Target ground object ai identification method based on multi-source aerial photographs

Abstract

The application discloses a kind of target ground object AI identification methods based on multi-source aerial photograph, it is related to aerial photograph processing technical field, including the following steps: S1, the panorama aerial photograph of the region to be identified is collected;S2, the first data set is extracted using the gray level co-occurrence matrix of panorama aerial photograph;S3, according to the response value of all corner points in panorama aerial photograph, the second data set is extracted;S4, the first data set and the second data set are all input to the preliminary screening condition model, and the pixel point meeting the preliminary screening condition model is regarded as the suspected target ground object of the region to be identified;S5, the suspected target ground object of the region to be identified is processed, and the accurate target ground object is obtained.The application is to overcome the limitation of single feature, and multiple information sources are fused, such as the texture features extracted by gray level co-occurrence matrix and the structural features extracted by corner detection, which can more comprehensively describe the characteristics of image control points in aerial photography, improve the accuracy and robustness of identification.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of aerial photograph processing, and in particular to an AI identification method for target ground objects based on multi-source aerial photographs. BACKGROUND

[0002] In the field of aerial photogrammetry and remote sensing image processing, ground control points (GCPs) are key reference points for image geometric correction, spatial positioning and three-dimensional reconstruction. The accurate identification of ground control points directly affects the accuracy and reliability of subsequent surveying and mapping results. For example, by identifying ground control points in the image, the mapping relationship between image pixels and geographic coordinates can be established, and the geometric distortion (such as perspective distortion and lens distortion) of the image can be eliminated. For example, if the ground control points are not accurately identified, the positions of buildings and roads and other geographic features in the image may be shifted or distorted, resulting in insufficient accuracy of the surveying and mapping results. For example, when a city is photographed by a drone, if ground control points are not used for correction, the position error of a building may be magnified to several meters, affecting city planning decisions.

[0003] However, the traditional ground control point identification method mainly relies on manual feature extraction, which has the following technical bottlenecks: 1. Low efficiency and easy to make mistakes, in large-scale aerial images, manual identification of ground control points requires a lot of time and manpower, and is easily affected by subjective experience, resulting in inconsistent or missing labeling results. Especially in complex terrain (such as urban dense areas, mountainous areas) or low-resolution images, the accuracy and efficiency of manual labeling are significantly reduced. 2. Poor adaptability of single image features, traditional methods are mostly based on texture, shape and other features of optical images, but different imaging conditions (such as light changes, occlusions and cloud cover) will cause feature loss or misjudgment. For example, the texture features of the ground control points may be covered by the shadow area, and the shape recognition may be disturbed by the water reflection. 3. Multi-source image information is not fully utilized, modern aerial remote sensing technology lacks the fusion and complementation of multi-source data. 4. The accuracy of automatic identification is insufficient, although some studies attempt to use computer vision techniques (such as edge detection and template matching) to realize automatic identification of ground control points, but these methods have poor robustness in complex scenes, and are difficult to deal with shape variations and background interference of ground control points. SUMMARY

[0004] The application is proposed to solve the above problems, and provides an AI identification method for target ground objects based on multi-source aerial photographs.

[0005] The technical scheme of the application is as follows: an AI identification method for target ground objects based on multi-source aerial photographs comprises the following steps:

[0006] S1, collecting panoramic aerial photographs of a region to be identified;

[0007] S2, extracting a first data set using a gray level co-occurrence matrix of the panoramic aerial photographs;

[0008] S3. Extract the second dataset based on the response values ​​of all corner points in the panoramic aerial photograph;

[0009] S4. Input both the first and second datasets into the initial screening condition model, and use the pixels that meet the initial screening condition model as suspected target features in the area to be identified.

[0010] S5. Process the suspected target features in the area to be identified to obtain accurate target features.

[0011] Furthermore, S2 includes the following sub-steps:

[0012] S21. Obtain the gray-level co-occurrence matrix of the panoramic aerial photograph, and extract the variance of all elements on the diagonal of the gray-level co-occurrence matrix as the adjustment threshold.

