Vehicle target detection and identification method and system based on deep learning
Through the deep learning-based vehicle object detection and recognition method, the improved yolov5 detection network and expert knowledge base are used to solve the problems of low vehicle object detection accuracy and high false alarm rate in SAR images, and efficient vehicle object recognition is achieved.
Patent Information
- Application Number
- CN202411210239.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-30
- Publication Date
- 2025-01-17
AI Technical Summary
In the field of SAR images, vehicle object detection and recognition have problems such as small number of samples, low detection accuracy, high false alarm rate and long detection time. The traditional method has a single feature category and is not universal, and deep learning methods require a large number of samples for train.
Deep learning-based vehicle object detection and recognition methods include obtaining SAR images to be tested for preprocessing and blocking processing, using the improved yolov5 detection network for detection, and changing it to the BOT module, FocalLoss loss function and DIOU_nms with category weights Prediction box filtering, and remove false alarms in combination with the expert knowledge base.
It improves the accuracy and efficiency of vehicle target detection and recognition, reduces false alarm rate, and is suitable for small sample target recognition, adapting to the detection of multi-scale and diverse vehicle targets.
Smart Images
Figure CN119323663A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image processing technology, and in particular to a vehicle target detection and recognition method and system based on deep learning. Background Art
[0002] In the field of SAR images, vehicle targets are important targets. There are many types of targets, various target sizes, and few target samples. Therefore, the detection and recognition of vehicle targets has always been an important research direction. At present, there are two main types of vehicle target detection and recognition methods: traditional target detection and recognition methods and deep learning-based target detection methods. Among them, the traditional SAR image target detection method is based on manual feature extraction, such as geometric features, polarization features, and background clutter features. The suspected targets in the image are selected, and then the features of the identification targets are compared with the candidate targets one by one. The targets that meet the conditions are finally determined as the correct targets. However, the feature categories extracted by the traditional method are single and not universal. They are often limited to certain specific places, and changing scenes often cannot achieve good results. The deep learning target detection and recognition method requires sufficient target samples for training to obtain a set of network weights, and then the network weights are applied to the image to be detected and recognized, so that the target can be automatically recognized from the image. Deep learning has significantly improved the detection and recognition effect of targets on visible light images compared with traditional algorithms.
[0003] Therefore, in order to address the current problems of small number of vehicle target samples, low target detection accuracy, high false alarm rate and long detection and recognition time, a SAR image vehicle target detection and recognition method based on deep learning is needed. Summary of the invention
[0004] In order to solve the technical problems existing in the background technology, the present invention proposes a vehicle target detection and recognition method and system based on deep learning.
[0005] The present invention proposes a vehicle target detection and recognition method based on deep learning, comprising:
[0006] Acquire the SAR image to be tested;
[0007] Preprocess the SAR image to be tested;
[0008] The preprocessed SAR image to be tested is processed into blocks to obtain slices of the SAR image to be tested;
[0009] The vehicle target detection and recognition model is used to detect the SAR image slice to be tested, and the detection and recognition result is obtained;
[0010] According to the preset expert knowledge base, false alarms are removed from the detection and recognition results.
[0011] Preferably, the vehicle target detection and recognition model is an improved yolov5 detection network;
[0012] Among them, the improved yolov5 detection network updates the data augmentation method at the input end, using random cropping and random arrangement to perform image splicing;
[0013] The BOT module is embedded after the SPP module of the yolov5 backbone network;
[0014] A FocalLoss loss function with category weights is introduced, and the nms of prediction box screening is changed to DIOU_nms.
[0015] Preferably, before detecting the SAR image after denoising using the vehicle target detection and recognition model to obtain the detection and recognition result, the method further includes:
[0016] Build training model;
[0017] Obtain a dataset of SAR image sample slices with target type annotations;
[0018] The SAR image sample slice dataset is divided into a training set and a test set;
[0019] Augment the training set:
[0020] Use the training set to train the training model to obtain a training model that has passed the training;
[0021] The training model that has passed the training is tested using the test set, and the training model that has passed the test is determined as the vehicle target detection and recognition model.
[0022] Preferably, obtaining a SAR image sample slice data set with target type annotations specifically includes:
[0023] Obtain a set of SAR images known to contain targets;
[0024] Preprocessing the SAR image samples in the SAR image set; wherein the preprocessing includes quantization and filtering denoising;
[0025] According to the preset expert knowledge base, target type annotation is performed on the preprocessed SAR image samples to obtain SAR image samples with target type annotation;
[0026] The SAR image samples with target type annotations are sliced to obtain SAR image sample slices with target type annotations, and a SAR image sample slice dataset is formed.
