Automatic identification method and equipment of insulator strings based on deep learning satellite remote sensing images

Abstract

The embodiment of the present invention discloses a method and device for automatically identifying insulator strings based on deep learning satellite remote sensing images. The method comprises: preprocessing remote sensing image data, super-resolution network training and testing to obtain four-fold super-resolution remote sensing images; post-processing the super-resolution remote sensing images, dividing and augmenting training sets and test sets to obtain a first training set and a first test set; building a pole tower network, using the first training set as input information for training; obtaining a pole tower data set, dividing and augmenting training sets and test sets to obtain a second training set and a second test set; building an insulator string network, using the second training set as input information for training; cascading the trained super-resolution network, the pole tower network and the insulator string network in sequence to form a cascade network, inputting the first test set and the second test set into the cascade network, obtaining output test images and geographic information of the pole tower and the insulator string, thereby realizing large-area inspection of the insulator string.

Description

Technical Field

[0001] The present invention relates to the fields of remote sensing image super-resolution, target recognition and semantic segmentation, and in particular to a method and device for automatically identifying insulator strings based on deep learning satellite remote sensing images. Background Art

[0002] Insulator strings are very important insulation control components in transmission lines. They mainly support transmission wires and prevent leakage tripping caused by electricity flowing back to the ground. In high-voltage transmission lines, a considerable number of insulators are required to form insulator strings. Once an insulator explodes, the transmission line will have hidden dangers in operation safety. In severe cases, it will reduce the operation cycle of the transmission line and even cause power outages, resulting in large-scale power outages and huge property losses. Therefore, it is particularly important to conduct large-scale inspections of insulator strings.

[0003] The existing technology usually adopts manual inspection or drone inspection. In manual inspection, the inspectors need to conduct safety inspections on each insulator string in each transmission line in turn, which is not only time-consuming and labor-intensive, but also has a lot of safety risks for the current ultra-high voltage and large-capacity transmission lines. In drone inspection, the work efficiency and personnel safety of drone inspection have been greatly improved compared with manual inspection. However, drones are expensive, the area that can be photographed at one time is limited, and they cannot conduct large-area inspections of insulator strings. Summary of the invention

[0004] Based on this, it is necessary to propose an automatic identification method and equipment for insulator strings based on deep learning satellite remote sensing images to address the above problems, so as to realize large-area inspection of insulator strings. In the long run, the inspection cost is not only low, but the safety of personnel is also guaranteed.

[0005] To achieve the above objectives, the present application provides, in a first aspect, a method for automatically identifying insulator strings based on deep learning satellite remote sensing images, the method comprising:

[0006] Preprocess the remote sensing image data by Gram-Schmidt fusion of multispectral image and panchromatic band image to obtain the preprocessed fused remote sensing image;

[0007] Performing super-resolution network training on the fused remote sensing image to obtain a trained super-resolution network, and testing to obtain a quadruple super-resolution remote sensing image;

[0008] Performing remote sensing image post-processing on the super-resolution remote sensing image to obtain a post-processed remote sensing image;

[0009] Dividing the post-processed remote sensing image into a training set and a test set, and performing augmentation processing to obtain a first training set and a first test set;

[0010] Building a pole tower target recognition network, and training the pole tower target recognition network using the first training set as input information to obtain a trained pole tower target recognition network;

[0011] Acquire a tower data set, divide the tower data set into a training set and a test set, and perform augmentation processing on the tower data set to obtain a second training set and a second test set;

[0012] Building an insulator string semantic segmentation and recognition network, and using the second training set as input information to train the insulator string semantic segmentation and recognition network to obtain a trained insulator string semantic segmentation and recognition network;

[0013] The trained super-resolution network, the pole tower target recognition network and the insulator string semantic segmentation recognition network are cascaded in sequence to form a cascade network, and the first test set and the second test set are input into the cascade network to obtain output test images and geographic information of the pole tower and insulator string.

[0014] Optionally, the remote sensing image data is remote sensing image data collected by high-resolution satellites, and the high-resolution satellites include Worldview-1 satellite and / or Worldview-3 satellite.

