DOM position precision quality inspection method and device based on deep learning, and medium
Automatically extract DOM image features through deep learning technology, solving the problems of inefficiency and error in DOM position accuracy quality inspection, achieving efficient and accurate quality inspection in complex scenarios, and enhancing the scope of application and robustness.
Patent Information
- Application Number
- CN202411358590.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-27
- Publication Date
- 2025-01-14
AI Technical Summary
The prior art is inefficient in DOM position accuracy quality inspection, relies on manual inspection and is easy to introduce errors, making it difficult to ensure accuracy and robustness in complex scenarios, and traditional algorithms have high computational complexity and limited scope of application.
Using a deep learning-based method, the deep features of DOM images are automatically extracted using visual big models and deep neural networks, and feature points and feature descriptors are extracted through the improved SuperPoint network, and image matching and position error calculation are combined with pre-trained deep learning models. , generate a quality inspection report.
It improves the efficiency and robustness of DOM position accuracy quality inspection, enhances the adaptability to complex scenarios, avoids the error of manual inspection and the computational complexity of traditional algorithms, and realizes comprehensive coverage inspection of large-scale data.
Smart Images
Figure CN119313730A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of digital image processing, and in particular to a method, device and medium for inspecting the quality of DOM position accuracy based on deep learning. Background Art
[0002] Digital Orthophoto Map (DOM) is a type of digital image product obtained by geometrically correcting and orthorectifying image data acquired by satellite or airborne sensing platforms. It is one of the important achievements in the field of surveying and mapping remote sensing technology. DOM combines the geometric accuracy of the map with the real texture of the image, and is widely used in the fields of digital earth, real-life 3D construction, environmental monitoring, disaster assessment, urban planning, etc.
[0003] The position accuracy of DOM refers to the accuracy of the plane position of the feature points on DOM, which reflects the consistency between the position information of the feature points on DOM and the actual position information. This position accuracy describes the geometric quality of DOM and is an important guarantee for the reliability and accuracy of DOM. However, due to errors introduced in the process of image acquisition, DOM production, algorithm processing, etc., the plane position of the feature points on DOM is inevitably subject to errors. These errors will cause the position of the feature on DOM to be inconsistent with the actual situation, thus affecting subsequent analysis and application. Therefore, the quality inspection of DOM position accuracy is very important to avoid the use of unqualified DOM in subsequent applications.
[0004] In engineering practice, the quality inspection of DOM position accuracy usually adopts the method of manual inspection, which determines the position accuracy of the ground object points on the DOM to be inspected by manually searching the control information of the corresponding position (generally the historical reference DOM or DLG) and manually visually comparing, which heavily relies on the subjective experience of the operators. The manual quality inspection method is inefficient, so for the quality inspection of large-scale DOM data, only spot checks can be adopted instead of full coverage. At the same time, the manual inspection of DOM position accuracy is very easy to introduce errors in the visual puncture points and other links. In recent years, there are some DOM position accuracy quality inspection methods based on traditional algorithms. Such methods usually extract manually designed features such as Harris corner points through traditional image matching algorithms such as AKAZE, SIFT, and ORB. However, such traditional algorithms are very sensitive to factors such as the radiation quality of DOM images, the type and complexity of ground objects, and the ambient light during image acquisition. It is difficult to ensure the accuracy and robustness of feature extraction and matching for DOM in complex scenes. In addition, the computational complexity of traditional algorithms is usually high, and the portability between scenes is poor, which is difficult to meet the actual application needs. Therefore, how to improve the quality inspection efficiency of DOM accuracy, expand its scope of application in various complex scenarios, and enhance its intelligence level has become a technical problem that needs to be solved. Summary of the invention
[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a DOM position accuracy quality inspection method, equipment and medium based on deep learning. In view of the needs of automation of DOM quality inspection and efficient processing of large-scale DOM data, the deep features of DOM images are automatically extracted by using a large visual model and a deep neural network, so that the invention has better generalization ability and robustness for various complex scenes.
