Illegal building identification, monitoring and detection method and system based on deep learning

Abstract

The present invention is applicable to the field of image processing technology, and provides a method and system for identifying, monitoring and detecting illegal buildings based on deep learning, the method comprising: step S1, image uploading step: acquiring image data through satellite or drone and uploading the image data to the server, and automatically completing the preprocessing of the image data during uploading; step S2, illegal building comparison step: selecting two periods of image data to be compared, and outputting the building change spots through comparison detection at the server; step S3, automatic data storage step: step S4, result verification step: by superimposing the extracted building change spots on the corresponding online image, further verifying the correctness of the extracted spot data. The present invention uses an algorithm to identify illegal buildings, and can quickly and accurately determine whether a building is an illegal building. The whole process is simple to implement, and only needs to train the building recognition model, which can quickly identify whether the outline and height of the building have changed, and then after identifying the building change spots of the illegal building, further verification and confirmation can be performed, which can significantly improve the recognition efficiency.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and in particular relates to a method and system for identifying, monitoring and detecting illegal buildings based on deep learning. Background Art

[0002] In recent years, the national policy level has successively issued a number of guiding opinions on the investigation and punishment of illegal buildings in urban management. The work of investigating and punishing illegal buildings has sprung up across the country. It is possible to determine whether a building is an illegal building by observing the changes in the building.

[0003] The current building image detection solution is generally to obtain high-resolution satellite images regularly. Several mainstream technical solutions have emerged:

[0004] (1) On-site photo collection solution based on mobile APP: Provide an APP program for collecting illegal construction information. During regular patrols, if staff find suspected illegal construction, they will take photos and use the APP to register and upload evidence.

[0005] (2) Visual interpretation and collection scheme based on satellite images: Regularly obtain high-resolution satellite images. Based on the satellite image display platform, two images from different periods are displayed on the same screen. Through manual visual judgment, vectorized outlines of suspected illegal buildings that have changed are generated, and illegal building maps are generated and uploaded to the server simultaneously.

[0006] (3) Visual patrol scheme based on drone video: Through the real-time video transmitted by the drone during flight, the staff visually identifies suspected illegal buildings and automatically takes screenshots and uploads them to the server.

[0007] (4) Identification solution based on remote sensing technology: The image data is extracted through remote sensing technology to obtain building patches, and the changed patches are found by manual visual comparison. The changed patches are uploaded to the server as suspected illegal buildings.

[0008] It can be seen that the current traditional judgment schemes have the following shortcomings:

[0009] (1) Traditional illegal building inspection schemes are easily affected by the difficulty of collecting evidence. The APP collection scheme cannot collect evidence of illegal buildings in buildings within the hospital or high-rise buildings, and requires the full participation of staff, which is a heavy workload.

[0010] (2) Although visual interpretation of satellite or UAV images has alleviated the difficulty of obtaining evidence, it still requires a lot of work to rely on pixel-by-pixel comparison by the human eye.

[0011] (3) The visual interpretation solution for video patrols requires a lot of manual work over a large area, and since the video lacks a bird's-eye view, it will be more troublesome to verify it through playback at a later stage.

[0012] (4) Traditional solutions based on remote sensing technology require high-precision geometric correction of images, but in real data situations, it is difficult to ensure that image data from different periods can be "perfectly" superimposed under geometric correction. Moreover, these algorithms only complete the automatic extraction of buildings, without further automatic comparison of changes between different buildings. Manual participation is still required to complete the task, which cannot free up manpower to the greatest extent.

[0013] (5) Most image-based solutions lack the emphasis on the storage, rapid sharing and visualization of massive images, resulting in a lot of manpower required for image management.

[0014] In short, the current judgment schemes require technical personnel intervention, increase judgment uncertainty, and are time-consuming and labor-intensive. Summary of the invention

[0015] In view of the above problems, the purpose of the present invention is to provide a method and system for identifying, monitoring and detecting illegal buildings based on deep learning, aiming to solve the technical problems that the existing illegal building judgment methods are low in accuracy and time-consuming and labor-intensive.

[0016] The present invention adopts the following technical solution:

[0017] On the one hand, the illegal building identification, monitoring and detection method based on deep learning includes the following steps:

[0018] Step S1, image uploading step: acquiring image data through satellite or drone and uploading the image data to the server, automatically completing image data preprocessing during uploading, including resolution resampling, geometric correction, and image slice pyramid construction;

[0019] Step S2, illegal building comparison step: selecting two periods of image data to be compared, and outputting a building change map including building outline changes and height changes through comparison detection at the server end;

[0020] Step S3, automatic data storage step: automatically importing the extracted building change spots into the achievement database;

[0021] Step S4, result verification step: by superimposing the extracted building change spots onto the corresponding online images, further verify the correctness of the extracted spot data, and automatically save the final result.