[0015] Optionally, the preprocessing of the remote sensing image data by Gram-Schmidt fusion of multispectral image and panchromatic band image to obtain the preprocessed fused remote sensing image includes:

[0016] Performing RPC geometric correction processing on the multispectral image and the panchromatic band image of the remote sensing image data to obtain a corrected image;

[0017] Performing Gram-Schmidt remote sensing image fusion processing on the corrected image to obtain the remote sensing image fusion;

[0018] The remote sensing image fusion is subjected to data compression processing in the range of 0-255 to obtain the fused remote sensing image.

[0019] Optionally, the super-resolution network is a super-resolution deep learning model.

[0020] Optionally, performing remote sensing image post-processing on the super-resolution remote sensing image to obtain a post-processed remote sensing image includes:

[0021] Performing false color output processing on the super-resolution remote sensing image to obtain a false color output image;

[0022] Grayscale stretching processing is performed on the false color output image to obtain the post-processed remote sensing image.

[0023] Optionally, the false color output replaces the green band with a near infrared band, and the grayscale stretching adopts a 1% linear grayscale stretching.

[0024] Optionally, the augmentation process includes: random cropping, rotation, and noise addition.

[0025] Optionally, the tower target recognition network is a tower target recognition network deep learning model.

[0026] Optionally, the insulator string semantic segmentation and recognition network is an insulator string semantic segmentation and recognition network deep learning model.

[0027] In a second aspect, the present application provides an automatic identification device for insulator strings based on deep learning satellite remote sensing images, the device comprising:

[0028] The preprocessing module is used to preprocess the remote sensing image data by Gram-Schmidt fusion of multispectral image and panchromatic band image to obtain the preprocessed fused remote sensing image;

[0029] A super-resolution network module is used to perform super-resolution network training on the fused remote sensing image to obtain a trained super-resolution network and test and obtain a four-fold super-resolution remote sensing image;

[0030] A post-processing module, used for performing remote sensing image post-processing on the super-resolution remote sensing image to obtain a post-processed remote sensing image;

[0031] A first division processing module is used to divide the post-processed remote sensing image into a training set and a test set, and perform augmentation processing to obtain a first training set and a first test set;

[0032] A tower recognition network module is used to build a tower target recognition network, and train the tower target recognition network using the first training set as input information to obtain a trained tower target recognition network;

[0033] A second partitioning processing module is used to obtain a tower data set, divide the tower data set into a training set and a test set, and perform augmentation processing to obtain a second training set and a second test set;

[0034] An insulator string recognition network module is used to build an insulator string semantic segmentation recognition network, and train the insulator string semantic segmentation recognition network using the second training set as input information to obtain a trained insulator string semantic segmentation recognition network;

[0035] The image output module is used to cascade the trained super-resolution network, the tower target recognition network and the insulator string semantic segmentation recognition network in sequence to form a cascade network, input the first test set and the second test set into the cascade network, and obtain the output test images and geographic information of the tower and insulator string.

[0036] The embodiment of the present invention has the following beneficial effects: a method and device for automatically identifying insulator strings based on deep learning satellite remote sensing images, the method comprising: preprocessing remote sensing image data by Gram-Schmidt fusion of multispectral images and panchromatic band images to obtain a preprocessed fused remote sensing image; performing super-resolution network training on the fused remote sensing image to obtain a trained super-resolution network, and testing to obtain a four-fold super-resolution remote sensing image; performing remote sensing image post-processing on the super-resolution remote sensing image to obtain a post-processed remote sensing image; dividing the post-processed remote sensing image into a training set and a test set, and performing augmentation processing to obtain a first training set and a first test set; building a pole tower target recognition network, using the first training set as input information to train the pole tower target recognition network, and obtaining a trained pole tower target recognition network; obtaining a pole tower The data set is divided into a training set and a test set for the pole tower data set, and augmented to obtain a second training set and a second test set; an insulator string semantic segmentation and recognition network is built, and the second training set is used as input information to train the insulator string semantic segmentation and recognition network to obtain the trained insulator string semantic segmentation and recognition network; the trained super-resolution network, the pole tower target recognition network and the insulator string semantic segmentation and recognition network are cascaded in sequence to form a cascade network, and the first test set and the second test set are input into the cascade network to obtain the output test images and geographic information of the pole tower and insulator string, and the insulator string recognition is output by inputting the satellite remote sensing image data into the test image output, thereby achieving large-area inspection, and in the long run, without the need for drones and manual inspections, the inspection cost is not only low, but also the safety of personnel is guaranteed. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0038] in:

[0039] Figure 1 This is an execution flow chart of the method for automatically identifying insulator strings based on deep learning satellite remote sensing images in an embodiment of the present application;

[0040] Figure 2 This is a schematic diagram of the structure of an insulator string automatic identification device based on deep learning satellite remote sensing images in an embodiment of the present application;

[0041] Figure 3 This is a schematic diagram of the structure of the control components of the insulator string automatic identification device based on deep learning satellite remote sensing images in an embodiment of the present application. DETAILED DESCRIPTION

[0042] 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 only 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 are within the scope of protection of the present invention.

