An intelligent detection method for towers and satellite image processing system

Abstract

The present invention relates to the technical field of intelligent operation and inspection of power transmission lines, and in particular to an intelligent detection method for towers and a satellite image processing system. The method comprises: acquiring tower spatial information by constructing a tower information database, performing complexity determination on the tower spatial information, and labeling the tower spatial information according to the determination result to acquire tower labeling data; training a tower recognition model according to the acquired tower labeling data; then correcting and fusing the tower detection information to acquire corrected detection information; identifying the corrected detection information according to the tower recognition model, and outputting a tower recognition result. The present invention effectively solves the problems of low recognition accuracy, low efficiency, and high labeling cost of traditional tower recognition methods in large-scale and high-complexity scenarios, thereby improving the accuracy of tower detection, enhancing the adaptability of the model in complex environments, and improving the intelligence level and operation and maintenance efficiency of power inspection.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and inspection of power transmission lines, and in particular to an intelligent detection method for towers and a satellite image processing system. Background Art

[0002] With the rapid development of the power system, the scale of transmission lines continues to expand. Traditional manual inspection methods have problems such as low efficiency, high cost, and high risk, and can no longer meet the efficient and accurate operation and maintenance needs of modern power grids. With the rapid development of satellite remote sensing technology, satellite remote sensing-based tower detection technology has emerged. By using the data information provided by satellites and artificial intelligence algorithms, it is possible to realize the automatic and intelligent identification and monitoring of transmission line towers.

[0003] However, existing tower detection methods based on satellite remote sensing still face many challenges. First, the shape and spatial distribution of towers are complex and diverse. Especially in areas with complex terrain or dense towers, the difficulty of identifying towers increases significantly. Second, data information is affected by factors such as atmospheric conditions, lighting changes, and geometric distortion, resulting in instability of single-source data information, which in turn affects the accuracy of tower detection.

[0004] The information disclosed in this background technology section is only intended to deepen the understanding of the overall background technology of the present disclosure and should not be regarded as an admission or any form of suggestion that the information constitutes the prior art known to those skilled in the art. Summary of the Invention

[0005] The present invention provides an intelligent detection method for a tower and a satellite image processing system, which can effectively solve the problems in the background technology.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] Building a tower information database, and acquiring tower space information according to the tower information database;

[0008] Performing a complexity determination on the pole tower spatial information, and performing a labeling process on the pole tower spatial information according to a complexity determination result to obtain pole tower labeling data;

[0009] Constructing a pole tower recognition model, and training the pole tower recognition model according to the pole tower annotation data;

[0010] The tower detection information is collected, the tower detection information is corrected and integrated to obtain corrected detection information, the corrected detection information is identified according to the trained tower recognition model, and a tower recognition result is output.

[0011] Furthermore, the correction detection information is identified according to the trained tower identification model, and a tower identification result is output, including:

[0012] Performing instance segmentation on the corrected detection information according to the trained tower recognition model to obtain a plurality of image detection results, wherein the image detection results include mask information, and the image detection results correspond to the mask information in a one-to-one manner;

[0013] Performing category screening on the plurality of image detection results to determine whether the image detection results are pole towers, and extracting pole tower contours from the pole tower determination results according to the mask information to obtain a plurality of pole tower image data;

[0014] Calculating confidence levels of a plurality of the tower image data respectively, and filtering the tower image data based on the confidence levels to obtain pixel-level segmentation results;

[0015] The tower detection information is identified according to the pixel-level segmentation result to obtain a tower identification result.

[0016] Furthermore, the confidence levels of the plurality of tower image data are calculated respectively, including:

[0017] Calculating the category probabilities of the plurality of tower image data respectively, wherein the category probabilities are used to indicate the likelihood that the detection result belongs to a tower category;

[0018] Calculating a corresponding target score based on the pole tower image data, wherein the target score indicates whether the pole tower image data contains a real pole tower object;

[0019] The confidence level of the pole tower image data is obtained by performing a weighted combination of the category probability and the target score. The confidence level is used to measure the credibility of the pole tower image data being the pole tower.

[0020] Furthermore, the tower spatial information is annotated according to the complexity determination result to obtain tower annotated data, including:

[0021] Obtaining regular tower information according to the tower space information;

[0022] Determine the boundary features of each tower based on the regular tower information and generate an initial boundary frame diagram corresponding to the tower;

[0023] Adaptively adjusting the initial rectangular frame diagram according to the structure and spatial distribution of the tower, and performing category labeling to obtain a labeled boundary frame diagram;

[0024] The marked boundary frame diagram is formatted to obtain tower marking data.

