DebuurGAN network-based high-speed unmanned aerial vehicle image motion blur removal identification method and system

Abstract

The invention discloses an identification method and system for removing motion blur of a high-speed unmanned aerial vehicle image based on a DeblurGAN network, and the method comprises the steps: collecting a motion clear image and a motion blur image of a high-speed unmanned aerial vehicle, carrying out the manual simulation of the motion clear image, and obtaining a simulated motion blur image; training the preprocessed DeblurGAN network model by using the motion clear image and the simulated motion blurred image to obtain a motion blurred removal network, and inputting the motion blurred image into the motion blurred removal network to obtain a final motion blurred removal image; and training and optimizing the edge convolutional neural network model by using an adaptive edge extraction weight technology to obtain an optimized edge convolutional neural network model, and inputting the final image subjected to motion blur removal into the model to obtain a classification result of the unmanned aerial vehicle. The method effectively solves the problem of motion blur in the high-speed unmanned aerial vehicle image, and improves the recognition and tracking capabilities of the unmanned aerial vehicle.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing and machine learning, and in particular to a method and system for identifying motion blur removal of high-speed unmanned aerial vehicle images based on a DeblurGAN network. Background Art

[0002] With the rapid development of drone technology, the use of LFS-UAVs (Low Fast and Small Unmanned Aerial Vehicles) has become more and more common and its application scope has become wider and wider. However, the widespread use of drones has also brought a series of safety risks and management challenges. The improper use of drones may pose a threat to public safety and may even mistakenly enter "no-fly" areas, collide with birds, airplanes and other aerial objects, cause serious safety accidents, disrupt air traffic control, and endanger air traffic safety.

[0003] At present, the identification and tracking of drones mainly rely on technical means such as radar and radio signal detection. However, these means have certain limitations in practical applications. For example, radar equipment is expensive and has blind spots when detecting at close range; radio signal detection may be affected by environmental interference, affecting recognition accuracy. Therefore, drone recognition technology based on video images has become an important research direction, but in the case of high-speed motion, drone images are easily affected by motion blur, resulting in a decrease in image quality, which in turn affects the identification and tracking of drones. Traditional deblurring methods often have poor results when dealing with high-speed motion blur and are difficult to meet the needs of practical applications. Summary of the invention

[0004] The purpose of the present invention is to provide a high-speed UAV image de-motion blur recognition method and system based on the DeblurGAN network, which can effectively solve the motion blur problem in high-speed UAV images and improve the recognition and tracking capabilities of UAVs by using advanced image processing and machine learning techniques.

[0005] The present invention adopts the following technical solution:

[0006] A method for identifying motion blur of high-speed UAV images based on a DeblurGAN network comprises the following steps:

[0007] S1. Collect motion clear images and motion blurred images of a high-speed drone, artificially simulate the motion clear images based on the motion blurred images, and obtain simulated motion blurred images. The motion clear images and the simulated motion blurred images constitute a data set of motion images.

[0008] S2. Preprocess the DeblurGAN network model using the motion clear image, train the preprocessed DeblurGAN network model using the data set in step S1 to obtain a de-motion blur network, perform background ablation and normalization on the motion blurred image in step S1 and input it into the de-motion blur network to obtain the final de-motion blurred image, and evaluate the image using Yolo5s.

[0009] S3. Use deep learning and computer image processing technology to build an edge convolutional neural network model; based on the final de-motion blurred image, use adaptive edge extraction weight technology to train and optimize the model to obtain an optimized edge convolutional neural network model.

[0010] S4. The final de-motion blurred image is input into the optimized edge convolutional neural network model. After feature extraction, feature fusion and classification processing, the classification result of the drone is obtained to realize the recognition of the drone.

[0011] Furthermore, in step S1, the data set constituting the motion blurred image includes the following contents:

[0012] S101, collecting clear motion images and motion blurred images of high-speed drones under different backgrounds, lighting conditions, and weather conditions, and performing random rotation, scaling, and cropping transformations on the clear motion images.