[0006] The purpose of the present invention can be achieved by the following technical solutions:
[0007] According to a first aspect of the present invention, a method for checking the quality of DOM position accuracy based on deep learning is provided, comprising the following steps:
[0008] S1. Extract the control point information from the historical reference data and unify the coordinate system. Use the improved SuperPoint network to extract historical DOM feature points as a supplement in the unevenly distributed areas, and intercept the images around the control points to form a control point image library.
[0009] S2, dividing the DOM to be inspected into blocks, searching for images corresponding to the DOM blocks to be inspected in the control point image library, and reviewing the search results in combination with the coordinate information of the corresponding images;
[0010] S3, using the improved SuperPoint network to extract feature points and feature descriptors of the DOM to be inspected, matching the control point image with the corresponding image obtained by the retrieval, obtaining matching point pairs, and eliminating erroneous and redundant matching point pairs;
[0011] S4. Calculate the position information of the control point on the DOM to be inspected according to the matching point pair, count the position error of the DOM to be inspected, and generate a quality inspection report on the DOM position accuracy.
[0012] Furthermore, the improved SuperPoint network in S1 uses a progressive training strategy: training with regular volume data to acquire the ability to extract feature points of regular objects; retraining with public image datasets; labeling a small amount of digital orthophoto feature point data, and fine-tuning the network after secondary training.
[0013] Furthermore, the historical reference data in S1 includes DOM, digital line drawing and control points. The control point image library generated in S1 includes the first control point, the second control point and the third control point: wherein the coordinates of the control point file are directly read as the first control point; the point primitives and line segment primitives in the digital line drawing file in the historical reference data are read, the intersection of the line segment primitives is calculated and the point primitives are added, and the total point primitive coordinates obtained are the second control point; in the DOM area lacking control points, the improved SuperPoint network is used to extract feature points, and the actual coordinates calculated by combining the position information and resolution of the DOM are the third control point.
[0014] Under the criterion of global uniformity of DOM blocks, evenly distributed control points are automatically generated by retaining all the first control points and screening the second and third control points, and the control point neighborhood images are intercepted to generate a control point image library for quality inspection.
[0015] Furthermore, the retrieval of the DOM to be inspected in S2 includes: freezing the pre-trained weights on the visual large model data set; building a feature adaptation module to convert the image features extracted by the visual large model into a DOM retrieval task; establishing a feature extractor for the DOM to be inspected and the control point image library, and outputting the final retrieval probability result; using the trained retrieval network to retrieve the DOM to be inspected in the control point image library, and verifying it with the position information of the DOM.
[0016] Furthermore, the S3 uses the improved SuperPoint network to extract feature points of the DOM to be inspected, matches the control point images in the library with the extracted feature points, and obtains a matching score; based on the matching score, the point distribution uniformity and local consistency constraints are checked, and the constraint form is:
[0017]
[0018] argmint r =argmin||p i ·H i→j -p j || 2 ,
[0019] Among them, H i→j is the homography matrix of the local DOM to be tested, p i and p j is a matching point pair, y r is a local consistency indicator.
[0020] Furthermore, in S4, the coordinates of the control points in the constructed library are converted to the coordinate system of the DOM to be inspected, and the position error between the control points and the matching points on the DOM to be inspected is calculated. The error expression is:
[0021]
[0022] Among them, Δ X and Δ Y is the plane position error of the DOM to be inspected, (X c ,Y c ) is the position of the checkpoint on the DOM to be checked, (X G ,T G ) is the position of the checkpoint on the historical reference data, and H is the homography transformation matrix.
[0023] According to a second aspect of the present invention, there is provided an electronic device, comprising a memory and a processor, wherein a computer program is stored in the memory, and the method described above is implemented when the processor executes the program.
[0024] According to a third aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, wherein the program implements the method described above when executed by a processor.