[0027] Preferably, the training set is expanded, specifically including:
[0028] According to the number of target sample slices in the training set, one or more of rotation, translation, contrast adjustment and target simulation are used to amplify the target sample slices with a smaller number.
[0029] Preferably, according to the number of target sample slices in the training set, target simulation is used to amplify target sample slices with a smaller number, specifically including:
[0030] Select a small number of target sample slices, use the electromagnetic simulation tool to select the corresponding three-dimensional model, set the band, size, azimuth and elevation information, and obtain the corresponding target simulation slices;
[0031] Perform image fusion on the obtained target simulation slice and the SAR image sample corresponding to the target sample slice;
[0032] The fused image is sampled and sliced to obtain the target simulation sample.
[0033] Preferably, the obtained target simulation slice is image-fused with the SAR image sample corresponding to the target sample slice, specifically comprising:
[0034] Determine the position to be fused of the target simulation slice according to the latitude and longitude range of the vehicle target parking area in the SAR image sample corresponding to the target sample slice;
[0035] The target simulation slice is fused with the SAR image sample using an image fusion method, and the pixel coordinates of the center point of the target simulation slice in the SAR image sample are recorded. According to the length and width of the target simulation slice, the corresponding target annotation information is generated.
[0036] Preferably, the SAR image to be measured is preprocessed, specifically including:
[0037] Quantify the SAR image to be tested;
[0038] The quantized SAR image to be tested is filtered and denoised.
[0039] Preferably, before the preprocessed SAR image is divided into blocks to obtain SAR image slices, the method further includes:
[0040] According to the latitude and longitude information of the existing vehicle target parking area, it is judged whether there is a parking area of key concern in the pre-processed SAR image;
[0041] If yes, then extract the image of the parking area of the focus, and process the extracted image in blocks;
[0042] If not, no processing is performed on the preprocessed SAR image, so as to facilitate block processing of the preprocessed SAR image.
[0043] Preferably, according to a preset expert knowledge base, false alarm removal is performed on the detection and recognition results, specifically including:
[0044] According to the detection and recognition results, the coordinate information of each identified target in the detection and recognition results in the SAR image to be tested is obtained;
[0045] According to the coordinate information of each identified target in the SAR image to be tested, the length, width and scattering characteristics of each identified target are calculated;
[0046] According to the length, width and scattering characteristics of the target in the preset expert knowledge base and the possible parking area position, each identified target is matched and analyzed to obtain each identified target and a reference target corresponding to each identified target;
[0047] The Euclidean distance is used to calculate the feature similarity between the identified target and the corresponding reference target;
[0048] Based on the feature similarity and the pre-set empirical threshold, it is determined whether the identified target is a vehicle and which type of vehicle it belongs to, thereby achieving the purpose of eliminating false alarms.
[0049] The present invention also proposes a vehicle target detection and recognition system based on deep learning, comprising:
[0050] The first acquisition module is used to acquire the SAR image to be tested;
[0051] A preprocessing module, used for preprocessing the SAR image to be tested;
[0052] The slicing module is used to perform block processing on the preprocessed SAR image to be tested to obtain slices of the SAR image to be tested;
[0053] The recognition module is used to detect the SAR image slice to be tested by using the vehicle target detection and recognition model to obtain the detection and recognition result;
[0054] The false alarm removal module is used to remove false alarms from the detection and recognition results according to a preset expert knowledge base.
[0055] Preferably, the vehicle target detection and recognition model is an improved yolov5 detection network;
[0056] Among them, the improved yolov5 detection network updates the data augmentation method at its input end, using random cropping and random arrangement to perform image splicing;
[0057] The BOT module is embedded after the SPP module of the yolov5 backbone network;
[0058] A FocalLoss loss function with category weights is introduced, and the nms of prediction box screening is changed to DIOU_nms.
[0059] Preferably, it also includes:
[0060] Model building module, used to build training models;
[0061] The second acquisition module is used to acquire a SAR image sample slice data set with target type annotations;
[0062] A partitioning module is used to partition the SAR image sample slice data set into a training set and a test set;
[0063] The augmentation module is used to augment the training set:
[0064] A training module is used to train the training model using the training set to obtain a training model that passes the training;
[0065] The testing module uses the test set to test the trained model, and determines the trained model that has passed the test as the vehicle target detection and recognition model.