[0025] Furthermore, the tower spatial information is annotated according to the complexity determination result to obtain tower annotated data, including:

[0026] Acquire special-shaped tower information according to the tower spatial information, wherein the special-shaped tower information includes overlapping and irregular tower images;

[0027] Generating a plurality of initial masks according to the information of the special-shaped tower, extracting tower boundaries respectively according to the plurality of initial masks, and optimizing mask contours based on the tower boundaries to obtain a plurality of optimized masks;

[0028] Performing overlap detection on a plurality of the optimized masks, calculating mask overlapping pixels, determining a mask overlapping state according to the mask overlapping pixels, and adjusting a mask boundary based on the mask overlapping state;

[0029] The adjusted optimization mask is formatted to obtain tower marking data.

[0030] Furthermore, the tower detection information is corrected and integrated to obtain corrected detection information, including:

[0031] Obtaining sensor calibration coefficients according to remote sensing satellites, and obtaining radiation brightness according to the sensor calibration coefficients and the tower detection information;

[0032] Construct an atmospheric correction model, correct atmospheric scattering and atmospheric absorption according to the atmospheric correction model, and correct the solar altitude angle;

[0033] Calculating the true reflectivity of the ground object according to the radiation brightness, adjusting the lighting conditions according to the true reflectivity of the ground object, and obtaining a radiation-corrected image;

[0034] Performing geometric correction on the radiation correction image to obtain a plurality of geometric correction images, and performing image fusion on the plurality of geometric correction images to obtain correction detection information.

[0035] Furthermore, geometric correction is performed on the radiation correction image to obtain a plurality of geometric correction images, including:

[0036] Extracting geometric distortion information from the radiometrically corrected image, selecting ground control points based on the geometric distortion information, and obtaining geometric deviations from the radiometrically corrected image and the ground control points;

[0037] constructing a geometric transformation matrix according to the geometric deviation, performing image reprojection according to the geometric transformation matrix, and obtaining a reprojected image, wherein the image reprojection is used to unify geographic coordinates;

[0038] The reprojected image is precision-checked to obtain a geometrically corrected image.

[0039] Furthermore, a tower information database is constructed, including:

[0040] Collecting historical remote sensing image information, and extracting historical tower images and historical tower spatial information based on the historical remote sensing image information;

[0041] Preprocessing the historical tower images, and making one-to-one correspondence between the preprocessed historical tower images and the historical tower spatial information to construct a tower spatial mapping relationship;

[0042] An index is constructed for the pole tower spatial mapping relationship, and a pole tower information database is established.

[0043] A satellite image processing system, comprising:

[0044] A tower database construction module is used to construct a tower information database and obtain tower space information based on the tower information database;

[0045] A tower labeling optimization module performs complexity determination on the tower spatial information, labels the tower spatial information according to the complexity determination result, and obtains tower labeling data;

[0046] A tower model training module is used to construct a tower recognition model and train the tower recognition model based on the tower annotation data;

[0047] The remote sensing image recognition module collects pole tower detection information, corrects and fuses the pole tower detection information, obtains corrected detection information, recognizes the corrected detection information according to the trained pole tower recognition model, and outputs a pole tower recognition result.

[0048] Furthermore, the tower marking optimization module includes:

[0049] A regular tower extraction unit, which obtains regular tower information according to the tower space information;

[0050] A tower boundary generation unit, which determines the boundary features of each tower based on the regular tower information and generates an initial boundary frame diagram corresponding to the tower;

[0051] A boundary optimization and labeling unit is configured to adaptively adjust the initial rectangular frame diagram according to the structure and spatial distribution of the tower, and to perform category labeling to obtain a labeled boundary frame diagram;

[0052] The annotation data conversion unit converts the format of the annotation boundary frame diagram to obtain the tower annotation data.

[0053] The technical solution of the present invention can achieve the following technical effects:

[0054] It effectively solves the problems of low recognition accuracy and low efficiency of traditional methods in large-scale and highly complex scenarios. By building a pole tower information database and combining deep learning instance segmentation and confidence calculation to optimize the recognition results, the accuracy of pole tower detection is improved. A complexity judgment strategy is adopted to optimize the annotation of regular pole towers and special-shaped pole towers respectively to enhance the adaptability of the model in complex environments. Radiation correction and geometric correction methods are used to improve the processing accuracy of remote sensing images and ensure the accuracy of pole tower location information. Overall, this method realizes efficient, accurate and automated pole tower identification, and improves the intelligence level and operation and maintenance efficiency of power inspection.