[0013] S102, based on image processing technology, using a computer script to randomly select the motion clear image transformed in step S101 as the foreground image, and select the corresponding blur kernel size according to the height of the foreground image. The specific selection process is as follows:

[0014] When the height pixel of the foreground image is less than or equal to 15, the blur kernel size is randomly selected between 2 and 4; when the height pixel of the foreground image is greater than 15 and less than or equal to 25, the blur kernel size is randomly selected between 2 and 8; when the height pixel of the foreground image is greater than 25 and less than or equal to 150, the blur kernel size is randomly selected between 5 and 20; when the height pixel of the foreground image is greater than 150, the blur kernel size is randomly selected between 5 and 50.

[0015] S103 , randomly overlaying the image corresponding to the blur kernel onto the background image in a vertical, horizontal, or vertical and horizontal mixed manner to obtain a simulated motion blurred image.

[0016] Furthermore, in step S2, the image after de-motion blurring includes the following contents:

[0017] S201, performing background ablation on the motion clear image in step S1 to obtain a foreground, normalizing the foreground and the original background and inputting them into the DeblurGAN network model input port to obtain a preprocessed DeblurGAN network model.

[0018] S202, select a set number of images in the data set of step S1 as a training set, input the training set into the preprocessed DeblurGAN network model, extract the feature map of the image at the conv3*3 layer through the VGG-19 model, and calculate the difference between the feature map of the simulated motion blurred image and the motion clear image. Based on the difference, use content loss and adversarial loss to iteratively optimize the generator until the set maximum number of iterations is reached, and then obtain the optimized generator. The optimized generator and discriminator constitute a de-motion blur network.

[0019] S203, performing background ablation on the collected motion blurred image to obtain a foreground moving object, a moving target and a background, normalizing the moving target and the background, processing the processed motion blurred image using a de-motion blur network to obtain a de-motion blurred image, performing brightness, contrast and color balance processing on the image to obtain a final de-motion blurred image.

[0020] Furthermore, in step S3, constructing the optimized edge convolutional neural network model includes the following contents:

[0021] S301, the edge convolutional neural network model includes an input layer, 3 convolutional layers, 2 fully connected layers and an output layer connected in sequence, wherein the design shape of the input layer is 64*64*4.

[0022] S302. Based on image processing technology, two different scale levels of edge extraction are used to obtain the weight of the final de-motion blurred image, and a 3x3 Gaussian kernel is used for smoothing, and the Canny function in the Opencv open source library is used for edge detection; a 7x7 Gaussian kernel is used for smoothing, and the Canny function is used for edge detection; the results of the two edge detections are weighted and summed.

[0023] S303, calling the non-maximum suppression and double threshold method in the opencv open source library to process the edge detection result after weighted summation to obtain the final edge abstract information.

[0024] S304, using an edge detection method to perform edge detection on the drone in the final de-motion blurred image, and using the final edge abstract information to train and optimize the edge convolutional neural network model to obtain an optimized edge convolutional neural network model.

[0025] Furthermore, in step S4, the classification result of the drone includes the following contents:

[0026] The final de-motion blurred image is divided into a training set and a test set in a ratio of 4:1. The training set is input into the optimized edge convolutional neural network model for training iteration until the maximum number of iterations is reached, and the trained model is obtained. The test set is resized, converted into Tensor, and standardized, and the processed test set and the final edge abstraction information are input into the trained model, where the processed test set passes through the first three channels of the model input layer, and the final edge abstraction information passes through the fourth channel of the model input layer. When passing through the first convolution layer, 19 19*19 filters are used for convolution operation to extract primary features. The first feature map is convolved with 22 15*15 filters when the first feature map passes through the second convolution layer, and 22 second feature maps are output; when the second feature map passes through the third convolution layer, 8 9*9 filters are used for convolution operation, and 8 third feature maps are output; the third feature map is nonlinearly transformed by using the ReLU activation function, and then input into the first fully connected layer for linear transformation, and a set number of nodes are output; the node is linearly transformed when passing through the second fully connected layer, and the final node is output; stochastic gradient descent is used as the optimizer, the learning rate and momentum parameters are set, and the loss and network back propagation are calculated to obtain the classification result of the drone.