[0022] Furthermore, the acquired data includes orthophotos and oblique photography data, and the specific process of step S2 is as follows:

[0023] S21. For the same area, two phases of orthophotos are obtained, and raster patches of building outline changes are obtained by building extraction and building outline change detection, and further converted into vector patches;

[0024] S22, for the area, obtaining oblique photography data of the period corresponding to the two periods of orthophotos, and obtaining a building height change vector map patch by building height change detection;

[0025] S23, merging the building outline change vector patch and the height change vector patch to obtain an overall building change pattern patch.

[0026] Furthermore, step 21 specifically includes the following steps:

[0027] The deep learning-based building recognition model is trained by collecting sample data, wherein the sample data collection includes collecting the required high-resolution remote sensing data, recording the shooting time and height of the high-resolution remote sensing data, preprocessing the data, and resampling to the required resolution; using the sample data image to be outlined as the electronic map base map, outlining the building spots in the image in the form of a vectorization toolbox, and realizing online vectorization operations. The vectorized building spots will be saved in the sample library for online call training;

[0028] Obtain two phases of orthophotos of the same area, use the building recognition model to predict the building images in the two phases of orthophotos, and obtain the building patches;

[0029] The building patch is used to generate a simplified building outline vector using Tensor Polygonization, and the building outline change vector patch is obtained based on the changes in the building outline vectors in two phases.

[0030] Furthermore, the process of generating a simplified building outline vector is as follows:

[0031] Generate a building thermal map through a building recognition model to obtain a building vector;

[0032] According to the principle of building shape characteristics, the building vector is smoothed to eliminate the jagged edges in the generated vector;

[0033] According to the directions generated in the building recognition model, the building corners are identified and the corner positions are marked;

[0034] Delete the vectors between corners and optimize the location of generated building outline vectors.

[0035] Furthermore, step S22 specifically includes the following steps:

[0036] Obtain the oblique photography data corresponding to the two periods of orthophotos, merge the triangulated networks of the oblique photography data, and form a single large-scale oblique photography data;

[0037] Use the building spot outline data to be superimposed on the merged oblique photography data, wherein the oblique photography data has high-level information of all pixels, obtain the highest high-level information within the oblique photography data range as the high-level value, and write the high-level value as an attribute field into the current building spot outline data;

[0038] After obtaining the building outline data with high-rise values ​​in two periods, the high-rise values ​​are compared after matching according to the range coordinates to obtain the height change vector map of the building where the high-rise has changed.

[0039] Furthermore, the process of constructing the image slice pyramid in step S1 is as follows: the image data is read into the memory at the server end at one time, and a spatial index is established to perform hierarchical resampling and slice output to establish an image slice pyramid.

[0040] Furthermore, the specific process of step S4 is as follows:

[0041] The identified building change spots are visually superimposed on the original image data in the form of an electronic map on the browser side. Through the image rolling comparison view or quick screenshot switch view mode, the accuracy of each suspected illegal building spot is verified, and confirmation or filtering operations are performed. The filtered data is updated to the original image data in real time.

[0042] On the other hand, the illegal building identification monitoring and detection system based on deep learning includes:

[0043] Image upload unit, used to obtain image data through satellite or drone and upload the image data to the server, automatically completing the pre-processing of image data during uploading, including resolution resampling, geometric correction, and image slice pyramid construction;

[0044] The illegal building comparison unit is used to select two phases of image data to be compared, and output building change patches including building outline changes and height changes through comparison detection on the server side;

[0045] The data automatic storage unit is used to automatically import the extracted building change spots into the results library;

[0046] The result checking unit is used to further check the correctness of the extracted building change spots by superimposing the extracted building change spots on the corresponding online images, and automatically save the final results.