[0055] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0057] Figure 1 Schematic diagram of the process of intelligent detection method for towers;

[0058] Figure 2 Schematic diagram of the process of obtaining tower identification results;

[0059] Figure 3 Schematic diagram of the process of obtaining the confidence of tower images;

[0060] Figure 4 Schematic diagram of the architecture obtained for the tower identification results. DETAILED DESCRIPTION

[0061] The technical solutions in the embodiments of the present invention will be clearly and completely described 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.

[0062] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used in this specification are for the purpose of describing specific embodiments only and are not intended to limit the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0063] Embodiment 1;

[0064] like Figure 1 As shown, the present application provides an intelligent detection method for a tower, the method comprising:

[0065] S100: Building a tower information database, and obtaining tower space information according to the tower information database;

[0066] S200: performing complexity determination on the tower space information, and labeling the tower space information according to the complexity determination result to obtain tower labeling data;

[0067] S300: Build a pole tower recognition model and train the pole tower recognition model based on the pole tower annotation data;

[0068] S400: collecting pole tower detection information, correcting and fusing the pole tower detection information, obtaining corrected detection information, identifying the corrected detection information according to the trained pole tower recognition model, and outputting the pole tower recognition result.

[0069] Specifically, a tower information database is first established to store tower spatial information, including: tower geographic location information, tower historical inspection data, tower category information (such as straight towers, corner towers, terminal towers, etc.), tower surrounding environment information (such as terrain features, vegetation coverage) and other relevant tower feature data; then the tower spatial information is judged for complexity to determine the regularity of tower-related data information; in some embodiments, shape matching technology and scale-invariant feature transformation are used to detect the regularity of tower shape and key point distribution; finally, spatial relationship analysis is performed, and GIS tools can be used to analyze the distance and arrangement between towers, and clustering algorithms can be used to detect Clustering pattern, thereby effectively evaluating the complexity of tower data information; then labeling is performed according to the complexity determination result to obtain tower labeling data; on this basis, a tower recognition model can be constructed based on deep learning technology, such as convolutional neural network, and deep learning training can be performed using labeled data. The training process includes steps such as data enhancement, feature extraction, target detection and instance segmentation optimization. In at least one embodiment, transfer learning technology is used to enable the model to adapt to remote sensing images of different resolutions, improve the generalization ability of the model, and combine the intersection-over-union loss to optimize the target positioning accuracy; finally, the trained tower recognition model is used to perform tower detection on the corrected detection information, and the tower recognition result is output.

[0070] The technical solution of the present invention effectively solves the problems of low recognition accuracy and low efficiency of traditional methods in large-scale and highly complex scenarios. By building a tower information database, the accuracy of tower detection is improved; a complexity judgment strategy is adopted to optimize the annotation of regular towers and special-shaped towers respectively, thereby enhancing the adaptability of the model in complex environments; radiation correction and geometric correction methods are used to improve the processing accuracy of remote sensing images and ensure the accuracy of tower location information. Overall, this method realizes efficient, accurate and automated tower identification, and improves the intelligence level and operation and maintenance efficiency of power inspection.

[0071] Further, if Figure 2 and Figure 4 As shown, the correction detection information is identified according to the trained tower recognition model, and the tower recognition results are output, including:

[0072] S410: performing instance segmentation on the correction detection information according to the trained tower recognition model to obtain a plurality of image detection results, wherein the image detection results include mask information, and the image detection results correspond to the mask information one-to-one;

[0073] S420: performing category screening on the plurality of image detection results, determining whether the image detection results are pole towers, and extracting the pole tower contours on the pole tower determination results according to the mask information to obtain a plurality of pole tower image data;

[0074] S430: Calculating the confidence of a plurality of tower image data respectively, and filtering the tower image data based on the confidence to obtain pixel-level segmentation results;

[0075] S440: Identify the tower detection information according to the pixel-level segmentation result to obtain the tower identification result.

