An aircraft recognition and detection system based on deep learning

Abstract

The present invention discloses an aircraft recognition and detection system based on deep learning, which includes an image processing module, a dataset construction module, a data acquisition module, an image acquisition module, a deep learning dataset module, a startup module, a target detection module, a power supply module, an identification algorithm module, a network connection module, an image preprocessing module, an image segmentation module, an image feature extraction module, an image model recognition module, a manual annotation label module, and an image enhancement module. The present invention first restores and processes the pre-classified image information, and then enhances and corrects the pixel points to increase the clarity of the image. The identification algorithm module improves the transmission efficiency of feature information through the design and optimization of the network structure model, conducts initial training using remote sensing images with different resolutions, and conducts training by randomly changing the image scale, realizing the automatic acquisition, interpretation, and result processing of images, and improving the test quality and efficiency of the aircraft system.

Description

Technical Field

[0001] The present invention belongs to the technical field of aviation testing, and particularly relates to an aircraft recognition and detection system based on deep learning. Background Art

[0002] With the rapid development of science and technology and its wide application in the field of aviation technology, aircraft have increasingly become an important means of transportation in human life and an important manifestation of national defense strength. Especially in the current complex international political and economic environment, whoever owns advanced aviation products has the right to speak in international politics, economy, trade, etc.

[0003] In recent years, the development of artificial intelligence technology has led to the promotion of actively using machines to replace repetitive physical labor in all walks of life. Among them, images are a form of information transmission and expression in people's life and work, and the recognition of targets in images has become an important research field of artificial intelligence. Image recognition is a process of processing, analyzing, and then understanding images, which usually requires extracting and matching features of objects in the images to determine the types of objects in the images. As an important function of modern aircraft, the aircraft recognition system is mainly to ensure the automatic and accurate recognition of various ground and air targets at long distances, large ranges, and high precision during the flight of the aircraft, and to provide flight and strike reconnaissance information for air, ground, and other users. In order to test and verify the functions and reliability of the aircraft recognition system, traditional aircraft image recognition relies on manual operation, the testing process is cumbersome and the accuracy is not high, it is impossible to automatically obtain, interpret, and process the results. At the same time, the aircraft image processing method is single, and the image clarity is not high enough, which affects the quality of subsequent image processing and analysis. Summary of the Invention

[0004] To solve the above-mentioned problems, a plane recognition and detection system based on deep learning is provided. In the present invention, the image restoration processing module, image enhancement module, image correction module, and image model noise reduction module inside the image processing module cooperate with each other to perform pre-restoration processing on the pre-classified image information, thereby improving the clarity of the image. Subsequently, through the enhancement and correction of pixel points, the effect of increasing the image clarity is achieved, and at the same time, the clarity and quality after image processing are ensured; the recognition algorithm module designs and optimizes the network structure model to ensure that the accuracy rate will not decrease, improves the transmission efficiency of feature information, and improves the training strategy during the training process. Remote sensing images with different resolutions are used for the initial training, and the image scale is randomly changed for training to improve the generalization ability and robustness of the model. The test results show that the optimized model has a significantly reduced model size, the detection speed is twice the original, the efficient transmission of feature information ensures the detection accuracy, improves the test quality of the plane, and at the same time realizes the automatic acquisition, interpretation, and result processing of images during the plane test, shortening the plane production cycle.

[0005] The technical solution adopted by this application to solve its technical problems is as follows:

[0006] A plane recognition and detection system based on deep learning, including an image processing module, a dataset construction module, a data acquisition module, an image acquisition module, a deep learning dataset module, a startup module, a target detection module, a power supply module, a recognition algorithm module, a network connection module, an image preprocessing module, an image segmentation module, an image feature extraction module, an image model recognition module, a manual annotation label module, and an image enhancement module. The output end of the startup module is connected to the input end of the deep learning dataset module, the output end of the deep learning dataset module is connected to the input end of the image acquisition module, the output end of the image acquisition module is connected to the input end of the data acquisition module, the output end of the data acquisition module is connected to the input end of the image processing module, the power supply module and the network connection module are fixedly installed outside the image processing module, the output end of the image processing module is connected to the input end of the dataset construction module, the output end of the dataset construction module is connected to the input end of the target detection module, and the output end of the target detection module is connected to the input end of the recognition algorithm module. Through the processing and analysis of the above modules, the recognition and detection of the plane target in the image are finally realized. The specific steps are as follows:

[0007] S1: After the startup module is started, the power supply module provides electrical energy for the entire system;

[0008] S2: The deep learning dataset module collects the picture information from the global image screenshots of Google Earth, the NWPU VHR-10 dataset, the DOTA dataset, and the RSOD dataset to form an image database;

[0009] S3: The image acquisition module collects the image data of the nature and state of the ground objects in real time through the on-board data interface;

[0010] S4: The image acquisition module transmits the image data collected according to the steps of S2, and finally inputs the data into the port of the image preprocessing module;

[0011] S5: The image preprocessing module first performs grayscale processing on the image, and then uses the method of smoothing filtering to denoise the image;

[0012] S6: The image segmentation module segments the image. The input image of this segmentation method is a grayscale image. After threshold operation, a binary image is output, and the contour of the segmented image is further extracted;

[0013] S7: The image feature extraction module uses the gray-level co-occurrence matrix to extract four parameters of energy, contrast, entropy, and correlation, as well as the shape features and invariant moment features of the image;

[0014] S8: The image model recognition module uses the SSD algorithm and combines the Inception V2 convolutional network operation as the target image recognition model of the aircraft;

[0015] S9: The network connection module provides network connection support for the entire system;

[0016] S10: After cleaning and enhancing the data, the manual annotation label module uniformly names the samples in the PASCAL VOC format and manually labels the images;

[0017] S11: The image enhancement module sets a test set to evaluate the effect of the model;

[0018] S12: The target detection module modifies some configuration options of the framework. For example, it removes the data enhancement method in the framework. In this target detection framework, the input size of the image is 300×300, and the non-linear activation function after the convolutional layer uses ReLU6 to achieve image target detection;

[0019] S13: After passing through the first convolutional layer Conv1, the size of the image is 120×120×64. The convolution uses the Same convolution. The principle of the second convolutional layer Conv2 is the same as that of Conv1, and the number of channels generated after the convolution operation remains unchanged. Next is a pooling layer Pool_1, which uses max pooling operation. After Pool_1, there are two consecutive convolutional layers Conv3 and Conv4, with the same principle as the previous convolutional layers, except that the number of channels is changed to 128. Then comes the second pooling layer Pool_2, with the same operation principle as the previous pooling layer. After Pool_2, there are two convolutional layers Conv5 and Conv6, with the number of channels set to 256. Then a pooling layer Pool_3 is added. After Pool_3, three convolutional layers are stacked, namely Conv7, Conv8, and Conv9, and the number of channels of these three convolutional layers is set to 512. After the three convolutional layers, a pooling layer Pool_4 is added, and then three convolutional layers Conv10, Conv11, and Conv12 are stacked, with their number of channels set to 1024, and then a pooling layer Pool_5 is added. After Pool_5, two fully connected layers are added to predict the category of the image. The number of neurons in the first fully connected layer is set to 2048, and at the same time, the Dropout principle is applied and its value is set to 0.5. The number of neurons in the second fully connected layer is set to the number of types of the image, and the Dropout value is set to 0.8. Finally, the Softmax function is used to output the predicted image type.

[0020] The image processing module is internally provided with an image preprocessing module, an image segmentation module, an image feature extraction module, and an image model recognition module. The output ends of the image preprocessing module, the image segmentation module, the image feature extraction module, and the image model recognition module are connected to the input end of the image processing module.

[0021] The dataset construction module is internally provided with a manual annotation label module and an image enhancement module. The output ends of the manual annotation label module and the image enhancement module are connected to the input end of the dataset construction module.

[0022] The deep learning dataset module is the UCAS - AOD dataset, and the UCAS - AOD dataset is an aerial remote sensing image target detection dataset.

[0023] Internally, the image preprocessing module first performs grayscale processing. The evCvColor function is selected for grayscale operation on the image. After that, the image preprocessing module will perform denoising internally, and the image preprocessing module uses the method of smoothing filtering for denoising.

[0024] The image segmentation module uses the code cvThreshold to segment the image.

[0025] The image feature extraction module extracts texture features, shape features, and invariant moment features. The texture features are extracted using the gray-level co-occurrence matrix to obtain four parameters: energy, contrast, entropy, and correlation.