[0066] Preferably, it also includes:
[0067] The key area extraction module is used to extract the key area of the preprocessed SAR image to be tested.
[0068] In the present invention, the proposed vehicle target detection and recognition method and system based on deep learning first preprocesses and blocks the acquired SAR image to be tested to obtain the SAR image slices to be tested, and then uses the vehicle target detection and recognition model to detect the SAR image slices to obtain the detection and recognition results, and then introduces expert knowledge to reduce false alarms, thereby ensuring the accuracy of target detection and recognition. BRIEF DESCRIPTION OF THE DRAWINGS
[0069] Figure 1 A schematic diagram of the flow of a vehicle target detection and recognition method based on deep learning in one embodiment of the present invention.
[0070] Figure 2 The figure is a schematic diagram of the structure of a vehicle target detection and recognition model in one embodiment of the present invention.
[0071] Figure 3 This is a schematic diagram of the structure of a BOT module in an embodiment of the present invention. DETAILED DESCRIPTION
[0072] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.
[0073] Reference Figure 1-3 The present invention proposes a vehicle target detection and recognition method based on deep learning, comprising:
[0074] Acquire the SAR image to be tested;
[0075] Preprocess the SAR image to be tested;
[0076] The preprocessed SAR image to be tested is processed into blocks to obtain slices of the SAR image to be tested;
[0077] The vehicle target detection and recognition model is used to detect the SAR image slice to be tested, and the detection and recognition result is obtained;
[0078] According to the preset expert knowledge base, false alarms are removed from the detection and recognition results.
[0079] The present invention first pre-processes and blocks the acquired SAR image to be tested to obtain slices of the SAR image to be tested, then uses a vehicle target detection and recognition model to detect the slices of the SAR image to be tested to obtain detection and recognition results, and then introduces expert knowledge to reduce false alarms, thereby ensuring the accuracy of target detection and recognition.
[0080] In this embodiment, before the SAR image after denoising is detected by using the vehicle target detection and recognition model to obtain the detection and recognition result, the following is also included:
[0081] Build training model;
[0082] Obtain a dataset of SAR image sample slices with target type annotations;
[0083] The SAR image sample slice dataset is divided into a training set and a test set;
[0084] Expand the training set: Use the training set to train the training model to obtain a training model that has passed the training;
[0085] The training model that has passed the training is tested using the test set, and the training model that has passed the test is determined as the vehicle target detection and recognition model.
[0086] The present embodiment is configured such that sample amplification is performed on target data with a small sample size in the training set before training, which can ensure that the number of targets of each type in the target data set is relatively small, and avoid the influence of small sample number, uneven sample distribution and other situations on the training and detection and recognition effects. It is suitable for the recognition of small sample targets under low security conditions and improves the target detection accuracy of small sample targets.
[0087] In further embodiments, the amplification includes one or more of rotation, translation, adjustment of contrast, and electromagnetic simulation.
[0088] Among them, the vehicle target detection and recognition model is an improved yolov5 detection network.
[0089] Specifically, the input end of the improved yolov5 detection network updates the data augmentation method, using random cropping and random arrangement to perform image splicing;
[0090] The BOT module is embedded after the SPP module of the yolov5 backbone network;
[0091] A FocalLoss loss function with category weights is introduced, and the nms of prediction box screening becomes DIOU_nms.
[0092] This embodiment is configured as follows: a BOT module is embedded after the SPP module of the yolov5 backbone network. The BOT module combines the advantages of CNN and Transformer, and can realize the extraction of global and local features. It effectively solves the problem of multi-scale and diverse vehicle targets and the influence of the background environment on target detection, and further ensures the accuracy of target detection; and by introducing a FocalLoss loss function containing category weights, it can effectively improve the problem of category imbalance in the sample data set; and by changing the nms of the prediction box screening to DIOU_nms, which conforms to the mechanism of target box regression, the distance between the target and the anchor, the overlap rate and the scale are taken into account, so that the target box regression becomes more stable, and will not have problems such as divergence during training like IoU and GIoU.
[0093] In a further embodiment, obtaining a SAR image sample slice data set with target type annotations specifically includes:
[0094] Obtain a set of SAR images known to contain targets;
[0095] Preprocessing the SAR image samples in the SAR image set; wherein the preprocessing includes quantization and filtering denoising;
[0096] According to the preset expert knowledge base, target type annotation is performed on the denoised SAR image samples to obtain SAR image samples with target type annotation;
[0097] The SAR image samples with target type annotations are sliced to obtain SAR image sample slices with target type annotations, and a SAR image sample slice dataset is formed.