[0027] Furthermore, the present invention also proposes a high-speed UAV image de-motion blur recognition system based on the DeblurGAN network, comprising:

[0028] The data set construction module is used to collect motion clear images and motion blurred images of high-speed UAVs, artificially simulate motion clear images based on motion blurred images, and obtain simulated motion blurred images. The motion clear images and simulated motion blurred images constitute a data set of motion images.

[0029] The image deblurring processing module is used to preprocess the DeblurGAN network model using the motion clear image, train the preprocessed DeblurGAN network model using the data set in step S1, obtain the de-motion blur network, perform background ablation and normalization processing on the motion blurred image in step S1, and then input it into the de-motion blur network to obtain the final de-motion blurred image, and evaluate the image using Yolo5s.

[0030] The model optimization module is used to construct an edge convolutional neural network model using deep learning and computer image processing technology; based on the final de-motion blurred image, the model is trained and optimized using adaptive edge extraction weight technology to obtain an optimized edge convolutional neural network model.

[0031] The recognition module is used to input the final de-motion blurred image into the optimized edge convolutional neural network model. After feature extraction, feature fusion and classification processing, the classification result of the drone is obtained to realize the recognition of the drone.

[0032] Furthermore, the present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the method for identifying motion blur of high-speed drone images based on the DeblurGAN network are implemented.

[0033] Furthermore, the present invention also proposes a computer-readable storage medium, which stores a computer program. When the computer program is run by a processor, it executes the recognition method for de-motion blur of high-speed drone images based on the DeblurGAN network.

[0034] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0035] 1. The present invention simplifies the image processing process and improves processing efficiency and accuracy.

[0036] 2. The present invention can effectively handle different types of blur, including local blur and global blur, and has better adaptability and robustness.

[0037] 3. Aiming at the problem of smear blur caused by high-speed UAVs in flight, the present invention proposes a method of first deblurring and then classifying and identifying. This deblurring method can greatly improve the recognition rate of UAV images.

[0038] 4. The present invention adds edge information to the drone detection and classification network for fusion. Edge information plays an important guiding role in the details and structure of the image, and can help the network better capture and restore the key features in the image, thereby improving the accuracy and reliability of recognition.

[0039] 5. The present invention is not only applicable to the recognition of high-speed UAVs, but can also be extended to the recognition of blurred images of other high-speed moving objects, and has wide applicability and flexibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is an overall implementation flow chart of the present invention.

[0041] Figure 2 is an optical image sample collected in an embodiment of the present invention.

[0042] Figure 3 It is the DeblurGAN network model after preprocessing of the present invention.

[0043] Figure 4 It is a recognition effect diagram of a clear image, a blurred image and a restored image in an optical image in an embodiment of the present invention.

[0044] Figure 5 is an edge detection result diagram of an embodiment of the present invention.

[0045] Figure 6 It is a diagram of the image input process of the input layer in the edge convolutional neural network model optimized by the present invention.

[0046] Figure 7 3 is a diagram showing the restoration effect of an optical image in an embodiment of the present invention.

[0047] Figure 8 It is a comparison chart of drone images and classification accuracy after deblurring by different networks in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0049] To achieve the above objectives, the present invention proposes a high-speed UAV image deblurring recognition method based on the DeblurGAN network. Figure 1 As shown, the specific steps are as follows:

[0050] S1. Collecting clear motion images of high-speed UAVs (Flying Drones) flying at low speeds, defocused lenses, and motion blurred images caused by excessive speed of high-speed UAVs (Flying Drones) in optical images; artificially simulating clear motion images based on motion blurred images of high-speed UAVs to obtain simulated motion blurred images, and the clear motion images and simulated motion blurred images constitute a data set of motion images; the specific contents are as follows:

[0051] S101. Collect clear motion images and motion blurred images of high-speed drones under different backgrounds, lighting conditions, and weather conditions, and perform random rotation, scaling, cropping and other image transformations on the clear motion images to simulate shooting effects at different angles and distances for data enhancement.