[0026] The manual annotation label module cleans and enhances the data, then uniformly names the samples in the PASCAL VOC format and manually labels the images.

[0027] The image enhancement module sets up a test set to evaluate the effect of the model, and the test set does not overlap with the training set.

[0028] The algorithm of the recognition algorithm module is as follows:

[0029] aircraft_data is the original aircraft training data set

[0030] learning_rate is the learning rate

[0031] epoch is the number of training iterations

[0032] batch_size is the sample size of the batch

[0033] y_true_d is the true aircraft class label

[0034] train_size, validation_size, and est_size are the sizes of the training set, validation set, and test set respectively

[0035] train_acc, validation_acc, and test_acc are the accuracy outputs of training, validation, and testing respectively

[0036] y_pred_cls is the predicted aircraft label

[0037] 1) Set the batch size, learning rate, and image input dimension;

[0038] 2) Determine the ratios of train_size, validation_size, and test_size;

[0039] 3) Load aircraft_data and generate aircraft class labels;

[0040] 4) Use data augmentation methods to expand aircraft_data;

[0041] 5) Build a 14-layer convolutional neural network;

[0042] 6) Use BN + ReLU + CReLU in the activation layer after convolution;

[0043] 7) Set the softmax cross-entropy loss function and the decay method of learning_rate;

[0044] 8) for i in range(1, 700)

[0045] 9) Take out samples of batch size and the corresponding

[0046] 10) y_re_cls

[0047] Take out samples of batch_size and the corresponding y_true_cls from the validation size;

[0048] 11) if i % (train size / batch size) == 0

[0049] 12) Calculate the current train ac and validation_acc;

[0050] 13) if train_acc > 0.9 and validation_acc > 0.8

[0051] 14) Keep the current model;

[0052] 15) Take out samples of batch size and the corresponding y_true_cls from the test size;

[0053] 16) Test the currently generated model and calculate test_ac;

[0054] 17) end

[0055] 18) end

[0056] 19) end.

[0057] The present invention has the following beneficial effects: 1) The image restoration processing module, image enhancement module, image correction module, and image model noise reduction module inside the image processing module cooperate with each other to perform pre-restoration processing on the pre-classified image information, thereby improving the clarity of the image. Subsequently, through the enhancement and correction of pixel points, the effect of increasing the image clarity is achieved, and at the same time, the clarity and quality after image processing are ensured; 2) The recognition algorithm module designs and optimizes the network structure model to ensure that the accuracy rate will not decrease, improves the transmission efficiency of feature information, and improves the training strategy during the training process. Remote sensing images with different resolutions are used for the initial training, and the image scale is randomly changed for training to improve the generalization ability and robustness of the model. The test results show that the optimized model has a significantly reduced model size, and the detection speed is twice that of the original. The efficient transmission of feature information ensures the detection accuracy, improves the test quality of the aircraft, and at the same time realizes the automatic acquisition, interpretation, and result processing of images during the aircraft test process, shortening the aircraft production cycle.

[0058] The following further describes the present system in conjunction with the drawings and embodiments. Description of the Drawings

[0059] Figure 1 is the system block diagram of the present invention;

[0060] Figure 2 is the system block diagram of the image processing module in the present invention;

[0061] Figure 3 is the system block diagram of the dataset construction module in the present invention; Detailed Embodiments

[0062] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the following further describes the present invention in detail in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0063] Embodiment: Refer to Figures 1-3, an aircraft recognition and detection system based on deep learning, including an image processing module, a dataset construction module, a data acquisition module, an image acquisition module, a deep learning dataset module, a startup module, a target detection module, a power supply module, a recognition algorithm module, a network connection module, an image preprocessing module, an image segmentation module, an image feature extraction module, an image model recognition module, a manual annotation label module, and an image enhancement module. The output end of the startup module is connected to the input end of the deep learning dataset module, the output end of the deep learning dataset module is connected to the input end of the image acquisition module, the output end of the image acquisition module is connected to the input end of the data acquisition module, the output end of the data acquisition module is connected to the input end of the image processing module, the power supply module and the network connection module are fixedly installed outside the image processing module, the output end of the image processing module is connected to the input end of the dataset construction module, and the output end of the dataset construction module is connected to the input end of the target detection module. The output end of the target detection module is connected to the input end of the recognition algorithm module. The specific steps are as follows:

[0064] S1: After the startup module is started, the power supply module provides electrical energy for the entire system;