[0098] This embodiment reduces the difficulty of network convergence in the model training process by quantizing the SAR image to be tested, reduces unnecessary calculations, and enhances the contrast and feature difference between the target and the background.
[0099] In this embodiment, the ratio of the training set to the test set is 7:3.
[0100] In a further embodiment, the training set is augmented, specifically comprising:
[0101] According to the number of target sample slices in the training set, one or more of rotation, translation, contrast adjustment and target simulation are used to amplify the target sample slices with a smaller number, thereby ensuring the amplification of small sample target samples, which is beneficial to improving the accuracy of the training model obtained during testing, thereby ensuring the target detection accuracy in subsequent detection and recognition.
[0102] Among them, the specific method of target simulation is:
[0103] Select a target with a small number of sample slices, use the electromagnetic simulation tool to select the corresponding three-dimensional model, set the band, size, azimuth and elevation information, and obtain the corresponding target simulation slice;
[0104] Perform image fusion on the obtained target simulation slice and the SAR image sample corresponding to the target sample slice;
[0105] The fused image is sampled and sliced to obtain the target simulation sample.
[0106] Among them, the specific steps of image fusion are:
[0107] Determine the position to be fused of the target simulation slice according to the latitude and longitude range of the vehicle target parking area in the SAR image sample corresponding to the target sample slice;
[0108] The target simulation slice is fused with the SAR image sample by using an image fusion method;
[0109] The pixel coordinates of the center point of the target simulation slice in the SAR image sample are recorded, and the corresponding target annotation information is generated according to the length and width of the target simulation slice.
[0110] With this setting, the simulation results are used to amplify samples and the corresponding annotation information is generated, without the need for manual annotation, which reduces the workload of data annotation.
[0111] In this embodiment, the SAR image to be measured is preprocessed, specifically including:
[0112] Quantify the SAR image to be tested to enhance the contrast and feature difference between the target and the background;
[0113] The quantized SAR image to be tested is subjected to filtering and denoising processing, so as to facilitate subsequent block processing of the SAR image to be tested after filtering and denoising processing to obtain slices of the SAR image to be tested.
[0114] In this embodiment, quantizing the SAR image specifically includes:
[0115] The grayscale histogram stretching method is used to quantize the SAR image to convert the 16-bit tiff format image into 8-bit; the quantized image will reduce the difficulty of network convergence during the model training process, reduce unnecessary calculations, and enhance the contrast and feature differences between the target and the background.
[0116] Among them, the grayscale histogram stretching method is used to quantify the SAR image, which specifically includes:
[0117] According to the distribution of pixel values of SAR images, the images were divided into 655 intervals with 100 values as one interval, and the number of pixels in each interval was counted;
[0118] Set the threshold of pixel value to 0 and 255;
[0119] Determine whether the number of intervals between two thresholds reaches a preset standard value;
[0120] If yes, the values of the pixels between the two thresholds are linearly stretched to 0-255;
[0121] If not, reset the pixel value thresholds to 0 and 255 until the number of intervals between the two thresholds reaches a preset standard value.
[0122] In this embodiment, filtering and denoising are performed on the quantized SAR image, specifically including:
[0123] Lee's improved filtering algorithm is used to filter and denoise the quantized SAR image, so as to perform filtering and denoising on the premise of retaining the target feature information.
[0124] In order to improve recognition efficiency, in this embodiment, before the pre-processed SAR image is divided into blocks to obtain SAR image slices, the following steps are further included:
[0125] Extract key areas from the preprocessed SAR image to obtain the latitude and longitude information of the existing vehicle target parking area;
[0126] According to the latitude and longitude information of the existing vehicle target parking area, it is judged whether there is a parking area of key concern in the SAR image after denoising;
[0127] If yes, the image range is extracted according to the parking area range, and the extracted influence range is processed in blocks;
[0128] If not, no processing is performed on the preprocessed SAR image, so as to facilitate block processing of the preprocessed SAR image.
[0129] The latitude and longitude information of the target parking area of the existing vehicle is pre-known information.
[0130] Among them, this embodiment adopts a sliding window slicing method to perform block processing on the SAR image after denoising, that is, a rectangular frame is used to slide the window according to a certain step size within the image area to be detected, and the previous slice and the next slice have a little overlapping area to ensure that the potential target can appear completely on at least one slice.