[0076] As a preferred embodiment of the above, first, instance segmentation is performed on the correction detection information according to the trained pole tower recognition model. The instance segmentation can be performed in the following manner: the collected satellite remote sensing image is preprocessed, including adjusting the image size and performing color correction to adapt to the input requirements of the model. Then, the processed image is input into the trained pole tower recognition model. The model first identifies potential pole tower candidate areas in the image through the region proposal network. For each candidate area, the model further uses a classifier to determine whether it is a pole tower and accurately adjusts its position through a bounding box regressor. At the same time, the mask branch in the pole tower recognition model is used to generate a corresponding pixel-level mask; the pixel-level mask is a two-dimensional array, in which the value of each pixel point is used to distinguish whether the pixel belongs to a pole tower object. For example, an area with a pixel value of 1 (or other specific value) represents the pixel range of the pole tower, and an area with a pixel value of 1 (or other specific value) represents the pixel range of the pole tower. The area with a pixel value of 0 represents the background or non-tower area. The pixel-level mask can accurately describe the outline of the tower. Finally, by applying non-maximum suppression and morphological operations, the mask is optimized and refined, and the final tower detection and segmentation results are output. Then, by analyzing the image detection results, the morphological features of each tower instance can be calculated, including shape integrity, size ratio, edge clarity, etc., and matched with the standard tower structure. If the shape of a certain detection result deviates seriously from the normal structure of the tower (such as being too slender, too small or too large in area, etc.), the result may be a false detection and needs to be further reviewed or directly eliminated to obtain several tower image data. Afterwards, the confidence of several tower image data is calculated respectively, and the tower image data is filtered based on the confidence to obtain pixel-level segmentation results. The tower detection information is identified according to the pixel-level segmentation results to obtain the tower recognition results.

[0077] Furthermore, if Figure 3 As shown in the figure, the confidence of several tower image data is calculated respectively, including:

[0078] S431: Calculate the category probabilities of a plurality of tower image data respectively, where the category probabilities are used to indicate the possibility that the detection result belongs to the tower category;

[0079] S432: Calculating a corresponding target score based on the tower image data, where the target score indicates whether the tower image data contains a real tower object;

[0080] S433: A confidence level of the pole tower image data is obtained based on a weighted combination of the category probability and the target score. The confidence level is used to measure the degree of credibility that the pole tower image data is a pole tower.

[0081] In this embodiment, the category probabilities of several tower image data can be obtained based on a deep learning network. For example, in target detection models such as Mask R-CNN or YOLO, after identifying a tower object, the network will output a category probability value, indicating the possibility that the object belongs to the tower category. The category probability is usually output by the softmax layer of the deep learning model and ranges from 0 to 1. The higher the category probability, the greater the possibility that the detection result belongs to the tower. Then, the corresponding target score is calculated based on the tower image data. The target score can be calculated based on multiple factors, including: whether the target morphological features (the geometric structure of the tower, such as height, shape, and outline) meet the tower characteristics, whether the target texture features (such as the edge of the tower, material reflectance characteristics, etc.) meet the tower texture pattern, and whether the confidence of the target area is higher than a set threshold. The target score calculation can adopt a target feature extraction method based on a convolutional neural network to extract the morphological and texture features of the tower, and calculate the target score through a classification network (such as an SVM or a fully connected layer). Then, the confidence is calculated by calculating the obtained category probability and the target score. The calculation formula is as follows: , is the confidence of the tower image data, is the class probability, is the target score, , is the weight coefficient, where the weight coefficient is set based on the historical tower recognition results; the confidence of the tower image data is obtained according to the calculation formula.

[0082] Furthermore, the tower spatial information is annotated based on the complexity determination result to obtain tower annotated data, including:

[0083] Obtain regular tower information based on tower space information;

[0084] According to the regular tower information, the boundary features of each tower are determined, and the initial boundary frame diagram of the corresponding tower is generated;

[0085] Adaptively adjust the initial rectangular frame diagram according to the structure and spatial distribution of the tower, perform category labeling, and obtain the labeled boundary frame diagram;

[0086] Convert the format of the annotated boundary box diagram to obtain the tower annotation data.