[0047] The beneficial effects of the present invention are as follows: the present invention realizes the rapid extraction of illegal building spots through deep learning algorithms in a targeted manner, and forms a highly integrated automation solution from image data uploading, building spots and illegal building spots extraction, to image and illegal building spots storage, image pyramid slicing, and spatial data service sharing; the solution further provides an online sample delineation function for buildings, and continuously improves the recognition accuracy of illegal buildings by deep learning models through online building sample delineation, illegal building result verification and other means. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flow chart of an illegal building identification, monitoring and detection method based on deep learning provided by an embodiment of the present invention;

[0049] Figure 2 This is a diagram of the selection interface for comparing images;

[0050] Figure 3 It is a building map that is used to outline and display the interface diagram;

[0051] Figure 4 is a schematic diagram of a building recognition model provided by an embodiment of the present invention;

[0052] Figure 5 It is a schematic diagram of adding directions to the building recognition model provided by an embodiment of the present invention;

[0053] Figure 6 Schematic diagram of the Tensor Polygonization vectorization process provided by an embodiment of the present invention;

[0054] Figure 7 This is the interface displaying the illegal construction list. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0056] In order to illustrate the technical solution of the present invention, a specific embodiment is provided below for illustration.

[0057] Figure 1 The process of the illegal building identification, monitoring and detection method based on deep learning provided by an embodiment of the present invention is shown. For the sake of convenience, only the parts related to the embodiment of the present invention are shown.

[0058] like Figure 1 As shown, the illegal building identification monitoring and detection method based on deep learning provided by this embodiment includes the following steps:

[0059] Step S1, image uploading step: image data is acquired through satellite or drone and uploaded to the server, and image data preprocessing is automatically completed during uploading, including resolution resampling, geometric correction, and image slice pyramid construction.

[0060] The present invention uses satellites or drones to obtain image data to check for illegal buildings. The collected image data will be automatically uploaded to the server. During the upload, the image data preprocessing is automatically completed, such as resolution resampling, geometric correction, and image slice pyramid construction, and the image metadata information is synchronously stored in the database.

[0061] Through automatic geometric correction, the spatial offset between two images in the same range can be further reduced. After selecting two images to be compared in the same area, the present invention automatically searches for the same-name points in the two images and performs automatic geometric correction on one of the images.

[0062] In addition, in order to improve the image slicing performance, the smoothness of the slice edge can be optimized. Traditional image slicing tools need to take into account issues such as the smoothness of the interface operation, and cannot fully utilize the computer's memory. In the embodiment of the present invention, the server does not need to consider the smoothness of the interface, and reads all the images into the memory at one time. By establishing a spatial index, it efficiently performs hierarchical resampling and slice output. Its biggest bottleneck is the IO output. Through comparative analysis, the slicing efficiency can be greatly improved. At the same time, when traditional slices are output at a high resolution, the mosaic phenomenon on the edge of the slice is very obvious. In order to make the slice edge softer, in the embodiment of the present invention, convolution resampling at the boundary is used to make the slice boundary smoother. The uploaded images can be previewed on the data line and queried, deleted, and modified.

[0063] Step S2, illegal building comparison step: select two phases of image data to be compared, perform comparison detection on the server, and output a building change map including building outline changes and height changes.

[0064] This embodiment compares the two phases of impact data of the same building to determine whether the building has changed in outline and height, and finally outputs the building change map. Figure 2 shown.

[0065] The impact data obtained here include orthophotos and oblique photography data. Orthophotos are remote sensing images with orthophoto properties. The original remote sensing images are affected by the changes in the internal state of the sensor, the external state and the surface conditions during imaging, and have different degrees of distortion and loss of stability. The geometric processing of remote sensing images can not only extract spatial information, but also resample the image grayscale according to the correct geometric relationship to form a new orthophoto.

[0066] For buildings, the original images taken by remote sensing are geometrically processed to obtain orthophotos of the buildings. The specific processing process is not the focus of this embodiment and will not be described in detail here.

[0067] The specific process of this step is as follows:

[0068] S21. For the same area, two phases of orthophotos are obtained, and raster patches of building outline changes are obtained through building extraction and building outline change detection, and further converted into vector patches.

[0069] This step obtains the building outline data through contour change detection. Finally, the building outline change vector map can be obtained through the orthophoto building outline data of two periods, including:

[0070] S211, train a deep learning-based building recognition model by collecting sample data;

[0071] The sample data collection includes collecting the required high-resolution remote sensing data, recording the shooting time and height of the high-resolution remote sensing data, preprocessing the data, resampling to the required resolution and importing it; then, based on the existing building and road data, training a building recognition model based on encoder-decoder deep learning, which is used for building extraction. The sample data image to be outlined is used as the electronic map base map, and the building patches in the image are outlined in the form of a vectorization toolbox to realize online vectorization operations. The vectorized building patches will be saved in the sample library for online training. The building patches are outlined and displayed as follows: Figure 3 shown.