[0025] Compared with the prior art, the present invention has the following beneficial effects:
[0026] (1) Improved accuracy and efficiency: By introducing deep learning into DOM position accuracy quality inspection, the quality inspection of large-scale DOM data can be completed quickly, which improves efficiency and avoids the problems of inspection errors caused by manual inspection, especially the difficulty of large-scale inspection and strong subjectivity.
[0027] (2) Enhanced robustness and adaptability: By using deep learning algorithms to automatically extract deep features of images, the ability to understand and generalize complex scenes is enhanced. In situations where scene types are rich and environmental conditions change, the pre-trained deep learning model replaces the complex algorithm models designed for multiple technical links. It also overcomes the sensitivity of traditional methods to different scenes and DOM image quality, and has better robustness and adaptability under different scenes and changes in DOM image quality.
[0028] (3) Comprehensive coverage: The present invention can achieve comprehensive coverage inspection of large-scale DOM data, avoiding the problem of missed inspections that may be caused by traditional spot checks. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 It is a flowchart of the steps of the DOM position accuracy quality inspection method based on deep learning;
[0030] Figure 2 It is a schematic diagram of the regular volume data in the progressive training strategy;
[0031] Figure 3 Schematic diagram of public image datasets in progressive training strategy;
[0032] Figure 4 A schematic diagram of a small amount of DOM feature point data in the progressive training strategy;
[0033] Figure 5 This is the architecture diagram of the DOM image retrieval network based on the visual big model and feature adaptation module;
[0034] Figure 6 A diagram of the feature extraction network architecture for DOM quality inspection tasks;
[0035] Figure 7 This is a diagram of the feature matching network architecture for DOM quality inspection tasks;
[0036] Figure 8 The results of other common methods on other image matching datasets are shown;
[0037] Fig. 9 Shown is a rendering of the method provided by the present invention;
[0038] Fig.10 Shown is the distribution map of control points on the historical reference image in DOM quality inspection;
[0039] Fig.11 Feature points provided for historical reference data;
[0040] Fig.12 This is a DOM visualization diagram of the city (#1) scene;
[0041] Fig.13 This is the DOM visualization diagram of the rural (#2) scene;
[0042] Fig.14 This is a comparison chart of the plane accuracy quality inspection results of digital orthophotos. DETAILED DESCRIPTION
[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.
[0044] In recent years, deep learning and artificial intelligence technologies have made breakthrough progress in the field of digital image processing, especially in the fields of image retrieval, image understanding, feature extraction, image matching, etc., showing performance far exceeding that of traditional algorithms. Deep learning algorithms can automatically extract deep features of images, have very strong scene understanding and generalization capabilities, and the deep features proposed can also show good robustness in image matching in complex scenes. Therefore, it is of great application value to introduce deep learning technology into the image retrieval and matching links involved in the quality inspection of the position accuracy of digital orthophotos. To this end, the present invention proposes a method, device and medium for DOM position accuracy quality inspection based on deep learning. This method not only avoids the inefficiency and error introduction of manual inspection, but also overcomes the sensitivity of traditional algorithms to different scenes and DOM image quality, and can be effectively applied to the quality inspection of large-scale DOM data.
[0045] like Figure 1 As shown, a DOM position accuracy quality inspection method based on deep learning is shown, which includes the following steps:
[0046] S1. Extract the control point information from the historical reference data and unify the coordinate system. Use the improved SuperPoint network to extract historical DOM feature points as a supplement in the unevenly distributed areas, and intercept the images around the control points to form a control point image library.
[0047] S2, dividing the DOM to be inspected into blocks, searching for images corresponding to the DOM blocks to be inspected in the control point image library, and reviewing the search results in combination with the coordinate information of the corresponding images;
[0048] S3, using the improved SuperPoint network to extract feature points and feature descriptors of the DOM to be inspected, matching the control point image with the corresponding image obtained by the retrieval, obtaining matching point pairs, and eliminating erroneous and redundant matching point pairs;
[0049] S4. Calculate the position information of the control point on the DOM to be inspected according to the matching point pair, count the position error of the DOM to be inspected, and generate a quality inspection report on the DOM position accuracy.