[0043] The embodiments of the present application will first describe a method for automatically identifying insulator strings based on deep learning satellite remote sensing images, and then describe a device for automatically identifying insulator strings based on deep learning satellite remote sensing images.

[0044] See also Figure 1 , is an execution flow chart of the method for automatic identification of insulator strings based on deep learning satellite remote sensing images in an embodiment of the present application, and the execution steps are as follows:

[0045] Step 110: Preprocess the remote sensing image data to obtain a fused remote sensing image.

[0046] The remote sensing image data in step 110 are remote sensing image data collected by high-resolution satellites, and the high-resolution satellites include the Worldview-1 satellite and / or the Worldview-3 satellite.

[0047] In a feasible implementation method, the remote sensing image processing software ENVI can be used to pre-process the remote sensing image. Specifically: the multispectral image and panchromatic band image of the remote sensing image data are subjected to RPC geometric correction processing to obtain the corrected image; the corrected image is subjected to Gram-Schmidt remote sensing image fusion processing to obtain remote sensing image fusion; and the remote sensing image fusion is subjected to data compression processing in the range of 0-255 to obtain a fused remote sensing image.

[0048] Step 120: Perform super-resolution network training on the fused remote sensing image to obtain a trained super-resolution network, and test to obtain a four-fold super-resolution remote sensing image.

[0049] A WDSR (Wide Deep Super Resolution, WDSR) super-resolution network can be built first. The super-resolution network can be a super-resolution deep learning model. In a feasible implementation, the super-resolution deep learning model includes: four input channels, twenty-three ResidentBlocks, a PixelShuffer layer, and a four-fold magnified PixelShuffer layer. The fused remote sensing image is trained with a super-resolution deep learning model to obtain a trained super-resolution deep learning model. The trained super-resolution deep learning model is used for testing. The data used for the test can be a fused remote sensing image. The fused remote sensing image is input into the trained super-resolution deep learning model to obtain a super-resolution remote sensing image.

[0050] The parameters for super-resolution deep learning model training are set as follows: the initial learning rate (LR) is 0.0002, the learning state (LS) attenuates LR at a ratio of 0.5 in the 25000th, 50000th and 75000th single cycles, the batch size (BS) is 16, the optimization algorithm uses the Adam algorithm, the loss function uses the L1 norm loss function, and the total number of training cycles is 100.

[0051] Step 130: performing remote sensing image post-processing on the super-resolution remote sensing image to obtain a post-processed remote sensing image.

[0052] Among them, after obtaining the four-fold super-resolution remote sensing image, the super-resolution remote sensing image can be post-processed, and the post-processing includes false color output and grayscale stretching. Specifically, the super-resolution remote sensing image can be first subjected to false color output processing to obtain a false color output image; then the false color output image is subjected to grayscale stretching processing to obtain a post-processed remote sensing image; wherein, the false color output is to replace the green band with the near-infrared band, and the grayscale stretching can adopt a 1% linear grayscale stretching.

[0053] It can be understood that: false color output replaces the green band with the near-infrared band, so that the super-resolution image is output in the red band, near-infrared band, and blue band to improve the recognition of the tower; grayscale stretching uses 1% linear grayscale stretching to improve image brightness.

[0054] Step 140: Divide the post-processed remote sensing image into a training set and a test set, and perform augmentation processing to obtain a first training set and a first test set.

[0055] The augmentation processing in step 140 includes: random cropping, rotation and noise addition.

[0056] It is understandable that through the augmentation processing methods of random cropping, rotation and noise addition, the training samples can be balanced and the amount of data can be increased.

[0057] Step 150: construct a pole tower target recognition network, and train the pole tower target recognition network using the first training set as input information to obtain a trained pole tower target recognition network.