[0052] S102, based on image processing technology, using a computer script to randomly select the motion clear image transformed in step S101 as the foreground image, and select a suitable blur kernel size according to the height of the foreground image. The specific selection process is as follows:

[0053] When the height pixel of the foreground image is less than or equal to 15, the blur kernel size is randomly selected between 2 and 4; when the height pixel of the foreground image is greater than 15 and less than 25, the blur kernel size is randomly selected between 2 and 8; when the height pixel of the foreground image is greater than 25 and less than 150, the blur kernel size is randomly selected between 5 and 20; when the height pixel of the foreground image is greater than 150, the blur kernel size is randomly selected between 5 and 50.

[0054] S103, randomly overlaying the image corresponding to the blur kernel onto the background image in a vertical, horizontal or mixed manner (horizontal and vertical superposition), so as to achieve artificial synthesis of motion blur effects of different degrees and obtain a simulated motion blurred image.

[0055] In this embodiment, the data set of motion images includes 500 sets of motion clear images and simulated motion blurred images. Figure 2 As shown, Figure 2 (a) is the optical image without adding blur kernel in the dataset. Figure 2 (b) is Figure 2 (a) corresponds to the optical image with blur kernel added, Figure 2 (c) and Figure 2 (d) are all drones randomly captured from the original images.

[0056] S2. Preprocess the DeblurGAN network model using the motion clear image, train the preprocessed DeblurGAN network model using the data set in step S1, obtain a de-motion blur network, perform background ablation and normalization on the motion blurred image collected in step S1, and input it into the de-motion blur network to obtain the final de-motion blurred image, and use Yolo5s to evaluate the image; the specific contents are:

[0057] The present invention uses the DeblurGAN network model as the basis of the image deblurring network. Specifically, by accessing the [KupynOrest] / [DeblurGAN] repository on the GitHub platform, download the source code of the DeblurGAN network model, which is based on the deblurring algorithm proposed by Goroshin et al. in 2018. The present invention makes fine adjustments based on the network model. Specifically, Figure 3As shown, since the network will perform downsampling operations on large-resolution images, in order to retain the blurred details of the drone, the background of the motion clear image in step S1 is ablated to obtain the foreground, and the foreground and the original background are normalized and input into the DeblurGAN network model input port to avoid the problem of detail loss caused by downsampling, and the preprocessed DeblurGAN network model is obtained.

[0058] 400 groups of images of the data set in step S1 are selected as training sets, and the training sets are input into the preprocessed DeblurGAN network model. The feature map of the image at the conv3*3 layer is extracted by the VGG-19 model, and the difference between the feature maps of the simulated motion blurred image and the motion clear image is calculated. Based on the difference, the generator is iteratively optimized using content loss and adversarial loss until 200 rounds are reached to obtain the optimized generator. The optimized generator and the discriminator constitute a de-motion blur network; background ablation is performed on the collected motion blurred image to obtain foreground moving objects, moving targets and backgrounds, the moving targets and backgrounds are normalized, and the processed motion blurred images are processed by the de-motion blur network to obtain de-motion blurred images. The image is processed for brightness, contrast and color balance to further enhance the visual effect of the image to obtain the final de-motion blurred image, which has clearer edges and details than the input blurred image.

[0059] Yolov5s was used to perform drone target detection on 100 groups of simulated motion blurred images, clear images and the final de-motion blurred images, and the ratio of the number of correctly identified targets to the total number of targets was counted. The experiment proved that the DeblurGAN network model can significantly improve the target recognition rate of blurred images. Specifically: the recognition rate of blurred images is between 50%-75%, with an average of 65%; the recognition rate of clear images is between 90%-95%, with an average of 92%; the recognition rate of restored images after deblurring is between 85%-95%, with an average of 88%. In the deblurring process of high-speed drones, the recognition rate of blurred images is significantly improved, with an average increase of 23%, and the recognition rate of restored images is close to that of clear images, with a difference of only 4%. Figure 4 As shown, Figure 4 (a) is the original image test result of the drone. Figure 4 (b) is the test result of the optical image of the drone to be restored. Figure 4 (c) is the image test result of the optical image restored by the trained network. It can be seen that after the blurred image is restored by the network, the drone recognition rate will be greatly improved and even higher than the original image.