[0072] like Figure 4 As shown in the figure, the building recognition model is mainly composed of an encoder network, a decoder network and a pixel-wise classification layer.

[0073] The encoding network converts high-dimensional vectors into low-dimensional vectors, thus realizing low-dimensional extraction of high-dimensional features. Although the encoding network can capture more translation invariant features through multiple maximum pooling operations, it will also lose more important segmentation basis such as boundary information of feature maps. Therefore, during the pooling process, the maximum pooling index information is recorded at the same time to save the location of the maximum eigenvalue, and then the input feature map is upsampled using the maximum pooling index information so that the boundary information can be preserved. The maximum pooling index can further optimize the accuracy of the boundary contour description, reduce the number of parameters, and can be trained end-to-end, further improving the accuracy of recognition.

[0074] The decoding network uses the maximum pooling index information of the corresponding feature layer saved when the encoder is down-sampled to map the low-resolution feature map to the high-spatial-resolution feature map, thus realizing the reconstruction of the low-dimensional vector to the high-dimensional vector. Reusing the maximum pooling index in the decoding process has several advantages: optimizing the description of the boundary contour; reducing the number of parameters and enabling end-to-end training; and the upsampling method can be applied to any encoding-decoding network.

[0075] The encoding network and the decoding network are connected via a band attention connection, so that the model can read the original high-dimensional data again during decoding, ensuring that the geometric features remain unchanged.

[0076] In the last layer, the decoder outputs a high-dimensional feature representation vector as the input of the trainable classifier. The encoding network can use the classic Resnet, Inception or Efficientnet.

[0077] In the specific implementation, this building recognition model is implemented based on pytorch, and is trained using a server with 3 GPU GeForce RTX2080Ti. The training batch size is 64 (batch_size), the learning rate is 1e-04, AdamW is used to optimize the parameters, and the cross entropy loss is used as the loss model to constrain the model training. The training process is as follows:

[0078] 1. The training data size is 512*512, and it is obtained from the original image using a sliding window method. The number of iteration rows is 256, ensuring that most buildings exist independently in one image. 60% of the total data is selected as training data and 40% as verification data.

[0079] 2. Random image enhancement is used in training to increase the generalization performance of the model and avoid overfitting. Images are randomly enhanced using image enhancement methods including inversion, scaling, random brightness, and contrast.

[0080] 3. The model training uses a fusion loss function, which includes two types of cross entropy and dice coefficient. The cross entropy is used to ensure the model's accurate recognition of building categories, and the dice coefficient is used to constrain the building outline.

[0081] 4. The model is trained for 400 cycles. The initial learning rate is 1e-04 and starts to decrease to 1e-07 after 10 cycles. The model with the highest accuracy in the test data is finally selected as the result model.

[0082] S212, obtaining two phases of orthophotos of the same area, using the building recognition model to predict the building map in the orthophotos of the two phases, and obtaining the building map.

[0083] To predict whether the outline of a building has changed, it is necessary to compare the orthophotos of the building in two phases. Then directly input them into the building recognition model trained in the previous step to obtain the building patch.

[0084] S213, using Tensor Polygonization to generate a simplified building outline vector from the building patch, and obtaining a building outline change vector patch according to the changes of the building outline vectors in two phases.

[0085] Deep learning image segmentation usually outputs segmentation results in raster format, while the application of geographic information systems often requires vector output. In order to optimize the process of raster image vectorization, this step uses Tensor Polygonization for vectorization. Figure 5 As shown in Figure 1, we first added a direction output to the building recognition model to make it consistent with the ground truth contour. This is used to predict the direction of each current target and extract a smooth vector effect. The vectorization process is shown in Figure 1. Figure 6 As shown, the process is as follows:

[0086] 1. Generate a building thermal map through the building recognition model to obtain the building vector;

[0087] 2. According to the principle of building shape characteristics, smooth the building vector and eliminate the jagged edges in the generated vector;

[0088] 3. According to the direction generated in the building recognition model, identify the building corners and mark the corner positions;

[0089] 4. Delete the vectors between corners and optimize the position of generating building outline vectors.

[0090] The simplified vector generation method (Tensor Polygonization) in this step proposes a vectorization method based on the building direction based on the characteristics that the building outlines are mostly right-angled or equiangular polygons, thereby greatly optimizing the accuracy and data volume of generating building vectors from image heat maps.

[0091] S22, for the area, obtain the oblique photography data corresponding to the two periods of orthophotos, and obtain the building height change vector map by building height change detection.