[0087] Specifically, first, the regular tower image is removed according to the tower spatial information and complexity judgment. After obtaining the regular tower information, it is necessary to generate an initial bounding box for each tower, that is, to determine its circumscribed contour in the image: then, based on the center coordinates of the tower, analyze the spatial footprint of the tower, obtain the outer contour of the tower through image segmentation (such as DeepLabV3, U-Net) or edge detection (Canny, Sobel), calculate the minimum circumscribed rectangle of the tower, and obtain the initial bounding box of the tower. For different types of towers, different initial bounding boxes need to be set. For example, transmission towers: usually have a quadrilateral structure, and the minimum circumscribed rectangle can be used directly. Cement poles: due to their large aspect ratio, the bounding box size needs to be adjusted according to the height of the tower; then based on the tower space The information on the spatial distribution and structure of the towers is obtained. If the towers are densely distributed, it is necessary to avoid overlapping bounding boxes. DBSCAN clustering can be used to optimize the bounding box spacing. If the towers are isolated, the bounding boxes are appropriately increased to ensure that the tower outline is completely covered, thereby obtaining a bounding box diagram. The towers are then categorized based on the bounding box diagram of the towers: the following methods can be used for category labeling: calculate the aspect ratio of the towers, then use the edge detection algorithm to extract the top outline of the towers, and finally combine the structural information of the towers to label the towers and obtain the labeled bounding box diagram of the towers. Finally, in order to adapt to different model training requirements, the labeling format needs to be converted. The format conversion is implemented through Python or labeling tools (such as LabelImg, Roboflow) to obtain the tower labeling data.

[0088] Furthermore, the tower spatial information is annotated based on the complexity determination result to obtain tower annotation data, including:

[0089] Obtaining special-shaped tower information based on tower spatial information, including overlapping and irregular tower images;

[0090] Generating a plurality of initial masks according to the information of the special-shaped towers, extracting the tower boundaries respectively according to the plurality of initial masks, and optimizing the mask contours based on the tower boundaries to obtain a plurality of optimized masks;

[0091] Performing overlap detection on a plurality of optimized masks, calculating mask overlapping pixels, determining a mask overlapping state according to the mask overlapping pixels, and adjusting a mask boundary based on the mask overlapping state;

[0092] Convert the adjusted optimized mask into a new format to obtain the tower marking data.

[0093] As a preferred embodiment of the above, it is first necessary to obtain data of special-shaped towers based on the spatial information of the towers, wherein the special-shaped towers include overlapping towers and tower images with irregular shapes. Due to the complex structure and diverse image presentation of special-shaped towers, it is impossible to directly use the conventional bounding box annotation method to process them. Preferably, an image segmentation method is used, such as the GrabCut algorithm based on edge detection, to obtain the optimized mask of the tower. After obtaining the optimized mask, overlap detection is further performed to identify whether there are multiple tower mask areas overlapping. Specifically, the overlapping area between each mask is calculated, and the number of pixels of the mask overlap is counted. If the number of pixels in the overlapping area accounts for a large proportion, it is determined that there may be overlapping targets in the tower area and segmentation optimization is required. For this purpose, a mask segmentation method can be used, such as a morphological watershed algorithm or a K-means clustering algorithm. Clustering) is used to split the mask of the overlapping area, making the mask area of each tower more independent. In addition, the geometric characteristics of the tower can be combined to further optimize the mask segmentation effect by analyzing the height of the tower, the image lighting characteristics, and the spatial distribution pattern of the tower to ensure that the tower target is not incorrectly merged or split. Finally, after the mask optimization is completed, it needs to be formatted to adapt to the needs of subsequent model training or database storage. Specifically, the optimized tower mask can be converted to a standard annotation format, such as COCO format (a universal annotation format for deep learning target detection and segmentation tasks) or GeoJSON format (suitable for spatial data storage and processing in GIS systems). The format conversion process includes steps such as normalizing the mask coordinates, converting the mask to a standard image size, and generating polygon outline data to ensure the reusability and compatibility of the annotation data and obtain the tower annotation data.

[0094] Furthermore, the tower detection information is corrected and integrated to obtain the corrected detection information, including:

[0095] Obtain sensor calibration coefficients based on remote sensing satellites, and obtain radiation brightness based on the sensor calibration coefficients and tower detection information;

[0096] Construct an atmospheric correction model, correct atmospheric scattering and atmospheric absorption according to the atmospheric correction model, and correct the solar altitude angle;

[0097] Calculate the true reflectivity of the ground object based on the radiation brightness, adjust the lighting conditions according to the true reflectivity of the ground object, and obtain the radiation correction image;

[0098] Perform geometric correction on the radiation correction image to obtain a number of geometric correction images, perform image fusion on the several geometric correction images, and obtain correction detection information.