[0060] S3. Use deep learning and computer image processing technology to build an edge convolutional neural network model; based on the final de-motion blurred image, use adaptive edge extraction weight technology to train and optimize the model to obtain an optimized edge convolutional neural network model; the specific content is:

[0061] S301. The present invention specifically designs a convolutional neural network for drone classification based on the Pytorch neural network library. When building a network model, first import the necessary data processing and model building library; then build a convolutional neural network model according to the defined model structure based on the API or function provided in the Pytorch library. The edge convolutional neural network model includes an input layer, 3 convolutional layers, 2 fully connected layers and an output layer connected in sequence, wherein the design shape of the input layer is 64*64*4.

[0062] S302, such as Figure 5 As shown in the figure, based on image processing technology, two different scale levels of edges are used to extract the weights of the final de-motion blurred image, and a smaller Gaussian kernel (3x3) is used for smoothing. The Canny function edge detection in the Opencv open source library is used to capture detailed features (propellers, wings, etc.). Figure 5 (a) shows that small-scale edge extraction can effectively obtain detailed features such as propellers. A larger Gaussian kernel (7x7) is then used for smoothing, followed by Canny edge detection to capture the main contours. Figure 5 (b) shows that large-scale edge extraction can clearly extract the main outline of the drone. Finally, the small-scale and large-scale edge results are weighted and summed.

[0063] S303, calling the non-maximum suppression and double threshold method in the opencv open source library to improve the accuracy and reliability of the edge result after weighted summation, and obtain the final edge abstract information, such as Figure 5 as shown in (c).

[0064] S304, using an edge detection method to perform edge detection on the drone in the final de-motion blurred image, and using the edge detection results to train and optimize the edge convolutional neural network model to obtain an optimized edge convolutional neural network model.

[0065] S4. Input the final de-motion blurred image into the optimized edge convolutional neural network model. After feature extraction, feature fusion and classification processing, the classification result of the drone is obtained to realize the recognition of the drone. The specific content is:

[0066] The final de-motion blurred image is divided into a training set and a test set in a ratio of 4:1. The training set is input into the optimized edge convolutional neural network model, and the number of training iterations is set to 100 and the sample size is set to 64. The forward propagation and loss function use the same parameters as in step S301. After 100 rounds of iterations, the test set is processed by a predefined transformation, including resizing (64*64), converting to Tensor, and normalizing. The processed test set and the final edge abstraction information are input into the trained optimized edge convolutional neural network model, as shown in Figure 6 As shown in the figure, the processed test set passes through the first three channels of the model input layer, and the final edge abstraction information passes through the fourth channel of the model input layer; when passing through the first convolution layer, 19 19*19 filters are used for convolution operation to extract primary features and output 19 46*46 first feature maps; when the first feature map passes through the second convolution layer, 22 15*15 filters are used for convolution operation to extract more complex features and output 22 32*32 second feature maps; when the second feature map passes through the third convolution layer, 8 9*9 filters are used for convolution operation The product operation is performed to extract higher-level features and output 8 24*24 third feature maps. The third feature map is nonlinearly transformed using the ReLU activation function and then input into the first fully connected layer, where it is flattened into 4608 nodes, linearly transformed, and 80 nodes are output. When the node passes through the second fully connected layer, it is linearly transformed and 3 nodes are output. Stochastic gradient descent is used as the optimizer, and the learning rate is set to 0.01 and the momentum parameter is set to 0.5. The loss and network backpropagation are calculated to obtain the classification results of drones, which are drones, flying birds, and other backgrounds (negative samples).

[0067] like Figure 7 As shown, Figure 7 (a) is the original image of the drone. Figure 7 (b) is the optical image of the drone to be restored. Figure 7 (c) shows the restoration effect of the trained network on the optical image. The shape of the drone is restored more completely and the smear is eliminated to a certain extent.

