuracy verification. The DSM before interpolation is used to evaluate the accuracy of the original 3D points, while the DSM after interpolation represents the accuracy evaluation of the final product.

[0101] Then, the acquired imaging parameters were used to conduct a grouping experiment. Table 1 lists the imaging parameters selected for this experiment.

[0102] Table 1 Imaging parameters

[0103]

[0104] (2) Evaluation and comparison: In order to verify the effectiveness of the model proposed in the present invention, the present invention selects support vector machine, random forest, K nearest neighbor and Transformer methods to predict the quality of multi-view stereo matching. These methods can establish a complex nonlinear relationship between stereo image quality and image parameters and process high-dimensional data. Table 2 shows the comparison results of the five models under different indicators. It can be found that the performance of the model of the present invention shows better performance under all three indicators.

[0105] Table 2 Performance comparison of different models

[0106]

[0107] Select R 2 There are three evaluation indicators: R-Squared, Root Mean Square Error (RMSE), and Relative Root Mean Square Error (rRMSE). R 2 It refers to the degree to which the independent variable explains the variation of the dependent variable. The value range is 0-1. The closer it is to 1, the higher the degree of explanation of the model for the data. It can be seen that among these five models, the model of the present invention obtains the highest R 2, which means that about 71.8% of the uncertainty can be explained by the selected imaging parameters. The root mean square error (RMSE) is the square root of the ratio of the square of the deviation between the predicted value and the true value to the number of observations n. It measures the deviation between the predicted value and the true value. The smaller the better. From Table 2, it can be seen that the model of the present invention obtains the smallest RMSE, indicating that the model has a good degree of fit. The relative root mean square error (rRMSE) is an indicator used to evaluate the accuracy of the prediction model. It is similar to RMSE, but the difference value is normalized relative to the range of the true value. rRMSE provides a relative evaluation standard that can better compare the performance of different models. A smaller rRMSE indicates that the model has better prediction performance. From Table 2, it can be seen that the model of the present invention obtains a smaller rRMSE value, indicating that the model has better performance.

[0108] Figure 3 The average values ​​of the reconstruction accuracy of the first 20 pairs of fused image pairs in the selected 42 regions were analyzed. Figure 3 It can be seen that when the number of fused image pairs is 13 to 20, the accuracy of DSM does not change much, and when the number of fused image pairs is 8, the accuracy of DSM is the lowest. Therefore, the present invention selects the first 8 image pairs for fusion to analyze the prediction performance of different models. It can be concluded that the fusion effect of the model of the present invention is smoother and the reconstruction accuracy is higher. The model of the present invention shows better results both visually and quantitatively, and the prediction quality is better.

[0109] Based on the same inventive concept, the present invention also proposes a multi-view stereo matching quality prediction system, comprising:

[0110] The acquisition module is used to obtain two image parameters when the scene to be tested is shot from different viewing angles, including left image parameters and right image parameters; the left image parameters and the right image parameters both include film altitude angle, film azimuth angle, solar altitude angle, and solar azimuth angle.

[0111] The prediction module is used to input two image parameters when the scene to be tested is shot from different perspectives into the multi-view stereo matching quality prediction model, encode the image parameters using an encoder, input the encoded features into a decoder for decoding, and obtain an image pair reconstruction quality prediction value; and predict the multi-view stereo matching quality based on the image pair reconstruction quality prediction value.

[0112] The encoder is used to encode the image parameters, and the process specifically includes:

[0113] The feature fusion unit is used to stitch and fuse the image altitude angle and image azimuth angle of the left and right image parameters with the sun altitude angle and sun azimuth angle of the left and right image parameters as independent features.

[0114] The screening unit is used to use the difference between the left image parameters and the right image parameters as the basic features and the spliced ​​and fused features as the context features to infer the relationship between the basic features and the context features; based on the relationship between the basic features and the context features, the features in the context features that are related to the predicted value of the image pair reconstruction quality are screened out.

[0115] The encoding unit is used to fuse the basic features with the filtered features as anchor points in the regression network, and fuse them with the intersection angle of the left image parameters and the right image parameters, and the number of SIFT same-name points between the left image parameters and the right image parameters to obtain the encoded features.

[0116] The present invention also proposes a multi-view stereo matching quality prediction computer device, comprising: a memory, a processor and a computer program stored in the memory, and the processor implements the steps of the multi-view stereo matching quality prediction method when executing the computer program.

[0117] The present invention further proposes a readable storage medium, which stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the multi-view stereo matching quality prediction method.

[0118] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A multi-view stereo matching quality prediction method, characterized in that: The following steps are involved: Obtaining two image parameters when shooting the scene to be measured at different viewing angles, which include left image parameters and right image parameters; the left image parameters and the right image parameters both include a picture altitude angle, a picture azimuth angle, a sun altitude angle, and a sun azimuth angle; Input two image parameters of the scene to be tested when photographed from different viewing angles into the multi-view stereo matching quality prediction model, encode the image parameters using an encoder, input the encoded features into a decoder for decoding, and obtain an image pair reconstruction quality prediction value; and predict the multi-view stereo matching quality according to the image pair reconstruction quality prediction value; The step of encoding the image parameters by using an encoder specifically includes the following steps: The image altitude angle and image azimuth angle of the left and right image parameters and the sun altitude angle and sun azimuth angle of the left and right image parameters are respectively used as independent features for splicing and fusion; The difference between the left image parameter and the right image parameter is used as the basic feature, and the spliced ​​and fused features are used as the context feature to infer the relationship between the basic feature and the context feature. Based on the relationship between the basic feature and the context feature, the features in the context feature that are related to the predicted value of the image pair reconstruction quality are screened out. The basic features are fused with the filtered features as anchor points in the regression network, and are fused with the intersection angle of the left image parameters and the right image parameters, and the number of SIFT same-name points between the left image parameters and the right image parameters to obtain the encoded features.