[0099] In this embodiment, first, based on the sensor characteristics of the remote sensing satellite, the sensor calibration coefficients are obtained, including the sensor's spectral response characteristics, gain, offset value and other parameters. These calibration coefficients are used to perform radiometric correction on the tower detection information: the digital value of the image is converted into a radiometric brightness value, and the formula is as follows: , is the radiance, 、 are the gain and offset coefficients of the sensor, is the pixel value of the original image, where the pixel value of the original image can be obtained through the tower detection information; through this conversion, the influence of the sensor characteristics of the remote sensing image is eliminated, so that the radiation values of different images are comparable; then, for the atmospheric interference of the remote sensing image, an atmospheric correction model is constructed to eliminate the influence of atmospheric scattering and absorption on the image spectrum. The atmospheric correction model can adopt the 6S radiation transfer model; based on the atmospheric correction model, the influence of atmospheric scattering on the image is first calculated, and the influence of aerosol scattering is removed through radiation correction, and the solar altitude angle is corrected. Then, according to the obtained radiation brightness, the radiation brightness is converted into the top atmospheric reflectance. The top atmospheric reflectance still includes the atmospheric scattering. Based on the influence of radiation and absorption, the atmospheric transmittance and atmospheric path radiation are calculated through the atmospheric correction model in combination with the aerosol type, atmospheric water vapor content, observation geometric parameters, etc. collected by remote sensing satellites, and the real reflectivity of the ground objects is calculated. Since the reflectivity of the ground objects should be consistent in different times and different images, it is also necessary to adjust the illumination consistency: histogram matching can be used to adjust the histogram of the image to be processed to the histogram distribution of the reference image to keep its brightness and contrast consistent. After completing the consistency adjustment, the radiation correction image is obtained, and then the radiation correction image is geometrically corrected to obtain several geometric correction images. Several geometric correction images are then fused to obtain correction detection information.

[0100] Furthermore, the radiation correction image is geometrically corrected to obtain several geometrically corrected images, including:

[0101] Extract geometric distortion information from the radiometric correction image, select ground control points based on the geometric distortion information, and obtain geometric deviations based on the radiometric correction image and the ground control points;

[0102] Constructing a geometric transformation matrix according to the geometric deviation, performing image reprojection according to the geometric transformation matrix, obtaining a reprojected image, and the image reprojection is used to unify geographic coordinates;

[0103] Perform accuracy check on the reprojected image to obtain a geometrically corrected image.

[0104] Specifically, first, geometric distortion information is extracted to identify the geometric deformation of remote sensing images during the imaging process. Geometric distortion usually comes from factors such as sensor distortion, platform motion and terrain influence. Sensor distortion can be obtained from sensor geometric parameters (such as focal length, principal point offset, tilt angle, etc.) provided by satellites. Platform motion distortion comes from attitude changes when satellites or drones take images, which can be extracted from satellite ephemeris data or attitude angles (pitch angle, yaw angle and roll angle) recorded by inertial navigation systems (INS). Terrain distortion requires the combination of digital elevation model (DEM) data to analyze the impact of terrain undulations on image geometry. After extracting geometric distortion information, the image is processed according to high-precision maps or GIS. The database selects ground control points, which are reference points for geometric correction. Next, the geometric deviation is calculated, that is, the deviation between the ground control points in the remote sensing image and their corresponding real geographic coordinates. Preferably, the pixel coordinates of the ground control points are calibrated in the image coordinate system, and the geographic longitude and latitude of the corresponding ground control points are obtained in the geographic coordinate system. The errors between the image coordinates and the geographic coordinates are calculated. These errors constitute the geometric deviation data. Subsequently, based on the calculated geometric deviation, a geometric transformation matrix is constructed to convert the original image coordinates into a standard geographic coordinate system. The geometric transformation matrix can be constructed by affine transformation or projection transformation. Then, according to the constructed transformation matrix, image reprojection is performed, that is, the pixels of the remote sensing image are remapped to the standard geographic coordinate system. Image reprojection usually adopts nearest neighbor interpolation, bilinear interpolation or cubic convolution interpolation to maintain image quality and spatial accuracy. After the reprojection is completed, the root mean square error can be used for accuracy verification to obtain a geometrically corrected image.

[0105] Furthermore, the tower information database is constructed, including:

[0106] Collect historical remote sensing image information, and extract historical tower images and historical tower spatial information based on the historical remote sensing image information;

[0107] Preprocess the historical tower images, and make a one-to-one correspondence between the preprocessed historical tower images and the historical tower spatial information to construct a tower spatial mapping relationship;

[0108] Build an index for the spatial mapping relationship of towers and establish a tower information database.