[0092] This step specifically includes:

[0093] S221, obtaining oblique photography data corresponding to the two periods of orthophotos, merging the triangulated networks of the oblique photography data, and forming a single large-scale oblique photography data.

[0094] Oblique photography data is relative to orthophotos. To determine whether a building is illegal, this step requires obtaining oblique photography data that is concurrent with the two orthophotos of the building. The two orthophotos of period A and period B are orthophoto A and orthophoto B, respectively, and the oblique photography data A and oblique photography data B obtained in the same period are also obtained.

[0095] Since oblique photography data is composed of multiple separate osgb data, when a building pattern spans multiple osgb data, there is a loss of efficiency in data reading and retrieval, so this embodiment merges the osgb data into a large osgb data to improve the data reading efficiency during retrieval. Oblique photography osgb data is generally a continuous triangular mesh. The merging of multiple osgb data requires merging the vertices and indexes in each data to ensure that the merged osgb data has only one triangular mesh and one texture.

[0096] This step reorganizes the fragmented original OSGB oblique photography data into a large-scale OSGB data. When reading the data at a high level, the number of traversals of the original data is reduced, and the data throughput is reduced.

[0097] S222. Use the building spot contour data to superimpose on the merged oblique photography data, that is, the orthophoto A is superimposed on the oblique photography data A, and the orthophoto B is superimposed on the oblique photography data B. The oblique photography data has high-level information of all pixels. The highest high-level information within the oblique photography data range is obtained as the high-level value. The high-level value is used as an attribute field and written into the current building spot contour data. The high-level information of the building can be obtained through high-level extraction.

[0098] S223. After acquiring the building outline data with high-rise values ​​in two periods, after matching according to the range coordinates, perform high-rise value comparison to obtain the height change vector map of the building whose high-rise has changed.

[0099] In the process of extracting the high-level of the image spots, since the original oblique photography data are stored in fragments, the preprocessing of the original oblique photography data is very important. In this step, the high-level information reading efficiency of the oblique photography data is further improved by reorganizing the oblique photography data.

[0100] S23, merging the building outline change vector patch and the height change vector patch to obtain an overall building change pattern patch.

[0101] The overall building change pattern represents the changes in the building's outline and height. As long as any change in the outline or height is found through the building change pattern, it can be preliminarily determined that the building is illegally constructed, which is a "suspected pattern".

[0102] Step S3, automatic data storage step: automatically import the extracted building change spots into the results library.

[0103] Step S4, result verification step: by superimposing the extracted building change spots onto the corresponding online images, further verify the correctness of the extracted spot data, and automatically save the final result.

[0104] The identified building change spots are visualized and superimposed on the original image data in the form of an electronic map on the browser side. The accuracy of each suspected illegal building spot is verified by rolling-up comparison of the image or quick screenshot switching viewing mode, and confirmation or filtering operations are performed. The filtered data is updated to the original image data in real time. During the verification of this step, the two images can be superimposed and displayed on the browser side in the form of a rolling-up screen, or superimposed by screenshot switching viewing mode, and all illegal building spots are superimposed on the image. The platform can control the visibility of the spots. For the building change spots identified in step S2, a list of illegal building spots is formed.

[0105] like Figure 7 In the interface shown, the backend staff selects the map to be checked, the map will be synchronously zoomed to the location of the map, and then the map is selected with the mouse to perform the check and verification operation in the pop-up box.

[0106] The present invention uses an algorithm to identify illegal buildings, and can quickly and accurately determine whether a building is an illegal building. The whole process is simple to implement. It only needs to train a building recognition model, and simultaneously obtain orthophotos and oblique photography data of the same period. Through data processing and detection and judgment, it can quickly identify the outline of the building and whether the height has changed. This method can significantly improve the recognition accuracy and efficiency. Then, after identifying the architectural change pattern of the illegal building, further verification and confirmation are carried out.

[0107] The present invention provides an online visual operation method for checking various "suspected spots" of illegal buildings, which is more intuitive and quick to assist users. At the visual end of illegal building checking, two operation modes, rolling shutter comparison and scaled image comparison, are provided, allowing users to quickly determine whether the illegal building spots are correct in combination with the image base map, and supports mouse and keyboard operations to further improve work efficiency.