[0131] In this embodiment, according to the preset expert knowledge base, false alarm removal is performed on the detection and recognition results, specifically including:
[0132] According to the detection and recognition results, the coordinate information of each identified target in the detection and recognition results in the SAR image to be tested is obtained;
[0133] According to the coordinate information of each identified target in the SAR image to be tested, the length, width and scattering characteristics of each identified target are calculated;
[0134] According to the length, width and scattering characteristics of the target in the preset expert knowledge base and the possible parking area position, each identified target is matched and analyzed to obtain each identified target and a reference target corresponding to each identified target;
[0135] The Euclidean distance is used to calculate the feature similarity between the identified target and the corresponding reference target;
[0136] Based on the feature similarity and the pre-set empirical threshold, it is determined whether the identified target is a vehicle and which type of vehicle it belongs to, thereby achieving the purpose of eliminating false alarms.
[0137] The present invention also provides a vehicle target detection and recognition system based on deep learning, comprising:
[0138] The first acquisition module is used to acquire the SAR image to be tested;
[0139] A preprocessing module, used for preprocessing the SAR image to be tested;
[0140] The slicing module is used to perform block processing on the preprocessed SAR image to be tested to obtain slices of the SAR image to be tested;
[0141] The recognition module is used to detect the SAR image slice to be tested by using the vehicle target detection and recognition model to obtain the detection and recognition result;
[0142] The false alarm removal module is used to remove false alarms from the detection and recognition results according to a preset expert knowledge base.
[0143] The vehicle target detection and recognition model in this embodiment is an improved yolov5 detection network; wherein the improved yolov5 detection network updates the data augmentation method at its input end, and uses random cropping and random arrangement to perform image splicing;
[0144] The BOT module is embedded after the SPP module of the yolov5 backbone network;
[0145] A FocalLoss loss function with category weights is introduced, and the nms of prediction box screening is changed to DIOU_nms.
[0146] Among them, preprocessing includes quantization and filtering denoising.
[0147] In this embodiment, a key area extraction module is also included. The key area extraction module is used to extract key areas from the preprocessed SAR image to be tested, so that the subsequent slicing module can perform block processing on the extracted key areas to improve efficiency.
[0148] In this embodiment, it also includes:
[0149] Model building module, used to build training models;
[0150] The second acquisition module is used to acquire a SAR image sample slice data set with target type annotations;
[0151] A partitioning module is used to partition the SAR image sample slice data set into a training set and a test set;
[0152] The augmentation module is used to augment the training set:
[0153] A training module is used to train the training model using the training set to obtain a training model that passes the training;
[0154] The testing module uses the test set to test the trained model, and determines the trained model that has passed the test as the vehicle target detection and recognition model.
[0155] The amplification module is used to amplify target sample slices with a smaller number in the training set by using one or more of rotation, translation, contrast adjustment and target simulation.
[0156] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.
Claims
1. A vehicle target detection and recognition method based on deep learning, characterized in that: include: Acquire the SAR image to be tested; Preprocess the SAR image to be tested; The preprocessed SAR image to be tested is processed into blocks to obtain slices of the SAR image to be tested; The vehicle target detection and recognition model is used to detect the SAR image slice to be tested, and the detection and recognition result is obtained; According to the preset expert knowledge base, false alarms are removed from the detection and recognition results.
2. The vehicle target detection and recognition method based on deep learning according to claim 1 is characterized in that: The vehicle target detection and recognition model is an improved yolov5 detection network; Among them, the improved yolov5 detection network updates the data augmentation method at the input end, using random cropping and random arrangement to perform image splicing; The BOT module is embedded after the SPP module of the yolov5 backbone network; A FocalLoss loss function with category weights is introduced, and the nms of prediction box screening is changed to DIOU_nms.
3. The vehicle target detection and recognition method based on deep learning according to claim 1 is characterized in that: Before using the vehicle target detection and recognition model to detect the SAR image after denoising and obtaining the detection and recognition result, the following steps are also included: Build a training model; Obtain a dataset of SAR image sample slices with target type annotations; The SAR image sample slice dataset is divided into a training set and a test set; Augment the training set: Use the training set to train the training model to obtain a training model that has passed the training; The training model that has passed the training is tested using the test set, and the training model that has passed the test is determined as the vehicle target detection and recognition model.