[0058] Among them, the tower target recognition network can be a tower target recognition network deep learning model, and the tower target recognition network deep learning model includes: three input channels, a feature extraction structure based on ResNet101, and top-up feature fusion, and after feature extraction, it is input into the RPN (Region Proposal Network) network and the Faster RCNN network for feature learning training of the tower target recognition network deep learning model, and the first training set is used as input information to train the tower target recognition network deep learning model, so as to obtain the trained tower target recognition network deep learning model.

[0059] Among them, the parameter setting parameters of the deep learning model of the tower target recognition network are: the number of iterations is 10000, the initial learning rate (LR) is 0.001, the attenuation is performed at 5000 and 7000 iterations, the attenuation factors are 0.1 and 0.001, the optimization algorithm uses the Adam algorithm, and the loss function uses the Smooth L1 loss function.

[0060] Step 160: Acquire a pole tower data set, divide the pole tower data set into a training set and a test set, and perform augmentation processing to obtain a second training set and a second test set.

[0061] In an embodiment of the present application, a pole tower data set can be obtained by manual screening, and the pole tower data set is divided into a training set and a test set, and the divided training set and test set are augmented based on GAN to obtain a second training set and a second test set.

[0062] Among them, the augmentation processing methods include: random cropping, rotation and noise addition.

[0063] It is understandable that through the augmentation processing methods of random cropping, rotation and noise addition, the training samples can be balanced and the amount of data can be increased.

[0064] In a feasible implementation, the tower data set in which the insulator strings can be clearly seen can be expanded to 2,500 images.

[0065] Step 170: construct an insulator string semantic segmentation and recognition network, and use the second training set as input information to train the insulator string semantic segmentation and recognition network to obtain a trained insulator string semantic segmentation and recognition network.

[0066] Among them, the insulator string semantic segmentation and recognition network can be an insulator string semantic segmentation and recognition network deep learning model, and the insulator string semantic segmentation and recognition network deep learning model includes: three input channels, based on the HR-Net (High-Resolution Networks, HR-Net) feature learning framework, building forty-eight layers of parallel connected hidden feature learning layers, and using the second training set as input information to train the insulator string semantic segmentation and recognition network deep learning model, thereby obtaining the trained insulator string semantic segmentation and recognition network deep learning model.

[0067] Among them, the parameters of the deep learning model of the insulator string semantic segmentation and recognition network are set as follows: the iteration round Epoch is 100 rounds, the SGD optimizer is used, the initial learning rate (Learning Rate, LR) is 0.001, the learning rate decay factor is 0.0001, each Epoch uses the Cityscapes category balancing method, and the OHEM (Online Hard Example Mining, OHEM) training method is used during the training process, and the OHEM threshold is 0.9.

[0068] Step 180: The trained super-resolution network, the tower target recognition network and the insulator string semantic segmentation recognition network are sequentially cascaded to form a cascade network, and the first test set and the second test set are input into the cascade network to obtain output test images and geographic information of the towers and insulator strings.

[0069] The cascade network is a cascade network into which geographic information has been introduced.

[0070] It is understandable that geographic information includes coordinates, tower numbers, and so on.

[0071] It should be noted that in the embodiment of the present application, the tower target recognition network deep learning model can adopt the FPN target recognition model, and the insulator string semantic segmentation recognition network deep learning model can adopt the HR-Net semantic segmentation recognition network model.

[0072] The cascade network of step 180, in a feasible implementation, may be formed by sequentially cascading the super-resolution network deep learning model trained in step 120, the pole tower target recognition network deep learning model trained in step 150, and the insulator string semantic segmentation recognition network deep learning model trained in step 170 to form a cascade network. By inputting the first test set and the second test set into the cascade network, the output test images and geographic information of the pole tower and insulator string are obtained, and the output test images and geographic information of the pole tower and insulator string are tested, and the test results obtained are as follows:

[0073] Among them, the tower target recognition network deep learning model of the cascade network adopts the FPN target recognition model, and the test results are: Precision is 0.916, Recall is 0.884, and F1-Score is 0.8997; the insulator string semantic segmentation recognition network deep learning model of the cascade network adopts the HR-Net semantic segmentation recognition network model, and the test results are: Precision is 0.800, Recall is 0.928, and F1-Score is 0.8593.