[0109] As a preferred embodiment of the above embodiment, historical remote sensing images are first collected from a historical image library. Multiple periods of high-resolution remote sensing satellite image data without cloud or snow cover, under different solar altitude and azimuth conditions, covering different seasons and time periods, are selected. After obtaining the historical remote sensing images, deep learning target detection technology (such as Faster R-CNN and YOLO) is used in combination with GIS data analysis to perform preliminary detection of tower targets in the images and extract tower image data. At the same time, the spatial information of the historical towers is obtained by combining the GIS spatial data of the towers (such as the latitude and longitude coordinates and elevation information of the towers). Since the historical remote sensing images come from different sources, there may be problems such as inconsistent resolution, different lighting conditions, and image geometric distortion. Therefore, data preprocessing is required, specifically including radiometric correction, geometric correction, and image fusion of the historical tower images. After image preprocessing is completed, a nearest neighbor matching algorithm can be used to find the closest tower location information in the GIS database based on the center coordinates of the tower images and establish a mapping relationship. The database can be constructed using a relational database such as PostgreSQL + PostGIS or a NoSQL database.

[0110] Embodiment 2;

[0111] Based on the same inventive concept as the intelligent detection method for towers and the satellite image processing system in the aforementioned embodiment, the present invention further provides a satellite image processing system, which includes:

[0112] A tower database construction module is used to construct a tower information database and obtain tower space information based on the tower information database;

[0113] The tower marking optimization module determines the complexity of the tower space information, and then marks the tower space information according to the complexity determination result to obtain the tower marking data;

[0114] The tower model training module builds a tower recognition model and trains the tower recognition model based on tower annotation data;

[0115] The remote sensing image recognition module collects pole tower detection information, corrects and fuses the pole tower detection information, obtains the corrected detection information, recognizes the corrected detection information based on the trained pole tower recognition model, and outputs the pole tower recognition result.

[0116] The above-mentioned adjustment system in the present invention can effectively realize an intelligent detection method for towers and a satellite image processing system, and the technical effects that can be achieved are as described in the above-mentioned embodiments and will not be repeated here.

[0117] Furthermore, the tower annotation optimization module includes:

[0118] A regular tower extraction unit obtains regular tower information based on tower space information;

[0119] The tower boundary generation unit determines the boundary features of each tower based on the regular tower information and generates an initial boundary frame diagram of the corresponding tower;

[0120] The boundary optimization annotation unit adaptively adjusts the initial rectangular frame diagram according to the structure and spatial distribution of the tower, and performs category annotation to obtain the annotated boundary frame diagram;

[0121] The annotation data conversion unit converts the format of the annotation boundary box diagram to obtain the tower annotation data.

[0122] Similarly, the above-mentioned optimization schemes for the system can also respectively achieve the corresponding optimization effects of the method in Example 1, which will not be repeated here.

[0123] Although the present application has been described with reference to specific features and embodiments thereof, it is apparent that various modifications and combinations may be made thereto without departing from the spirit and scope of the present application. Accordingly, this specification and drawings are merely illustrative of the present application as defined herein and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, those skilled in the art may make various modifications and variations to the present application without departing from the scope of the present application. Thus, the present application is intended to include such modifications and variations as fall within the scope of the present application and its equivalents.

Claims

1. An intelligent detection method for a tower, characterized in that: The method comprises: Building a tower information database, and acquiring tower space information according to the tower information database; Performing a complexity determination on the tower space information, and labeling the tower space information according to the complexity determination result to obtain tower labeling data, including: Obtaining regular tower information according to the tower space information; Determine the boundary features of each tower based on the regular tower information and generate an initial boundary frame diagram corresponding to the tower; Adaptively adjusting the initial boundary frame diagram according to the structure and spatial distribution of the tower, and performing category labeling to obtain a labeled boundary frame diagram; Convert the marked boundary frame diagram into a format to obtain tower marking data; Also includes: Acquire special-shaped tower information according to the tower spatial information, wherein the special-shaped tower information includes overlapping and irregular tower images; Generating a plurality of initial masks according to the information of the special-shaped tower, extracting tower boundaries respectively according to the plurality of initial masks, and optimizing mask contours based on the tower boundaries to obtain a plurality of optimized masks; Performing overlap detection on a plurality of the optimized masks, calculating mask overlapping pixels, determining a mask overlapping state according to the mask overlapping pixels, and adjusting a mask boundary based on the mask overlapping state; Convert the adjusted optimization mask into a new format to obtain tower marking data; Constructing a tower recognition model, and training the tower recognition model according to the tower annotation data; The tower detection information is collected, the tower detection information is corrected and integrated to obtain corrected detection information, the corrected detection information is identified according to the trained tower recognition model, and a tower recognition result is output.