[0050] The historical reference data in S1 include DOM, digital line drawing and control points. The control point image library generated in S1 includes the first control point, the second control point and the third control point: among them, the coordinates of the control point file are directly read as the first control point; the point primitives and line segment primitives in the digital line drawing file in the historical reference data are read, the intersection of the line segment primitives is calculated and the point primitives are added, and the total point primitive coordinates obtained are the second control point; in the DOM area where control points are missing, the improved SuperPoint network is used to extract feature points, and the actual coordinates calculated by combining the DOM position information and resolution are the third control point.
[0051] The intersection calculation expression of the line segment primitive in the second control point is:
[0052]
[0053] Among them, (x i ,y i ) is the intersection coordinate, (x 1 ,y 1 ) and (x 2 ,y 2 ) is the line segment L 1 The endpoint, (x 3 ,y 3 ) and (x 4 ,y 4 ) is the line segment L 2 The endpoint, u 2×1 It is the ratio of the intersection point to the length of the line segment.
[0054] The coordinates of the two types of control points are converted to the image coordinate system of the historical reference DOM. The conversion expression is:
[0055]
[0056] Among them, (X G ,Y G ,Z G ) is the object coordinate of the control point, The object coordinates of the control point converted to the DOM coordinate system, R gcp→DOM is the rotation matrix for transforming between coordinate systems, T gcp→DOM is the translation vector for conversion between coordinate systems, The corresponding DOM pixel coordinates, (X 0 ,Y 0 ) is the object coordinate corresponding to the upper left corner pixel of DOM, and Res is the spatial resolution of DOM.
[0057] Divide the DOM image of the historical reference data into blocks and count the number of control points on the DOM image blocks. n and uniformity index t ρ , the expression of the uniformity index of the control point distribution is:
[0058]
[0059] where num b is the minimum number of grids for uniformity evaluation, which depends on the size of the minimum grid. is the number of control points in the ith grid, is the average number of points in each grid.
[0060] For tn With t ρ For DOM image blocks that do not reach the threshold, feature points on the image are extracted based on the improved SuperPoint deep neural network. A multi-layer pyramid feature fusion module is added before the SuperPoint network to enhance the network's ability to extract features from scale-difference images.
[0061] The improved SuperPoint network uses a progressive training strategy: using regular volume data for training to obtain the ability to extract regular object feature points; using public image data sets for retraining; annotating a small amount of digital orthophoto feature point data to fine-tune the network after the second training. Figure 2 , Figure 3 and Figure 4 shown.
[0062] Under the criterion of global uniformity of DOM blocks, evenly distributed control points are automatically generated by retaining all the first control points and screening the second and third control points, and the control point neighborhood images are intercepted to generate a control point image library for quality inspection.
[0063] The steps for retrieving the DOM to be inspected in S2 are:
[0064] The visual large model taking DINO as an example is used as the backbone for deep feature extraction of DOM images, and in this embodiment, the pre-trained weights frozen on the large-scale image dataset ImageNet do not participate in the training; a feature adaptation module is built to convert the image features extracted by the visual large model into DOM retrieval tasks; a feature extractor for the DOM to be inspected and the control point image library is established by constructing a twin network architecture that does not share weights, and the cosine similarity score of the descriptor obtained by feature extraction is input into the classifier to output the category tensor corresponding to the final classification score, which is output as the final retrieval probability result through a normalized exponential function (softmax); the trained retrieval network is used to retrieve the DOM to be inspected in the control point image library, and is verified with the location information of the DOM. The network architecture diagram of DOM image retrieval is as follows: Figure 5 shown.