[0074] Among them, Precision is the accuracy, which represents the false detection rate of model recognition, and the value range is [0, 1]. The higher the value, the lower the false detection rate; Recall is the recall rate, which represents the missed detection rate of model recognition, and the value range is [0, 1]. The higher the value, the lower the missed detection rate; F1-Score is a balance parameter of Precision and Recall, and represents both the false detection rate and the missed detection rate of the model.

[0075] It can be seen from the test results that the FPN target recognition model maintains a low false positive rate and a low missed positive rate when identifying pole towers, and can well achieve the purpose of identifying most pole towers; the HR-Net semantic segmentation recognition network model also maintains a low false positive rate and a low missed positive rate when identifying insulator strings, and can well achieve the purpose of identifying most insulator strings.

[0076] In the embodiment of the present application, by identifying insulator strings from satellite remote sensing image data input to test image output, large-area inspections are achieved. In the long run, there is no need for drone inspections and manual inspections, so not only is the inspection cost low, but the safety of personnel is also guaranteed.

[0077] It is understandable that compared with manual inspections and drone inspections, inspection methods based on satellite remote sensing images and deep learning models are faster, safer, and have a wider prediction range. They can identify insulator strings on a large scale for transmission lines under the needs of power grid applications.

[0078] See also Figure 2 , is a schematic diagram of the structure of an insulator string automatic identification device based on deep learning satellite remote sensing images in an embodiment of the present application, the device includes:

[0079] The preprocessing module 210 is used to preprocess the remote sensing image data by Gram-Schmidt fusion of multispectral image and panchromatic band image to obtain a preprocessed fused remote sensing image;

[0080] The super-resolution network module 220 is used to perform remote sensing image post-processing on the super-resolution remote sensing image to obtain a post-processed remote sensing image;

[0081] A post-processing module 230 is used to perform remote sensing image post-processing on the super-resolution remote sensing image to obtain a post-processed remote sensing image;

[0082] A first division processing module 240 is used for post-processing the remote sensing image to divide the training set and the test set, and perform augmentation processing to obtain a first training set and a first test set;

[0083] The tower recognition network module 250 is used to build a tower target recognition network, and train the tower target recognition network using the first training set as input information to obtain a trained tower target recognition network;

[0084] A second partitioning processing module 260 is used to obtain a tower data set, partition the tower data set into a training set and a test set, and perform augmentation processing to obtain a second training set and a second test set;

[0085] An insulator string recognition network module 270 is used to build an insulator string semantic segmentation recognition network, and train the insulator string semantic segmentation recognition network using the second training set as input information to obtain a trained insulator string semantic segmentation recognition network;

[0086] The image output module 280 is used to cascade the trained super-resolution network, the tower target recognition network and the insulator string semantic segmentation recognition network in sequence to form a cascade network, input the first test set and the second test set into the cascade network, and obtain the output test images and geographic information of the tower and insulator string.

[0087] In the embodiment of the present application, the relevant contents of the above-mentioned pre-processing module 210, super-resolution network module 220, post-processing module 230, first division processing module 240, tower identification network module 250, second division processing module 260, insulator string identification network module 270 and image output module 280 can be referred to. Figure 1 The contents in the illustrated embodiment will not be described in detail here.

[0088] See also Figure 3, which is a structural diagram of the control components of the insulator string automatic identification device based on deep learning satellite remote sensing images in an embodiment of the present application, and the device includes a processor, a memory and a network interface connected through a system bus.

[0089] The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the isolation ward stores an operating system and may also store a computer program, which, when executed by a processor, enables the processor to automatically identify an insulator string. The internal memory may also store a computer program, which, when executed by a processor, enables the processor to automatically identify an insulator string.

[0090] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a part of the structure related to the scheme of the present application, and does not constitute a limitation on the automatic identification device of the insulator string to which the scheme of the present application is applied. The specific automatic identification device of the insulator string may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0091] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing related hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods.