4. The vehicle target detection and recognition method based on deep learning according to claim 3 is characterized in that: Obtain a dataset of SAR image sample slices with target type annotations, including: Obtain a set of SAR images known to contain targets; Preprocessing the SAR image samples in the SAR image set; wherein the preprocessing includes quantization and filtering denoising; According to the preset expert knowledge base, target type annotation is performed on the preprocessed SAR image samples to obtain SAR image samples with target type annotation; Slice the SAR image samples with target type annotations to obtain SAR image sample slices with target type annotations, and form a SAR image sample slice data set; Preferably, the training set is expanded, specifically including: According to the number of each target sample slice in the training set, one or more of rotation, translation, contrast adjustment and target simulation are used to amplify the target sample slices with a smaller number; Preferably, according to the number of target sample slices in the training set, target simulation is used to amplify targets with a smaller number, specifically including: Select a small number of targets, use the electromagnetic simulation tool to select the corresponding three-dimensional model, set the band, size, azimuth and elevation information, and obtain the corresponding target simulation slice; Perform image fusion on the obtained target simulation slice and the SAR image sample corresponding to the target sample slice; Perform sample slicing on the fused image to obtain the target simulation sample; Preferably, the obtained target simulation slice is image-fused with the SAR image sample corresponding to the target sample slice, specifically comprising: Determine the position to be fused of the target simulation slice according to the latitude and longitude range of the vehicle target parking area in the SAR image sample corresponding to the target sample slice; The target simulation slice is fused with the SAR image sample using an image fusion method, and the pixel coordinates of the center point of the target simulation slice in the SAR image sample are recorded. According to the length and width of the target simulation slice, the corresponding target annotation information is generated.
5. The vehicle target detection and recognition method based on deep learning according to claim 1 is characterized in that: Preprocess the SAR image to be tested, including: Quantify the SAR image to be tested; Perform filtering and denoising on the quantized SAR image to be tested; Preferably, before the preprocessed SAR image is divided into blocks to obtain SAR image slices, the method further includes: According to the latitude and longitude information of the existing vehicle target parking area, it is judged whether there is a parking area of key concern in the pre-processed SAR image; If so, an image of the focused parking area is extracted, and the extracted image is processed in blocks.
6. The vehicle target detection and recognition method based on deep learning according to claim 1 is characterized in that: According to the preset expert knowledge base, false alarms are removed from the detection and recognition results, including: According to the detection and recognition results, the coordinate information of each identified target in the detection and recognition results in the SAR image to be tested is obtained; According to the coordinate information of each identified target in the SAR image to be tested, the length, width and scattering characteristics of each identified target are calculated; According to the length, width and scattering characteristics of the target in the preset expert knowledge base and the possible parking area position, each identified target is matched and analyzed to obtain each identified target and a reference target corresponding to each identified target; The Euclidean distance is used to calculate the feature similarity between the identified target and the corresponding reference target; Based on the feature similarity and the pre-set empirical threshold, it is determined whether the identified target is a vehicle and which type of vehicle it belongs to, thereby achieving the purpose of eliminating false alarms.
7. A vehicle target detection and recognition system based on deep learning, characterized in that: include: The first acquisition module is used to acquire the SAR image to be tested; A preprocessing module, used for preprocessing the SAR image to be tested; The slicing module is used to perform block processing on the preprocessed SAR image to be tested to obtain slices of the SAR image to be tested; The recognition module is used to detect the SAR image slice to be tested by using the vehicle target detection and recognition model to obtain the detection and recognition result; The false alarm removal module is used to remove false alarms from the detection and recognition results according to a preset expert knowledge base.
8. The vehicle target detection and recognition system based on deep learning according to claim 7 is characterized in that: The vehicle target detection and recognition model is an improved yolov5 detection network; Among them, the improved yolov5 detection network updates the data augmentation method at its input end, using random cropping and random arrangement to perform image splicing; The BOT module is embedded after the SPP module of the yolov5 backbone network; A FocalLoss loss function with category weights is introduced, and the nms of prediction box screening is changed to DIOU_nms.
9. The vehicle target detection and recognition system based on deep learning according to claim 7, characterized in that: Also includes: Model building module, used to build training models; The second acquisition module is used to acquire a SAR image sample slice data set with target type annotations; A partitioning module is used to partition the SAR image sample slice data set into a training set and a test set; The augmentation module is used to augment the training set: A training module is used to train the training model using the training set to obtain a training model that passes the training; The testing module uses the test set to test the trained model, and determines the trained model that has passed the test as the vehicle target detection and recognition model.
10. The vehicle target detection and recognition system based on deep learning according to claim 7, characterized in that: Also includes: The key area extraction module is used to extract the key area of the preprocessed SAR image to be tested.