[0068] like Figure 8 Shown is the comparison data of the accuracy of the optimized edge convolutional neural network model and the convolutional neural network without edge information for classifying deblurred drone images and the convolutional neural network for blurry drone images. The experimental results show that the classification results output by the optimized edge convolutional neural network model on the data set produced by the present invention can achieve an accuracy of 90% or more, and the classification accuracy of drones is improved by about 2 to 3 percentage points compared with ordinary convolutional neural networks.

[0069] The embodiment of the present invention also proposes a high-speed UAV image deblurring recognition system based on the DeblurGAN network, including a data set construction module, an image deblurring processing module, a model optimization module, a recognition module, and a computer program that can be run on a processor. It should be noted that each module in the above system corresponds to the specific steps of the method provided in the embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method. For technical details not described in detail in this embodiment, please refer to the method provided in the embodiment of the present invention.

[0070] The embodiment of the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. It should be noted that when the processor executes the computer program, the specific steps of the method provided in the embodiment of the present invention are corresponding to the specific steps of the method provided in the embodiment of the present invention, and the processor has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided in the embodiment of the present invention.

[0071] The embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program. It should be noted that when the computer program is executed by the processor, it corresponds to the specific steps of the method provided in the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided in the embodiment of the present invention.

[0072] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for deblurring high-speed UAV images based on the DeblurGAN network, characterized in that: include: S1. Collect motion clear images and motion blurred images of a high-speed UAV, artificially simulate the motion clear images based on the motion blurred images, and obtain simulated motion blurred images. The motion clear images and the simulated motion blurred images constitute a data set of motion images. S2, preprocessing the DeblurGAN network model using the motion clear image, training the preprocessed DeblurGAN network model using the data set in step S1, obtaining a de-motion blur network, performing background ablation and normalization processing on the motion blurred image in step S1, and inputting the de-motion blur network to obtain the final de-motion blurred image, and evaluating the image using Yolo5s; S3. Use deep learning and computer image processing technology to build an edge convolutional neural network model; based on the final de-motion blurred image, use adaptive edge extraction weight technology to train and optimize the model to obtain an optimized edge convolutional neural network model; S4. The final de-motion blurred image is input into the optimized edge convolutional neural network model. After feature extraction, feature fusion and classification processing, the classification result of the drone is obtained to realize the recognition of the drone.

2. The method for deblurring high-speed UAV images based on the DeblurGAN network according to claim 1 is characterized in that: In step S1, the data set constituting the motion blurred image includes the following contents: S101, collecting clear motion images and motion blurred images of high-speed drones under different backgrounds, lighting conditions, and weather conditions, and performing random rotation, scaling, and cropping transformations on the clear motion images; S102, based on image processing technology, using a computer script to randomly select the motion clear image transformed in step S101 as the foreground image, and select the corresponding blur kernel size according to the height of the foreground image. The specific selection process is as follows: When the height pixel of the foreground image is less than or equal to 15, the blur kernel size is randomly selected between 2 and 4; when the height pixel of the foreground image is greater than 15 and less than or equal to 25, the blur kernel size is randomly selected between 2 and 8; when the height pixel of the foreground image is greater than 25 and less than or equal to 150, the blur kernel size is randomly selected between 5 and 20; when the height pixel of the foreground image is greater than 150, the blur kernel size is randomly selected between 5 and 50; S103 , randomly overlaying the image corresponding to the blur kernel onto the background image in a vertical, horizontal, or vertical and horizontal mixed manner to obtain a simulated motion blurred image.

3. The method for deblurring high-speed UAV images based on the DeblurGAN network according to claim 1 is characterized in that: In step S2, the image after de-motion blurring is obtained includes the following contents: S201, performing background ablation on the motion clear image in step S1 to obtain a foreground, normalizing the foreground and the original background and inputting them into the DeblurGAN network model input port to obtain a preprocessed DeblurGAN network model; S202, select a set number of images in the data set of step S1 as a training set, input the training set into the preprocessed DeblurGAN network model, extract the feature map of the image at the conv3*3 layer through the VGG-19 model, and calculate the difference between the feature map of the simulated motion blurred image and the motion clear image, based on the difference, use the content loss and adversarial loss to iteratively optimize the generator until the set maximum number of iterations is reached, and the optimized generator is obtained. The optimized generator and the discriminator constitute a de-motion blur network; S203, performing background ablation on the collected motion blurred image to obtain a foreground moving object, a moving target and a background, normalizing the moving target and the background, processing the processed motion blurred image using a de-motion blur network to obtain a de-motion blurred image, performing brightness, contrast and color balance processing on the image to obtain a final de-motion blurred image.