[0092] Among them, any reference to memory, storage, database or other media used in the embodiments provided in the present application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0093] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0094] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for automatic identification of insulator strings based on deep learning satellite remote sensing images, characterized in that: The method comprises: Preprocess the remote sensing image data by Gram-Schmidt fusion of multispectral image and panchromatic band image to obtain the preprocessed fused remote sensing image; Performing super-resolution network training on the fused remote sensing image to obtain a trained super-resolution network, and testing to obtain a quadruple super-resolution remote sensing image; Performing remote sensing image post-processing on the super-resolution remote sensing image to obtain a post-processed remote sensing image; Dividing the post-processed remote sensing image into a training set and a test set, and performing augmentation processing to obtain a first training set and a first test set; Building a pole tower target recognition network, and training the pole tower target recognition network using the first training set as input information to obtain a trained pole tower target recognition network; Acquire a tower data set, divide the tower data set into a training set and a test set, and perform augmentation processing on the tower data set to obtain a second training set and a second test set; Building an insulator string semantic segmentation and recognition network, and using the second training set as input information to train the insulator string semantic segmentation and recognition network to obtain a trained insulator string semantic segmentation and recognition network; The trained super-resolution network, the tower target recognition network and the insulator string semantic segmentation recognition network are sequentially cascaded to form a cascade network, and the first test set and the second test set are input into the cascade network to obtain output test images and geographic information of the tower and the insulator string; The step of performing remote sensing image post-processing on the super-resolution remote sensing image to obtain a post-processed remote sensing image includes: Performing false color output processing on the super-resolution remote sensing image to obtain a false color output image; Grayscale stretching processing is performed on the false color output image to obtain the post-processed remote sensing image.

2. The method according to claim 1, characterized in that The remote sensing image data is remote sensing image data collected by high-resolution satellites, and the high-resolution satellites include Worldview-1 satellite and / or Worldview-3 satellite.

3. The method according to claim 1, characterized in that The preprocessing of the remote sensing image data by Gram-Schmidt fusion of multispectral image and panchromatic band image to obtain the preprocessed fused remote sensing image includes: Performing RPC geometric correction processing on the multispectral image and the panchromatic band image of the remote sensing image data to obtain a corrected image; Performing Gram-Schmidt remote sensing image fusion processing on the corrected image to obtain the remote sensing image fusion; The remote sensing image fusion is subjected to data compression processing in the range of 0-255 to obtain the fused remote sensing image.

4. The method according to claim 1, characterized in that: The super-resolution network is a super-resolution deep learning model.

5. The method according to claim 1, characterized in that The false color output is to replace the green band with the near infrared band, and the grayscale stretching is to adopt a 1% linear grayscale stretching.

6. The method according to claim 1, characterized in that The augmentation process includes: random cropping, rotation, and noise addition.

7. The method according to claim 1, characterized in that The tower target recognition network is a tower target recognition network deep learning model.

8. The method according to claim 1, characterized in that The insulator string semantic segmentation and recognition network is an insulator string semantic segmentation and recognition network deep learning model.

9. An automatic identification device for insulator strings based on deep learning satellite remote sensing images, characterized in that: The device comprises: The preprocessing module is used to preprocess the remote sensing image data by Gram-Schmidt fusion of multispectral image and panchromatic band image to obtain the preprocessed fused remote sensing image; A super-resolution network module is used to perform super-resolution network training on the fused remote sensing image to obtain a trained super-resolution network and test and obtain a four-fold super-resolution remote sensing image; A post-processing module, used for performing remote sensing image post-processing on the super-resolution remote sensing image to obtain a post-processed remote sensing image; A first division processing module is used to divide the post-processed remote sensing image into a training set and a test set, and perform augmentation processing to obtain a first training set and a first test set; A tower recognition network module is used to build a tower target recognition network, and train the tower target recognition network using the first training set as input information to obtain a trained tower target recognition network; A second partitioning processing module is used to obtain a tower data set, divide the tower data set into a training set and a test set, and perform augmentation processing to obtain a second training set and a second test set; An insulator string recognition network module is used to build an insulator string semantic segmentation recognition network, and train the insulator string semantic segmentation recognition network using the second training set as input information to obtain a trained insulator string semantic segmentation recognition network; An image output module, used to sequentially cascade the trained super-resolution network, the tower target recognition network and the insulator string semantic segmentation recognition network to form a cascade network, input the first test set and the second test set into the cascade network, and obtain output test images and geographic information of the tower and insulator string; The step of performing remote sensing image post-processing on the super-resolution remote sensing image to obtain a post-processed remote sensing image includes: Performing false color output processing on the super-resolution remote sensing image to obtain a false color output image; Grayscale stretching processing is performed on the false color output image to obtain the post-processed remote sensing image.

Images