[0065] S2: The deep learning dataset module collects the picture information from the global image screenshots of Google Earth, the NWPU VHR-10 dataset, the DOTA dataset, and the RSOD dataset to form an image database;

[0066] S3: The image acquisition module collects the image data of the ground object properties and states in real time through the on-board data interface;

[0067] S4: The image acquisition module transmits the image data collected according to step S2, and finally inputs the data to the image preprocessing module port;

[0068] S5: The image preprocessing module first performs grayscale processing on the image, and then uses the method of smoothing filtering to denoise the image;

[0069] S6: The image segmentation module segments the image. The input image of this segmentation method is a grayscale image. After threshold operation, a binary image is output, and the contour of the segmented image is further extracted;

[0070] S7: The image feature extraction module uses the gray-level co-occurrence matrix to extract four parameters of energy, contrast, entropy, and correlation, as well as the shape features and invariant moment features of the image;

[0071] S8: The image model recognition module uses the SSD algorithm and combines it with the Inception V2 convolutional network operation as the target image recognition model of the aircraft;

[0072] S9: The network connection module provides network connection support for the entire system;

[0073] S10: After the manual annotation label module cleans and enhances the data, it uniformly names the samples in the PASCAL VOC format and performs manual label annotation on the images;

[0074] S11: The image enhancement module sets up a test set to evaluate the effect of the model.

[0075] S12: The object detection module modifies some configuration options of the framework. For example, it removes the data enhancement method in the framework. In this object detection framework, the input size of the image is 300×300, and the non-linear activation function after the convolutional layer uses ReLU6 to achieve image object detection;

[0076] S13: After passing through the first convolutional layer Conv1, the size of the image in the recognition algorithm module is 120×120×64. The convolution uses the Same convolution. The principle of the second convolutional layer Conv2 is the same as that of Conv1, and the number of channels generated after the convolution operation remains unchanged; Next is a pooling layer Pool_1, which uses the max pooling operation. After Pool_1, there are two consecutive convolutional layers Conv3 and Conv4, and the principle is the same as the previous convolutional layers, except that the number of channels is changed to 128; Then comes the second pooling layer Pool_2, and the operation principle is the same as the previous pooling layer. After Pool_2, there are two consecutive convolutional layers Conv5 and Conv6, and the number of channels is set to 256; Then a pooling layer Pool_3 is added. After Pool_3, three convolutional layers are stacked, namely Conv7, Conv8, and Conv9, and the number of channels of these three convolutional layers is set to 512; After the three convolutional layers, a pooling layer Pool_4 is added, and then three convolutional layers Conv10, Conv11, and Conv12 are stacked and their number of channels is set to 1024. After that, a pooling layer Pool_5 is added; After Pool_5, two fully connected layers are added to predict the category of the image; The number of neurons in the first fully connected layer is set to 2048, and at the same time, the Dropout principle is applied and its value is set to 0.5; The number of neurons in the second fully connected layer is set to the number of image types, and the Dropout value is set to 0.8. Finally, the Softmax function is used to output the predicted image type.

[0077] The image processing module internally has an image preprocessing module, an image segmentation module, an image feature extraction module, and an image model recognition module. The output ends of the image preprocessing module, the image segmentation module, the image feature extraction module, and the image model recognition module are connected to the input end of the image processing module.

[0078] The internal of the dataset construction module is equipped with a manual annotation label module and an image enhancement module. The output ends of the manual annotation label module and the image enhancement module are connected to the input end of the dataset construction module.

[0079] The deep learning dataset module is the UCAS-AOD dataset. The UCAS-AOD dataset is an aerial remote sensing image target detection dataset, annotated by the Pattern Recognition and Intelligent System Development Laboratory of the University of Chinese Academy of Sciences, first released in 2014 and supplemented in 2015, mainly containing two types of targets: airplanes and cars; its data comes from the global image screenshots of Google Earth, with a total of 1000 airplane images, including 7482 airplane target samples and 7114 car samples; the NWPU VHR-10 dataset is a publicly available 10-class geospatial object detection remote sensing dataset; the DOTA dataset is an aerial photography image dataset for target detection released by Wuhan University in 2017. The images mainly come from Google Earth and satellite images, and the dataset contains 15 types of targets, with a total of 2806 images; another dataset released by Wuhan University is the RSOD dataset, which contains four types of targets: airplanes, playgrounds, overpasses, and oil drums. The 446 airplane images contain a total of 4993 airplane target samples, the 189 playground images contain 191 playground targets, the 176 overpass images contain 180 overpasses, and the 165 oil drum images contain 1586 oil drum targets.