[0013] S22. Generate pixel representation parameters for each pixel in the panoramic aerial photograph based on the adjusted threshold.

[0014] S23. Use the pixel representation parameters of all pixels as the first dataset of the panoramic aerial photograph.

[0015] The beneficial effects of the above-mentioned further solutions are as follows: In this invention, the gray-level co-occurrence matrix (GLCM) is a statistical matrix used to describe the spatial dependencies between gray levels in an image, and is generally a symmetric matrix. Extracting texture features from the image using the GLCM captures the spatial dependencies between gray levels, which helps distinguish control points from background features. Using the variance of the elements on the GLCM diagonal as an adjustment threshold makes the feature extraction process more adaptive, capable of handling different lighting and texture variations in images. Through linear transformation, pixel values ​​and texture features (reflected by adjusting the threshold) are combined to generate more discriminative pixel representation parameters.

[0016] Furthermore, in S22, the pixel representation parameters of the pixel point The calculation formula is:

[0017] ;

[0018] In the formula, This represents the maximum grayscale value of all pixels in the panoramic aerial photograph. This represents the grayscale value of a pixel in a panoramic aerial photograph. This represents the first value on the diagonal of the gray-level co-occurrence matrix. This represents the second value on the diagonal of the gray-level co-occurrence matrix. The first element on the diagonal of the gray-level co-occurrence matrix represents the first element. One value, This indicates finding the maximum value. This indicates finding the minimum value. This indicates that the threshold is being adjusted.

[0019] Furthermore, S3 includes the following sub-steps:

[0020] S31. Extract all corner points of the panoramic aerial photograph and determine the Harris response value of each corner point;

[0021] S32. Construct a neighborhood feature column matrix for the corner point based on the Harris response value of the corner point and the corresponding neighboring pixels;

[0022] S33. Construct a weighted feature row matrix for the corner points;

[0023] S34. Construct a second dataset for panoramic aerial photographs based on the neighborhood feature column matrix and weight feature row matrix of each corner point.

[0024] The beneficial effects of the above-mentioned further scheme are as follows: In this invention, the basic idea of ​​the Harris corner response function is to determine the corner point by calculating the gray-level change rate of the pixel in different directions, thus the corner point has a corresponding response value. Combining the Harris response value of the corner point, the gray-level values ​​of the four neighboring pixels, and the average gray-level values ​​of the eight neighboring pixels, multi-scale feature information can be fused, improving the richness and accuracy of feature representation. Dynamically assigning weights based on whether neighboring pixels are corner points allows the feature matrix to more accurately reflect the relationship between the corner point and its neighboring pixels, enhancing the discriminative power of the features.

[0025] Furthermore, S32 includes the following sub-steps:

[0026] S321. Use the gray values ​​of the four neighboring pixels around the corner point as the first column matrix;

[0027] S322. Use the Harris response values ​​at the corner points as linear coefficients;

[0028] S323. Take the average gray value of the eight neighboring pixels corresponding to each four neighboring pixel as the second column matrix.

[0029] S324. Multiply the linear coefficients by the second column matrix and add the result to the first column matrix to obtain the neighborhood feature matrix of the corner point.

[0030] In S321, the first column matrix The expression is: , This represents the grayscale value of the pixel above the corner point. This represents the grayscale value of the pixel below the corner point. This represents the grayscale value of the pixel to the left of the corner point. This represents the grayscale value of the pixel to the right of the corner point.

[0031] In S323, the second column matrix The expression of the second column matrix is: , represents the average value of the gray scale of the 8-neighborhood pixel points corresponding to the pixel point above the corner point, represents the average value of the gray scale of the 8-neighborhood pixel points corresponding to the pixel point below the corner point, represents the average value of the gray scale of the 8-neighborhood pixel points corresponding to the pixel point left to the corner point, represents the average value of the gray scale of the 8-neighborhood pixel points corresponding to the pixel point right to the corner point.