[0065] The overall loss function of the network is Loss, which is composed of the contrast loss L C and classification loss L CE It consists of two parts, and the expression is:
[0066]
[0067] Loss = L C +L CE ,
[0068] Among them, N is the number of training samples, y iis the sample label, p i is the sample prediction probability, d i is the feature distance between sample image pairs, and m is the interval parameter.
[0069] For multiple periods of DOM data at the same location in the historical reference data, a DOM image retrieval dataset is constructed based on DOM location information, and the dataset is used to train the DOM retrieval network constructed above.
[0070] S3 uses the image feature point extraction and description based on the improved SuperPoint deep neural network and the feature matching method based on the SuperGlue network to establish the correspondence between the plane points in the control image in the retrieved library and the DOM to be inspected.
[0071] S3 uses the improved SuperPoint network to extract the feature points of the DOM to be inspected, matches the control point images in the library with the extracted feature points, and obtains the matching score. The network architecture of feature extraction is shown in the figure below. Figure 6 As shown, the network architecture diagram of feature matching is as follows Figure 7 As shown. Based on the matching score, combined with the checkpoint distribution uniformity and local consistency constraints, the wrong and redundant matching point pairs are eliminated to obtain the refined checkpoint pairs used for position accuracy evaluation. The constraint form is:
[0072]
[0073] argmint r =argmin||p i ·H i→j -p j || 2 ,
[0074] Among them, H i→j is the homography matrix of the local DOM to be tested, p i and p j is a matching point pair, t r is a local consistency indicator.
[0075] like Figure 8 Shown are the effects of other common methods on other image matching datasets, such as Fig. 9 Shown is a rendering of the method provided by the present invention, such as Fig.10 The figure shows the distribution of control points on the historical reference image in DOM quality inspection. Fig.11 Feature points provided for historical reference data. Figure 8 and Fig. 9 By comparison, it can be seen that the method proposed in the present invention has advantages in both the accuracy of feature matching and the uniformity of matching point distribution. Fig.10 and Fig.11It can be seen that the present invention also performs well in checkpoint matching in the DOM quality inspection task.
[0076] S4 uses the homography transformation model fitted by matching point pairs as the evaluation model for DOM plane point position errors, and automatically generates quality inspection scores and inspection reports for large-scale DOM data. The specific steps are as follows:
[0077] First, the position coordinates of the checkpoint pairs on the historical reference data and the digital orthophoto to be checked are used to fit the homography transformation model between the images based on the random sampling consistency algorithm. The expression for solving the model parameters is:
[0078]
[0079] Among them, (X c ,Y c ) is the position of the checkpoint on the digital orthophoto to be checked, (X G ,Y G ) is the position of the checkpoint on the historical reference data, H 3×3 is the homography transformation matrix, R 2×2 is the rotation transformation parameter, T 2×1 is the translation transformation parameter, V 1×2 is the transformed edge intersection relationship, S 1×1 is the scale factor.
[0080] The coordinates of the control points in the built library are converted to the coordinate system of the DOM to be inspected, and the position error between the control points and the matching points on the DOM to be inspected is calculated. The error expression is:
[0081]
[0082] Among them, Δ X and Δ Y is the plane position error of the DOM to be inspected, (X c ,Y c ) is the position of the checkpoint on the DOM to be checked, (X G ,Y G ) is the position of the checkpoint on the historical reference data, and H is the homography transformation matrix.
[0083] Finally, the position errors of all checkpoints of the digital orthophoto to be inspected are counted, the total error of the overall position accuracy is calculated, and the proportion of position errors within the 1-fold, 2-fold, and 3-fold total error intervals, as well as the number of checkpoints with gross errors (i.e., the number of checkpoints exceeding the 3-fold total error interval) are counted. Based on the quality inspection standards for surveying and mapping results, the inspection score of the digital orthophoto is calculated, and a quality inspection report on the position accuracy of the digital orthophoto is generated. Assume that the number of checkpoints within the 1-fold, 2-fold, and 3-fold mean error is (n 1 n 2 n3 ), the variance of the plane position error corresponding to each type of point is (S 1 S 2 S 3 ), the number of checkpoints outside the 3-fold mean error is n 4 According to the quality inspection standards for surveying and mapping results, when n 4 When the value is >6, there is no need to calculate the inspection score, and the DOM to be inspected is directly judged as unqualified; for other cases, the formula for calculating the DOM position accuracy quality inspection score is as follows:
[0084]
[0085] Where Tol is the tolerance of the DOM position accuracy in the embodiment. When score<60, it is also determined that the quality of the DOM to be inspected is unqualified, and in other cases, a quality inspection report and inspection score are output.