2. The intelligent detection method for a tower according to claim 1, characterized in that: Identifying the correction detection information according to the trained tower identification model and outputting a tower identification result includes: Performing instance segmentation on the corrected detection information according to the trained tower recognition model to obtain a plurality of image detection results, wherein the image detection results include mask information, and the image detection results correspond to the mask information in a one-to-one manner; Performing category screening on the plurality of image detection results to determine whether the image detection results are pole towers, and extracting pole tower contours from the pole tower determination results according to the mask information to obtain a plurality of pole tower image data; Calculating confidence levels of a plurality of the tower image data respectively, and filtering the tower image data based on the confidence levels to obtain pixel-level segmentation results; The tower detection information is identified according to the pixel-level segmentation result to obtain a tower identification result.

3. The intelligent detection method for a tower according to claim 2, characterized in that: Calculating the confidence of the plurality of tower image data respectively, including: Calculating the category probabilities of the plurality of tower image data respectively, wherein the category probabilities are used to indicate the likelihood that the detection result belongs to a tower category; Calculating a corresponding target score based on the tower image data, wherein the target score indicates whether the tower image data contains a real tower object; The confidence level of the pole tower image data is obtained by performing a weighted combination of the category probability and the target score. The confidence level is used to measure the credibility of the pole tower image data being the pole tower.

4. The intelligent detection method for a tower according to claim 1, characterized in that: Correcting and fusing the tower detection information to obtain corrected detection information includes: Obtaining sensor calibration coefficients according to remote sensing satellites, and obtaining radiation brightness according to the sensor calibration coefficients and the tower detection information; Construct an atmospheric correction model, correct atmospheric scattering and atmospheric absorption according to the atmospheric correction model, and correct the solar altitude angle; Calculating the true reflectivity of the ground object according to the radiation brightness, adjusting the lighting conditions according to the true reflectivity of the ground object, and obtaining a radiation-corrected image; The radiation correction image is geometrically corrected to obtain a plurality of geometric correction images, and the plurality of geometric correction images are image-fused to obtain correction detection information.

5. The intelligent detection method for a tower according to claim 4, characterized in that: Performing geometric correction on the radiation correction image to obtain a plurality of geometric correction images, including: Extracting geometric distortion information from the radiometrically corrected image, selecting ground control points based on the geometric distortion information, and obtaining geometric deviations from the radiometrically corrected image and the ground control points; constructing a geometric transformation matrix according to the geometric deviation, performing image reprojection according to the geometric transformation matrix, and obtaining a reprojected image, wherein the image reprojection is used to unify geographic coordinates; The reprojected image is precision-checked to obtain a geometrically corrected image.

6. The intelligent detection method for a tower according to claim 1, characterized in that: Build a tower information database, including: Collecting historical remote sensing image information, and extracting historical tower images and historical tower spatial information based on the historical remote sensing image information; Preprocessing the historical tower images, and making one-to-one correspondence between the preprocessed historical tower images and the historical tower spatial information to construct a tower spatial mapping relationship; An index is constructed for the pole tower spatial mapping relationship, and a pole tower information database is established.

7. A satellite image processing system, characterized in that: Using the tower intelligent detection method according to claim 1, the system includes: A tower database construction module is used to construct a tower information database and obtain tower space information based on the tower information database; A tower labeling optimization module performs complexity determination on the tower spatial information, labels the tower spatial information according to the complexity determination result, and obtains tower labeling data; A tower model training module is used to construct a tower recognition model and train the tower recognition model based on the tower annotation data; A remote sensing image recognition module collects pole tower detection information, performs correction and fusion on the pole tower detection information, obtains corrected detection information, recognizes the corrected detection information according to the trained pole tower recognition model, and outputs a pole tower recognition result; The tower marking optimization module includes: A regular tower extraction unit, which obtains regular tower information according to the tower space information; A tower boundary generation unit, which determines the boundary features of each tower based on the regular tower information and generates an initial boundary frame diagram corresponding to the tower; A boundary optimization and labeling unit is configured to adaptively adjust the initial boundary frame diagram according to the structure and spatial distribution of the tower, and to perform category labeling to obtain a labeled boundary frame diagram; The annotation data conversion unit converts the format of the annotation boundary frame diagram to obtain the tower annotation data.

Images