[0032] In S324, the neighborhood feature matrix of the corner point The expression of the neighborhood feature matrix of the corner point is: ; in the expression, represents the Harris response value of the corner point.

[0033] The beneficial effect of the further scheme is that: in the application, the first column matrix (the gray scale of the 4-neighborhood pixel points) directly reflects the gray scale distribution of the pixels around the corner point and captures the local texture information. The second column matrix (the average value of the gray scale of the 8-neighborhood pixel points): by calculating the average value of the gray scale of the 8-neighborhood pixel points, more extensive neighborhood information is introduced, and the robustness of the feature is enhanced. The linear coefficient (the Harris response value): the Harris response value is used as the linear coefficient, the contribution of the second column matrix to the feature matrix is dynamically adjusted, and the feature matrix can more accurately reflect the saliency of the corner point.

[0034] Further, in S33, the weight of the 4-neighborhood pixel point of the corner point is used as the weight feature row matrix.

[0035] The method for determining the weight of the 4-neighborhood pixel point is that: if the neighborhood pixel point is the corner point, the weight of the neighborhood pixel point is , otherwise, the weight of the neighborhood pixel point is , wherein, represents the gradient size value of the corner point, represents a constant.

[0036] The beneficial effect of the further scheme is that: in the application, the weight feature row matrix The expression of the weight feature row matrix is: ; in the expression, represents the weight of the pixel point above the corner point, represents the weight of the pixel point below the corner point, represents the weight of the pixel point left to the corner point, represents the weight of the pixel point right to the corner point, and the gradient size R of the corner point is used as the weight basis, so that the weight distribution is more reasonable and can reflect the change degree of the pixel points around the corner point.

[0037] Further, in S34, the neighborhood feature column matrix of the corner point is multiplied by the weight row matrix to obtain a feature matrix of the corner point, and the mean value of all feature values corresponding to the feature matrix of each corner point is taken as the second data set of the panoramic aerial photograph.

[0038] Further, in S4, the expression of the preliminary screening condition model is:

[0039] ;

[0040] In the formula, represents the gray value of a pixel point in the panoramic aerial photograph, represents the mean value of all parameters in the first data set, represents the mean value of all parameters in the second data set, represents the first weight of the preliminary screening condition model, represents the second weight of the preliminary screening condition model.

[0041] By using reasonable weight parameters, the screening performance of the preliminary screening condition model can be improved.

[0042] Further, in S5, morphological operations are performed on the suspected target ground object to obtain a precise target ground object.

[0043] The beneficial effects of the above further scheme are that: in the present application, morphological operations (such as erosion, dilation, opening operation, closing operation, etc.) are performed on the suspected target ground object, which can remove noise, fill holes and separate adjacent objects, and obtain a more precise target ground object.

[0044] The beneficial effects of the present application are:

[0045] (1) In order to overcome the limitation of a single feature, the present application fuses multiple information sources, such as texture features extracted by a gray level co-occurrence matrix and structure features extracted by corner point detection, which can more comprehensively describe the characteristics of the image control points in aerial photography and improve the accuracy and robustness of recognition.

[0046] (2) The present application fuses texture features and structure features into a preliminary screening condition model, which can fully utilize the advantages of the two kinds of features to improve the accuracy of recognition, and the preliminary screening condition model can be adjusted and optimized according to specific needs to adapt to different image control point types (such as circular, cross-shaped and square).

[0047] (3) The present application can accurately locate the target ground object through further processing of the suspected target ground object, and is suitable for land surveying and mapping, disaster monitoring, urban planning and agricultural yield estimation, etc., especially in emergency mapping, which can quickly extract key image control points from multi-source images to support rapid response and further eliminate false positives and missed detections. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 A flowchart of a target ground object AI recognition method based on multi-source aerial photographs. DETAILED DESCRIPTION

[0049] The embodiments of the present application will be further described below with reference to the drawings.