[0086] In order to verify the effectiveness of the method, DOM data of various complex scenes including cities and villages are selected as the experimental data of this section, and these data are used to test the whole process of digital orthophoto position accuracy quality inspection proposed in this patent. Specifically, the DOM (number of frames: 70), DLG, and control point data (number of points: 1531) measured and obtained in 2021-2022 are used as historical reference data to check the position accuracy quality of the DOM product (number of frames: 500) in 2023. Fig.12 The cities shown (#1) and Fig.13 A DOM product is used for each of the rural (#2) scenes shown for visual display, and the quality inspection reports are shown in Tables 1 and 2.
[0087] Table 1 Urban scene (#1) digital orthophoto position accuracy quality inspection report
[0088]
[0089]
[0090] Table 2 Rural scene (#2) digital orthophoto position accuracy quality inspection report
[0091]
[0092] The method proposed in this embodiment takes about 5 minutes to establish the control point image library. After the database is built, the query, feature extraction, matching, and quality report generation links take less than 30 seconds for each DOM product to be inspected, which has good quality inspection efficiency. The number of matching points for each DOM product to be inspected used to fit the homography change model used for quality inspection is more than 1,000 points. The matching success rate is statistically calculated based on the matching point pair error within 1 pixel. The matching success rate is more than 95%, and the RMSE (Root Mean Square Error) between matching point pairs is about 0.85 pixel. Compared with the operator's image relationship establishment and precision quality inspection strategy of the puncture points, this method has advantages in both efficiency and accuracy. The quality inspection report generated by this method is compared with the 50 DOM image quality inspection reports generated by manual sampling. The inspection conclusions are the same. For the qualified DOM, the score error is within 5 points, which is considered to be the same. The comparison results of the inspection scores are as follows. Fig.14 As shown in the figure, the inspection score -10 means that the DOM inspection is unqualified, and the accuracy of the quality inspection conclusion generated by this method is 96%. In summary, this method can provide an effective and reliable intelligent solution for the position accuracy quality inspection of digital orthophotos.
[0093] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.
[0094] The electronic device of the present invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) or loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The CPU, ROM and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.
[0095] A plurality of components in the device are connected to the I / O interface, including: an input unit, such as a keyboard, a mouse, etc.; an output unit, such as various types of displays, speakers, etc.; a storage unit, such as a disk, an optical disk, etc.; and a communication unit, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit allows the device to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks. The processing unit performs the various methods and processes described above, such as the method of the present invention. For example, in some embodiments, the method of the present invention may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via a ROM and / or a communication unit. When the computer program is loaded into the RAM and executed by the CPU, one or more steps of the method of the present invention described above may be performed. Alternatively, in other embodiments, the CPU may be configured to perform the method of the present invention in any other appropriate manner (e.g., by means of firmware).
[0096] The functions described above herein may be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.
[0097] The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code can be executed entirely on the machine, partially on the machine, partially on the machine as a stand-alone software package and partially on a remote machine, or entirely on a remote machine or server.
[0098] In the context of the present invention, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.
[0099] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.