2. A multi-view stereo matching quality prediction method according to claim 1, characterized in that: After obtaining the two image parameters of the scene to be measured when photographed from different viewing angles, the image parameters are preprocessed; the preprocessing includes performing satellite image plane correction on the image parameters, performing feature matching on the corrected imaging parameter data, and performing regional network adjustment processing on the matched imaging parameter data.

3. The method for predicting multi-view stereo matching quality according to claim 1, characterized in that: Before the relationship between the basic features and the context features is inferred, the context features are preliminarily screened, and the preliminarily screening includes the following steps: The basic feature parameters With context feature parameters Send them into the linear layer to get the basic features and contextual features : ， ， in, and They are basic feature parameters and context feature parameters, is the weight matrix, is the offset vector, is the activation function; Will and Perform feature fusion, and use the exponential linear unit as the activation function after the normalization layer to obtain the input for inferring the relationship between basic features and context features: ， in are the linear layer weights, is the bias term, is a function used to connect two or more arrays, ELU is an activation function, is the normalization function.

4. A multi-view stereo matching quality prediction method according to claim 3, characterized in that: According to the relationship between the basic features and the context features, the gate selection mechanism is used to filter out the features in the context features that are related to the predicted values ​​of the image pair reconstruction quality, which specifically includes the following steps: The original features are obtained through the linear layer: ， in are the linear layer weights, is the bias term; By adding the basic features and contextual features After being sent to the linear layer for Hadamard product calculation of each dimensional feature product, the feature is compressed through one-dimensional convolution to obtain feature correlation ; Use linear layer to map it to obtain parameter features : ， ， in, and are the weight matrices of basic features and context features respectively, and are the offsets of the basic features and context features respectively, is the weight matrix of feature correlation, is the offset of feature correlation; Then we get the initial features of the relationship between the basic features and the context features and ; Further feature extraction and expression of the relationship between basic features and context features, inference to obtain deep relationship features : ， ， in, for and The result after feature concatenation and mapping through a linear layer, is the weight matrix, is the offset, and are weight matrices and offsets respectively; ; Based on deep relationship Use a gate selection mechanism to select features in the context that are relevant to the prediction of the reconstruction quality of the image pair : ， ， in is the Sigmoid function, It is the feature selected by the gate selection mechanism; The features related to the predicted value of image pair reconstruction quality Add to the basic features and normalize, we get: ， in is the addition of corresponding indexes, is the normalized feature, is the normalization function.

5. The method for predicting multi-view stereo matching quality according to claim 1, characterized in that: The cross attention mechanism is used to fuse the basic features with the filtered features and then use them as anchor points in the regression network, and fuse them with the intersection angle of the left image parameters and the right image parameters, and the number of SIFT same-name points between the left image parameters and the right image parameters.

6. A multi-view stereo matching quality prediction system, characterized in that: include: An acquisition module is used to acquire two image parameters when the scene to be measured is photographed at different viewing angles, including left image parameters and right image parameters; the left image parameters and the right image parameters both include a picture altitude angle, a picture azimuth angle, a sun altitude angle, and a sun azimuth angle; A prediction module is used to input two image parameters of the scene to be tested when photographed from different viewing angles into a multi-view stereo matching quality prediction model, encode the image parameters using an encoder, input the encoded features into a decoder for decoding, and obtain a predicted value of image pair reconstruction quality; and predict the multi-view stereo matching quality according to the predicted value of image pair reconstruction quality; The process of encoding the image parameters by using an encoder specifically includes: A feature fusion unit is used to combine and fuse the image altitude angle and image azimuth angle of the left and right image parameters with the sun altitude angle and sun azimuth angle of the left and right image parameters as independent features; A screening unit is used to use the difference between the left image parameter and the right image parameter as the basic feature and the spliced ​​and fused feature as the context feature to infer the relationship between the basic feature and the context feature; based on the relationship between the basic feature and the context feature, screen out the features in the context feature that are related to the predicted value of the image pair reconstruction quality; The encoding unit is used to fuse the basic features with the filtered features as anchor points in the regression network, and fuse them with the intersection angle of the left image parameters and the right image parameters, and the number of SIFT same-name points between the left image parameters and the right image parameters to obtain the encoded features.

7. A computer device for predicting the quality of multi-view stereo matching, characterized in that: include: A memory, a processor, and a computer program stored in the memory, wherein when the processor executes the computer program, the steps of the multi-view stereo matching quality prediction method according to any one of claims 1 to 5 are implemented.

8. A readable storage medium, characterized in that: The readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, they are used to execute the steps of the multi-view stereo matching quality prediction method according to any one of claims 1 to 5.

Images