[0050] As shown in the drawings, the present application provides a target ground object AI recognition method based on multi-source aerial photographs, comprising the following steps: Figure 1

[0051] S1, collecting panoramic aerial photographs of a to-be-recognized region;

[0052] S2, extracting a first data set by using a gray level co-occurrence matrix of the panoramic aerial photographs;

[0053] S3, extracting a second data set according to response values of all corner points in the panoramic aerial photographs;

[0054] S4, inputting the first data set and the second data set into a preliminary screening condition model, and taking pixel points meeting the preliminary screening condition model as suspected target ground objects of the to-be-recognized region;

[0055] S5, processing the suspected target ground objects of the to-be-recognized region to obtain accurate target ground objects.

[0056] In the embodiments of the present application, S2 comprises the following sub-steps:

[0057] S21, obtaining a gray level co-occurrence matrix of the panoramic aerial photographs, and extracting variances of all elements on a diagonal line of the gray level co-occurrence matrix as an adjustment threshold;

[0058] S22, generating a pixel representation parameter for each pixel point in the panoramic aerial photographs according to the adjustment threshold;

[0059] S23, taking the pixel representation parameters of all pixel points as the first data set of the panoramic aerial photographs.

[0060] In the present application, the gray level co-occurrence matrix is a statistical matrix used to describe the spatial dependence relationship between gray levels in an image, and is generally a symmetric matrix. The texture features of the image are extracted by the gray level co-occurrence matrix (GLCM), and the spatial dependence relationship between the gray levels is captured, which is helpful for distinguishing the control points from the background ground objects. The variances of the elements on the diagonal line of the GLCM are used as the adjustment threshold, so that the feature extraction process is more adaptive and can cope with the changes in illumination and texture of different images. The pixel value and the texture features (reflected by the adjustment threshold) are combined by linear transformation to generate more discriminative pixel representation parameters.

[0061] In the embodiments of the present application, in S22, the calculation formula of the pixel representation parameter of the pixel point is: ​

[0062] ;

[0063] In the formula, This represents the maximum grayscale value of all pixels in the panoramic aerial photograph. This represents the grayscale value of a pixel in a panoramic aerial photograph. This represents the first value on the diagonal of the gray-level co-occurrence matrix. This represents the second value on the diagonal of the gray-level co-occurrence matrix. The first element on the diagonal of the gray-level co-occurrence matrix represents the first element. One value, This indicates finding the maximum value. This indicates finding the minimum value. This indicates that the threshold is being adjusted.

[0064] In this embodiment of the invention, S3 includes the following sub-steps:

[0065] S31. Extract all corner points of the panoramic aerial photograph and determine the Harris response value of each corner point;

[0066] S32. Construct a neighborhood feature column matrix for the corner point based on the Harris response value of the corner point and the corresponding neighboring pixels;

[0067] S33. Construct a weighted feature row matrix for the corner points;

[0068] S34. Construct a second dataset for panoramic aerial photographs based on the neighborhood feature column matrix and weight feature row matrix of each corner point.

[0069] In this invention, the basic idea of ​​the Harris corner response function is to determine corners by calculating the rate of grayscale change of pixels in different directions, thus each corner has a corresponding response value. Combining the Harris response value of a corner, the grayscale values ​​of its four neighboring pixels, and the average grayscale value of its eight neighboring pixels allows for the fusion of multi-scale feature information, improving the richness and accuracy of feature representation. Dynamically assigning weights based on whether neighboring pixels are corners enables the feature matrix to more accurately reflect the relationship between corners and their neighboring pixels, enhancing the discriminative power of the features.

[0070] In this embodiment of the invention, S32 includes the following sub-steps:

[0071] S321. Use the gray values ​​of the four neighboring pixels around the corner point as the first column matrix;

[0072] S322. Use the Harris response values ​​at the corner points as linear coefficients;

[0073] S323. Take the average gray value of the eight neighboring pixels corresponding to each four neighboring pixel as the second column matrix.