[0080] The internal of the image preprocessing module first performs grayscale processing; for grayscale processing, the evCvColorsr, gray_img, CV-RGB2GRAY function is selected to perform grayscale operation on the image; after that, the internal of the image preprocessing module will denoise. The internal of the image preprocessing module uses the method of smoothing filtering for denoising. It mainly uses the method of filtering mask to determine the average gray value of the neighborhood pixels to replace each pixel value of the image, so as to reduce the "sharp" change of the image gray level - noise P. The implementation code is: evSmothsrc_ing, smooth_img, CV_GAUSSAN, 3, 1, 0.

[0081] The image segmentation module uses the code cvThresholdsrc, threshold_img, 100, 255, CV_THRESH_BINARY_INV to segment the image. This function performs a fixed threshold operation on a single-channel number. The input image is a grayscale image, and a binary image is output after the threshold operation; for the segmented image, contour extraction is performed. This study uses the function evFindContours to retrieve contours from the binary image and returns the number of retrieved contour points and extracts the contours.

[0082] The image feature extraction module mainly extracts texture features, shape features, and invariant moment features. The texture features mainly use the gray-level co-occurrence matrix to extract four parameters: energy, contrast, entropy, and correlation.

[0083] After the manual annotation label module cleans and enhances the data, it uniformly names the samples in the PASCAL VOC format and manually labels the images; the PASCAL VOC format selects the PASCAL VOC 2007 format, and the main file is VOCdevkit. Set the folder VOC2007 in it, and set four folders: Annotations, ImageSets, JPEGImages, and Labels in it; the Annotations folder mainly stores xml-format files, that is, the files generated after the annotation tool annotates the samples. The naming requirement of its xml file is the same as the corresponding image naming in the JPEGImages folder, so that the model can correctly read during training and testing; the JPEGImage folder is mainly used to store the training and test images. The test set test.txt, training set train.txt, validation set val.txt, and training and validation set trainval.txt are stored in the Main folder. The content of these 4 files is mainly generated by a Python script; the annotation tool selects the open-source tool labelImg provided by the github website. labelImg is a dataset production software suitable for image detection tasks. Its label storage function is convenient to use and can directly store the results as xml files, which is convenient for the dataset to be unified into the PASCAL VOC format; before annotation, the save location of the xml file needs to be reset, and then click "OpenDir" to open the folder to be marked, and start marking from the first one by default; when performing manual annotation, the category name must be in English, and the annotation box generated by the annotation is a rectangle. It is necessary to accurately annotate the targets in the annotated samples. If the annotation area is too large, other interference information will be mixed in during target discrimination, making it difficult for the model training to achieve the ideal effect; if the annotation area is too small, less learning information will be provided to the model, with a large error, and it is also difficult to achieve the expected effect; only by placing the target exactly within the annotation box as much as possible can a better training effect be achieved, which belongs to correct annotation.

[0084] The image enhancement module sets a test set to evaluate the effect of the model, and the test set should not overlap with the training set; common data partitioning methods in deep learning are: hold-out method, cross-validation method, and bootstrap method; for the hold-out method, first partition the dataset D into two mutually exclusive sets, and use one of them as the test set T and the other as the training set S; D = S ∪ T. Among them, the training set S is used to train the model, and the test set T is used to evaluate its error as an estimate of the generalization error. When dividing the dataset D, it is necessary to maintain the consistency of the data distribution to avoid errors caused by the division. For example, when performing stratified sampling on a dataset D containing 1000 samples, 70% is selected as the training set S, and the remaining 30% is used as the test set. If the positive and negative samples in D each account for 50%, then the proportion of positive and negative samples in the training set S and the test set T should also each account for 50%. If the sample distributions in the training set and the test set are quite different, it will lead to bias in training. At the same time, the training set and the test set are usually divided in an 8:2 ratio. If the proportion of the training set is too large, the trained model will be more in line with the samples in the dataset D, but too few samples in the test set will lead to unstable test results. On the contrary, if too many samples are used for the test set, the trained model will deviate from the dataset D.