[0108] In addition, this embodiment also provides an illegal building identification monitoring and detection system based on deep learning, including:

[0109] Image upload unit, used to obtain image data through satellite or drone and upload the image data to the server, automatically completing the pre-processing of image data during uploading, including resolution resampling, geometric correction, and image slice pyramid construction;

[0110] The illegal building comparison unit is used to select two phases of image data to be compared, and output building change patches including building outline changes and height changes through comparison detection on the server side;

[0111] The data automatic storage unit is used to automatically import the extracted building change spots into the results library;

[0112] The result checking unit is used to further check the correctness of the extracted building change spots by superimposing the extracted building change spots on the corresponding online images, and automatically save the final results.

[0113] Each of the above functional units correspondingly implements the above steps S1-S4, and the specific implementation process is not repeated here.

[0114] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying, monitoring and detecting illegal buildings based on deep learning, characterized in that: The method comprises the following steps: Step S1, image uploading step: acquiring image data through satellite or drone and uploading the image data to the server, automatically completing image data preprocessing during uploading, including resolution resampling, geometric correction, and image slice pyramid construction; Step S2, illegal building comparison step: selecting two periods of image data to be compared, and outputting a building change map including building outline changes and height changes through comparison detection at the server end; Step S3, automatic data storage step: automatically importing the extracted building change spots into the achievement database; Step S4, result checking step: further checking the correctness of the extracted building change spots by superimposing the extracted building change spots on the corresponding online image, and automatically saving the final result; The acquired data includes orthophotos and oblique photography data, and the specific process of step S2 is as follows: S21. For the same area, two phases of orthophotos are obtained, and raster patches of building outline changes are obtained by building extraction and building outline change detection, and further converted into vector patches; S22, for the area, obtaining oblique photography data of the period corresponding to the two periods of orthophotos, and obtaining a building height change vector map patch by building height change detection; S23, fusing the building outline change vector spot with the height change vector spot to obtain an overall building change spot; Wherein step S21 specifically includes the following steps: The deep learning-based building recognition model is trained by collecting sample data, wherein the sample data collection includes collecting the required high-resolution remote sensing data, recording the shooting time and height of the high-resolution remote sensing data, preprocessing the data, and resampling to the required resolution; The sample data image to be outlined is used as the electronic map base map, and the building spots in the image are outlined through the vectorization toolbox. The vectorization toolbox provides operation rollback, data preservation, data modification, and spot statistics functions to realize online vectorization operations. The vectorized building spots will be saved in the sample library for online call training; Obtain two phases of orthophotos of the same area, use the building recognition model to predict the building images in the two phases of orthophotos, and obtain the building patches; The building patch is used to generate a simplified building outline vector using Tensor Polygonization, and the building outline change vector patch is obtained based on the changes in the building outline vectors in two phases.

2. The method for identifying, monitoring and detecting illegal buildings based on deep learning as claimed in claim 1, characterized in that: The process of generating a simplified building outline vector is as follows: Generate a building thermal map through a building recognition model to obtain a building vector; According to the principle of building shape characteristics, the building vector is smoothed to eliminate the jagged edges in the generated vector; According to the directions generated in the building recognition model, the building corners are identified and the corner positions are marked; Delete the vectors between corners and optimize the location of generated building outline vectors.

3. The method for identifying, monitoring and detecting illegal buildings based on deep learning as claimed in claim 1, characterized in that: Step S22 specifically includes the following steps: Obtain the oblique photography data corresponding to the two periods of orthophotos, merge the triangulated networks of the oblique photography data, and form a single large-scale oblique photography data; Use the building spot outline data to be superimposed on the merged oblique photography data, wherein the oblique photography data has high-level information of all pixels, obtain the highest high-level information within the oblique photography data range as the high-level value, and write the high-level value as an attribute field into the current building spot outline data; After obtaining the building outline data with high-rise values ​​in two periods, the high-rise values ​​are compared after matching according to the range coordinates to obtain the height change vector map of the building where the high-rise has changed.

4. The method for identifying, monitoring and detecting illegal buildings based on deep learning as claimed in claim 1, characterized in that: The process of constructing the image slice pyramid in step S1 is as follows: the image data is read into the memory at the server end at one time, and a spatial index is established to perform hierarchical resampling and slice output to establish an image slice pyramid.

5. The method for identifying, monitoring and detecting illegal buildings based on deep learning as claimed in claim 1, characterized in that: The specific process of step S4 is as follows: The identified building change spots are visually superimposed on the original image data in the form of an electronic map on the browser side. Through the image rolling comparison view or quick screenshot switch view mode, the accuracy of each suspected illegal building spot is verified, and confirmation or filtering operations are performed. The filtered data is updated to the original image data in real time.

Images