4. The method for deblurring high-speed UAV images based on the DeblurGAN network according to claim 1 is characterized in that: In step S3, constructing the optimized edge convolutional neural network model includes the following contents: S301, the edge convolutional neural network model includes an input layer, three convolutional layers, two fully connected layers and an output layer connected in sequence, wherein the design shape of the input layer is 64*64*4; S302, based on image processing technology, using two different scale levels of edge extraction to obtain the weight of the final de-motion blurred image, and using a 3x3 Gaussian kernel for smoothing, and using the Canny function in the Opencv open source library for edge detection; then using a 7x7 Gaussian kernel for smoothing, and using the Canny function for edge detection; and weighted summing the results of the two edge detections; S303, calling the non-maximum suppression and double threshold method in the opencv open source library to process the edge detection result after weighted summation to obtain the final edge abstract information; S304, using an edge detection method to perform edge detection on the drone in the final de-motion blurred image, and using the final edge abstract information to train and optimize the edge convolutional neural network model to obtain an optimized edge convolutional neural network model.

5. The method for deblurring high-speed UAV images based on the DeblurGAN network according to claim 4 is characterized in that: In step S4, the classification results of the drones include the following: The final de-motion blurred image is divided into a training set and a test set in a ratio of 4:

1. The training set is input into the optimized edge convolutional neural network model for training iteration until the maximum number of iterations is reached to obtain the trained model. The test set is resized, converted into Tensor and standardized, and the processed test set and the final edge abstraction information are input into the trained model, wherein the processed test set passes through the first three channels of the input layer of the model, and the final edge abstraction information passes through the fourth channel of the input layer of the model. When passing through the first convolution layer, Use 19 19*19 filters to perform convolution operations, extract primary features, and output 19 first feature maps; when the first feature map passes through the second convolution layer, use 22 15*15 filters to perform convolution operations, and output 22 second feature maps; when the second feature map passes through the third convolution layer, use 8 9*9 filters to perform convolution operations, and output 8 third feature maps; use the ReLU activation function to perform nonlinear transformation on the third feature map, input it into the first fully connected layer for linear transformation, and output a set number of nodes; when the node passes through the second fully connected layer, it undergoes linear transformation and outputs the final node; Use stochastic gradient descent as the optimizer, set the learning rate and momentum parameters, calculate the loss and network backpropagation, and obtain the classification results of the drone.

6. A system for the high-speed UAV image deblurring recognition method based on the DeblurGAN network according to claim 1, characterized in that: include: A data set construction module is used to collect motion clear images and motion blurred images of a high-speed UAV, artificially simulate the motion clear images based on the motion blurred images, and obtain simulated motion blurred images. The motion clear images and the simulated motion blurred images constitute a data set of motion images; An image deblurring processing module is used to preprocess the DeblurGAN network model using a motion clear image, train the preprocessed DeblurGAN network model using the data set in step S1, obtain a de-motion blurring network, perform background ablation and normalization processing on the motion blurred image in step S1, and then input the processed image into the de-motion blurring network to obtain a final de-motion blurred image, and evaluate the image using Yolo5s; The model optimization module is used to construct an edge convolutional neural network model using deep learning and computer image processing technology; based on the final de-motion blurred image, the model is trained and optimized using the adaptive edge extraction weight technology to obtain an optimized edge convolutional neural network model; The recognition module is used to input the final de-motion blurred image into the optimized edge convolutional neural network model. After feature extraction, feature fusion and classification processing, the classification result of the drone is obtained to realize the recognition of the drone.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method for identifying motion blur removal of high-speed drone images based on the DeblurGAN network as described in any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the recognition method for de-motion blurring of high-speed UAV images based on the DeblurGAN network described in any one of claims 1 to 5 is executed.

Images