[0074] S324, multiplying the linear coefficient with the multiplication result of the second column matrix, and adding the first column matrix to obtain a neighborhood feature matrix of the corner point.

[0075] In S321, the first column matrix is expressed as: , represents a gray value of a pixel point above the corner point, represents a gray value of a pixel point below the corner point, represents a gray value of a pixel point left to the corner point, represents a gray value of a pixel point right to the corner point.

[0076] In S323, the second column matrix is expressed as: , represents a mean value of gray values of 8-neighborhood pixel points corresponding to the pixel point above the corner point, represents a mean value of gray values of 8-neighborhood pixel points corresponding to the pixel point below the corner point, represents a mean value of gray values of 8-neighborhood pixel points corresponding to the pixel point left to the corner point, represents a mean value of gray values of 8-neighborhood pixel points corresponding to the pixel point right to the corner point.

[0077] In S324, the neighborhood feature matrix of the corner point is expressed as: ; in the formula, represents a Harris response value of the corner point.

[0078] In the application, the first column matrix (gray values of 4-neighborhood pixel points) directly reflects the gray distribution of the pixels around the corner point, and captures the local texture information. The second column matrix (mean values of gray values of 8-neighborhood pixel points): by calculating the mean values of gray values of 8-neighborhood pixel points, more extensive neighborhood information is introduced, and the robustness of the feature is enhanced. The linear coefficient (Harris response value): the Harris response value is taken as the linear coefficient, the contribution of the second column matrix to the feature matrix is dynamically adjusted, so that the feature matrix can more accurately reflect the saliency of the corner point.

[0079] In the embodiment of the application, in S33, the weight of the 4-neighborhood pixel point of the corner point is taken as a weight feature row matrix.

[0080] The method for determining the weight of the 4-neighborhood pixel point is: if the neighborhood pixel point is the corner point, the weight of the neighborhood pixel point is , otherwise the weight of the neighborhood pixel point is , wherein, represents a gradient size value of the corner point, represents a constant.

[0081] In the present application, the expression of the weight feature row matrix is: ; wherein, represents the weight of the pixel point above the corner point, represents the weight of the pixel point below the corner point, represents the weight of the pixel point on the left side of the corner point, represents the weight of the pixel point on the right side of the corner point, and the gradient size R of the corner point is used as the weight basis to make the weight distribution more reasonable and to reflect the change degree of the pixel points around the corner point.

[0082] In the embodiment of the present application, in S34, the neighborhood feature column matrix of the corner point is multiplied by the weight row matrix to obtain the feature matrix of the corner point, and the mean value of all feature values corresponding to the feature matrix of each corner point is taken as the second data set of the panoramic aerial photograph.

[0083] In the embodiment of the present application, in S4, the expression of the preliminary screening condition model is:

[0084] ;

[0085] ; wherein, represents the gray value of the pixel point in the panoramic aerial photograph, represents the mean value of all parameters in the first data set, represents the mean value of all parameters in the second data set, represents the first weight of the preliminary screening condition model, represents the second weight of the preliminary screening condition model.

[0086] By using reasonable weight parameters, the screening performance of the preliminary screening condition model can be improved.

[0087] In the embodiment of the present application, in S5, morphological operations are performed on the suspected target ground object to obtain the accurate target ground object.

[0088] In the present application, morphological operations (such as erosion, dilation, opening operation, closing operation, etc.) are performed on the suspected target ground object, which can remove noise, fill holes, and separate adjacent objects to obtain more accurate target ground objects.

[0089] Those skilled in the art will appreciate that the embodiments described herein are intended to help the reader understand the principles of the present application and should be understood as not limiting the scope of protection of the present application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations according to the technical inspirations disclosed in the present application without departing from the essence of the present application, and these modifications and combinations are still within the scope of protection of the present application.