Claims
1. A DOM position accuracy quality inspection method based on deep learning, characterized in that: The following steps are involved: S1. Extract the control point information from the historical reference data and unify the coordinate system. In the unevenly distributed areas, use the improved SuperPoint network to extract the feature points of the historical digital orthophotos as a supplement, and intercept the images around the control points to form a control point image library. S2, dividing the digital orthophoto to be inspected into blocks, searching for images corresponding to the digital orthophoto blocks to be inspected in the control point image library, and reviewing the search results in combination with the coordinate information of the corresponding images; S3, using the improved SuperPoint network to extract the feature points and feature descriptors of the DOM to be inspected, inputting the feature extraction results into the SuperGlue network constructed by the attention graph neural network and the optimal matching layer, obtaining matching point pairs, and screening and refining the matching point pairs; S4. Calculate the position information of the control point on the digital orthophoto to be inspected according to the matching point pairs, count the position errors of the digital orthophoto to be inspected, and generate a quality inspection report on the position accuracy of the digital orthophoto.
2. According to a method for quality inspection of DOM position accuracy based on deep learning in claim 1, it is characterized in that: The improved SuperPoint network in S1 uses a progressive training strategy; Use regular volume data for training to acquire the ability to extract feature points of regular objects; use public image datasets for retraining; label a small amount of digital orthophoto feature point data and fine-tune the network after secondary training.
3. According to a method for quality inspection of DOM position accuracy based on deep learning in claim 1, it is characterized in that: The historical reference data in S1 include digital orthophotos, digital line drawings and control points.
4. According to a method for quality inspection of DOM position accuracy based on deep learning in claim 1, it is characterized in that: The control point image library generated in S1 includes a first control point, a second control point and a third control point; Among them, the coordinates of the control point file are directly read as the first control point; the point primitives and line segment primitives in the digital line drawing file in the historical reference data are read, the intersection points of the line segment primitives are calculated and the point primitives are added, and the total point primitive coordinates obtained are the second control point; in the digital orthophoto area where control points are missing, the improved SuperPoint network is used to extract feature points, and the actual coordinates solved by combining the position information and resolution of the digital orthophoto are the third control point.
5. A method for inspecting DOM position accuracy based on deep learning according to claim 4, characterized in that: Under the criterion of global uniformity of digital orthophoto blocks, evenly distributed control points are automatically generated by retaining all first control points and screening second and third control points, and the control point neighborhood images are intercepted to generate a control point image library for quality inspection.
6. A method for inspecting DOM position accuracy based on deep learning according to claim 1, characterized in that: The retrieval of the digital orthophoto to be inspected in S2 includes: Freeze the pre-trained weights on the visual large model dataset; build a feature adaptation module to convert the image features extracted by the visual large model into digital orthophoto retrieval tasks; establish a feature extractor for the digital orthophotos to be inspected and the control point image library, and output the final retrieval probability results; use the trained retrieval network to retrieve the digital orthophotos to be inspected in the control point image library, and verify it with the location information of the digital orthophotos.
7. A method for inspecting DOM position accuracy based on deep learning according to claim 1, characterized in that: The S3 uses the improved SuperPoint network to extract the feature points of the digital orthophoto to be inspected, matches the control point image in the library with the extracted feature points, and obtains a matching score; based on the matching score, the point distribution uniformity and local consistency constraints are checked, and the constraint form is: argmin t r =argmin||p i ·H i→j -p j ||2, Among them, H i→j is the homography matrix of the local digital orthophoto to be inspected, p i and p j is a matching point pair, t r is a local consistency indicator.
8. A method for inspecting DOM position accuracy based on deep learning according to claim 1, characterized in that: In S4, the coordinates of the control points in the built database are converted to the coordinate system of the digital orthophoto to be checked, and the position error between the control points and the matching points on the digital orthophoto to be checked is calculated. The error expression is: Among them, Δ X and Δ Y is the plane position error of the DOM to be inspected, (X c ,Y c ) is the position of the checkpoint on the digital orthophoto to be checked, (X G ,Y G ) is the position of the checkpoint on the historical reference data, and H is the homography transformation matrix.
9. An electronic device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 9 is implemented.