[0085] The algorithm of the recognition algorithm module is as follows:

[0086] aircraft_data is the original aircraft training dataset

[0087] learning_rate is the learning rate

[0088] epoch is the number of training iterations

[0089] batch_size is the sample size of the batch

[0090] y_true_d is the true aircraft class label

[0091] train_size, validation_size, and est_size are the sizes of the training set, validation set, and test set respectively

[0092] train_acc, validation_acc, and test_acc are the accuracy outputs of training, validation, and testing respectively

[0093] y_pred_cls is the predicted aircraft label

[0094] 1) Set the batch size, learning rate, and image input dimension;

[0095] 2) Determine the ratios of train_size, validation_size, and test_size;

[0096] 3) Load aircraft_data and generate aircraft class labels;

[0097] 4) Augment aircraft_data using data augmentation methods;

[0098] 5) Build a convolutional neural network with a depth of 14 layers;

[0099] 6) Use BN+ReLU+CReLU in the activation layer after convolution;

[0100] 7) Set the softmax cross-entropy loss function and the decay method of learning_rate;

[0101] 8) for i in range(1,700)

[0102] 9) Take samples of batch size and the corresponding

[0103] 10) y_re_cls

[0104] Take samples of batch_size and the corresponding y_true_cls from the validation size;

[0105] 11) if i % (train size / batch size) == 0

[0106] 12) Calculate the current train ac and validation_acc;

[0107] 13) if train_acc > 0.9 and validation_acc > 0.8

[0108] 14) Keep the current model;

[0109] 15) Take samples of batch size and the corresponding y_true_cls from the test size;

[0110] 16) Test the currently generated model and calculate test_ac;

[0111] 17) end

[0112] 18) end

[0113] 19) end

[0114] In the present invention, the image restoration processing module, image enhancement module, image correction module, and image model noise reduction module inside the image processing module cooperate with each other to perform pre-restoration processing on the pre-classified image information, thereby improving the clarity of the image. Subsequently, through the enhancement and correction of pixel points, the effect of increasing the image clarity is achieved, and at the same time, the clarity and quality after image processing are ensured.

[0115] In the present invention, the recognition algorithm module designs and optimizes the network structure of the model to ensure that the accuracy rate does not decrease, improves the transmission efficiency of feature information, and improves the training strategy during the training process. Remote sensing images with different resolutions are used for initial training, and the image scale is randomly changed for training to improve the generalization ability and robustness of the model. The test results show that for the optimized model, the model size is greatly reduced, the detection speed is twice the original, the efficient transmission of feature information ensures the detection accuracy, improves the test quality of the aircraft, and at the same time realizes the automatic acquisition, interpretation and result processing of images during the aircraft test, shortening the aircraft production cycle.

[0116] It should be noted that in the present invention, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "comprising a..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