Claims

1. A method for AI recognition of target ground objects based on multi-source aerial images, characterized in that, The method comprises the following steps: S1, collecting a panoramic aerial photograph of a to-be-identified region; S2, extracting a first data set by using a gray level co-occurrence matrix of the panoramic aerial photograph; S3, extracting a second data set according to response values of all corner points in the panoramic aerial photograph; S4, inputting the first data set and the second data set into a preliminary screening condition model, and taking pixel points meeting the preliminary screening condition model as suspected target ground objects of the to-be-identified region; S5, processing the suspected target ground objects of the to-be-identified region to obtain accurate target ground objects. The S2 comprises the following sub-steps: S21, obtaining a gray level co-occurrence matrix of the panoramic aerial photograph, and extracting variances of all elements on a diagonal line of the gray level co-occurrence matrix as an adjusted threshold value; S22, generating a pixel representation parameter for each pixel point in the panoramic aerial photograph according to the adjusted threshold value; S23, taking the pixel representation parameters of all pixel points as the first data set of the panoramic aerial photograph; In the S22, the pixel representation parameter of the pixel point The calculation formula is: ； wherein, represents the maximum gray value of all pixel points in the panoramic aerial photograph, represents the gray value of a pixel point in the panoramic aerial photograph, represents the first value on the diagonal line of the gray level co-occurrence matrix, represents the second value on the diagonal line of the gray level co-occurrence matrix, represents the third value on the diagonal line of the gray level co-occurrence matrix, represents the fourth value on the diagonal line of the gray level co-occurrence matrix, represents finding the maximum value, represents finding the minimum value, represents adjusting the threshold value.

2. The multi-source aerial photograph-based target ground object AI identification method according to claim 1, characterized in that, The S3 comprises the following sub-steps: S31, extracting all corner points of the panoramic aerial photograph, and determining Harris response values of the corner points; S32, constructing a neighborhood feature column matrix for a corner point according to a Harris response value of the corner point and corresponding neighborhood pixel points; S33, constructing a weight feature row matrix for the corner point; S34, constructing a second data set for the panoramic aerial photograph according to the neighborhood feature column matrix and the weight feature row matrix of each corner point.

3. The multi-source aerial photograph-based target ground object AI identification method according to claim 2, characterized in that, The S32 comprises the following sub-steps: S321, taking gray level values of four-neighborhood pixel points around the corner point as a first column matrix; S322, taking a Harris response value of the corner point as a linear coefficient; S323, taking mean values of gray levels of eight-neighborhood pixel points corresponding to each four-neighborhood pixel point as a second column matrix; S324, adding a multiplication result of the linear coefficient and the second column matrix to the first column matrix to obtain a neighborhood feature matrix of the corner point.

4. The multi-source aerial photograph-based target ground object AI identification method according to claim 2, characterized in that, In the S33, weights of the four-neighborhood pixel points of the corner point are taken as the weight feature row matrix; The method for determining the weight of the four-neighborhood pixel point is: if the neighborhood pixel point is a corner point, the weight of the neighborhood pixel point is , otherwise, the weight of the neighborhood pixel point is , wherein, represents the gradient size value of the corner point, represents a constant.

5. The multi-source aerial photograph-based target ground object AI identification method according to claim 2, characterized in that, In the S34, the neighborhood feature column matrix and the weight row matrix of the corner point are multiplied to obtain a feature matrix of the corner point, and mean values of all feature values corresponding to the feature matrix of each corner point are taken as the second data set of the panoramic aerial photograph.

6. The multi-source aerial photograph-based target ground object AI identification method according to claim 1, characterized in that, In the S4, an expression of the preliminary screening condition model is: ； wherein represents the gray value of a pixel point in the panoramic aerial photograph, represents the mean value of all parameters in the first data set, represents the mean value of all parameters in the second data set, represents the first weight of the preliminary screening condition model, represents the second weight of the preliminary screening condition model.

7. The multi-source aerial photograph-based target ground object AI identification method according to claim 1, characterized in that, In the S5, morphological operations are performed on the suspected target ground objects to obtain the accurate target ground objects.

Images