[0117] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An aircraft recognition and detection system based on deep learning, comprising an image processing module, a dataset construction module, a data acquisition module, an image acquisition module, a deep learning dataset module, a startup module, a target detection module, a power supply module, a recognition algorithm module, a network connection module, an image preprocessing module, an image segmentation module, an image feature extraction module, an image model recognition module, a manual annotation label module, and an image enhancement module, characterized in that: The output end of the startup module is connected to the input end of the deep learning dataset module. The output end of the deep learning dataset module is connected to the input end of the image acquisition module. The output end of the image acquisition module is connected to the input end of the data acquisition module. The output end of the data acquisition module is connected to the input end of the image processing module. A power supply module and a network connection module are fixedly installed outside the image processing module. The output end of the image processing module is connected to the input end of the dataset construction module. The output end of the dataset construction module is connected to the input end of the target detection module. The output end of the target detection module is connected to the input end of the recognition algorithm module. Through the processing and analysis of the above modules, the recognition and detection of aircraft targets in images are finally realized. The specific steps are as follows: S1: After the startup module is started, the power supply module provides electrical energy for the entire system; S2: The deep learning dataset module collects the picture information in the global image screenshots from Google Earth, the NWPU VHR-10 dataset, the DOTA dataset, and the RSOD dataset to form an image database; S3: The image acquisition module collects the image data of the ground object properties and states in real time through the on-board data interface; S4: The image acquisition module transmits the image data collected according to step S2, and finally inputs the data to the image preprocessing module port; S5: The image preprocessing module first performs grayscale processing on the image, and then uses the method of smoothing filtering to denoise the image; S6: The image segmentation module segments the image. The input image of the segmentation method is a grayscale image. After threshold operation, a binary image is output, and the contour of the segmented image is extracted; S7: The image feature extraction module uses the gray-level co-occurrence matrix to extract four parameters of energy, contrast, entropy, and correlation, as well as the shape features and invariant moment features of the image; S8: The image model recognition module uses the SSD algorithm and combines it with the Inception V2 convolutional network operation as the target image recognition model of the aircraft; S9: The network connection module provides network connection support for the entire system; S10: After the manual annotation label module cleans and enhances the data, it uniformly names the samples in the PASCAL VOC format and performs manual label annotation on the images; S11: The image enhancement module sets a test set to evaluate the effect of the model; S12: The target detection module modifies some configuration options of the framework, removes the data enhancement method in the framework. In the framework, the input size of the image is 300×300, and the non-linear activation function after the convolutional layer uses ReLU6, so as to realize image target detection; S13: After passing through the first convolutional layer Conv1, the size of the image is 120×120×64. The convolution uses the Same convolution. The principle of the second convolutional layer Conv2 is the same as that of Conv1, and the number of channels generated after the convolution operation remains unchanged. Next is a pooling layer Pool_1, which uses the max pooling operation. After Pool_1, there are two consecutive convolutional layers Conv3 and Conv4, with the same principle as the previous convolutional layers, except that the number of channels is changed to 128. Then comes the second pooling layer Pool_2, with the same operation principle as the previous pooling layer. After Pool_2, there are two consecutive convolutional layers Conv5 and Conv6, with the number of channels set to 256. Then, a pooling layer Pool_3 is added. After Pool_3, three convolutional layers are stacked, namely Conv7, Conv8, and Conv9, and the number of channels of these three convolutional layers is set to 512. After the three convolutional layers, a pooling layer Pool_4 is added, and then three convolutional layers Conv10, Conv11, and Conv12 are stacked, with their number of channels set to 1024. After that, a pooling layer Pool_5 is added. After Pool_5, two fully connected layers are added to predict the category of the image. The number of neurons in the first fully connected layer is set to 2048, and at the same time, the Dropout principle is applied and its value is set to 0.

5. The number of neurons in the second fully connected layer is set to the number of image types, and the Dropout value is set to 0.

8. Finally, the Softmax function is used to output the predicted image type.

2. The aircraft recognition and detection system based on deep learning according to claim 1, characterized in that: Inside the image processing module, there are an image preprocessing module, an image segmentation module, an image feature extraction module, and an image model recognition module. The output ends of the image preprocessing module, the image segmentation module, the image feature extraction module, and the image model recognition module are connected to the input end of the image processing module.

3. The aircraft recognition and detection system based on deep learning according to claim 1, characterized in that: Inside the dataset construction module, there are a manual annotation label module and an image enhancement module. The output ends of the manual annotation label module and the image enhancement module are connected to the input end of the dataset construction module.

4. The aircraft recognition and detection system based on deep learning according to claim 1, characterized in that: The deep learning dataset module is the UCAS-AOD dataset, which is an aerial remote sensing image target detection dataset.

5. The aircraft recognition and detection system based on deep learning according to claim 1, characterized in that: Inside the image preprocessing module, grayscale processing is first performed; the evCvColor function is selected for grayscale operation on the image; after that, noise reduction is performed inside the image preprocessing module, and the image preprocessing module uses the method of smoothing filtering for noise reduction.

6. The aircraft recognition and detection system based on deep learning according to claim 1, characterized in that: The image segmentation module uses the code cvThreshold to segment the image.

7. The aircraft recognition and detection system based on deep learning according to claim 1, characterized in that: The image feature extraction module extracts texture features, shape features, and invariant moment features. The texture features use the gray-level co-occurrence matrix to extract four parameters: energy, contrast, entropy, and correlation.

8. The aircraft recognition and detection system based on deep learning according to claim 1, characterized in that: After the manual annotation label module cleans and enhances the data, it uniformly names the samples in the PASCAL VOC format and manually labels the images.

9. The aircraft recognition and detection system based on deep learning according to claim 1, characterized in that:The image enhancement modules set up a test set to evaluate the effect of the model, and the test set does not overlap with the training set.

Images