A SAR image target detection method combining deep learning and CFAR

Abstract

The present invention relates to the field of radar target recognition, and more specifically to a SAR image target detection method that combines deep learning with CFAR. The present invention significantly improves the target detection rate of SAR images by combining a CNN network, a YOLO V5 network, and a CFAR algorithm. A network model with few parameters and a fast CFAR detection algorithm are used, as well as a CNN network that filters out complex scene slices that do not contain targets, thereby eliminating them from YOLO V5 and CFAR detection. This reduces the detection time and improves the detection efficiency of the entire target recognition process. The YOLO V5 neural network structure is used, and data sets are preprocessed through normalization, data enhancement, and adaptive calculations to obtain multi-layer features of SAR images. This method is more robust than traditional recognition methods and target recognition methods using network structures such as Fast RCNN.

Description

Technical Field

[0001] The present invention relates to the field of radar target recognition, and in particular to a SAR image target detection method combining deep learning and CFAR. Background Art

[0002] SAR (Synthetic Aperture Radar) is a high-resolution coherent imaging radar system. It is independent of sunlight and capable of all-day, all-weather Earth observation. It can effectively detect the spatial morphological characteristics of target objects even in extremely low-visibility weather conditions. Consequently, SAR systems are increasingly widely used in both civilian and military applications. With the continuous development of SAR technology and the continuous improvement of its imaging resolution, the field of SAR image target recognition faces increasing opportunities, but also many challenges.

[0003] There are many existing target recognition methods for SAR images, which can be roughly divided into two categories. One category is traditional SAR image target recognition methods, mainly represented by the CFAR (Constant False-Alarm Rate) detection algorithm. The CFAR algorithm is simple, fast, and highly practical. In the field of target recognition, the CFAR algorithm can maintain a set false alarm rate to detect the presence of targets in the image. However, due to the special nature of SAR imaging, the CFAR algorithm has limitations in suppressing clutter noise and controlling false alarms in complex areas during the SAR image detection process. Although some improved CFAR algorithms have been proposed, they still have difficulty resolving the false alarm problem in complex areas.

[0004] Another category is SAR image target recognition methods based on deep learning. In recent years, deep learning has been rapidly applied to various fields, including medicine, industry, and radar. The CNN (Convolutional Neural Network), a classic deep learning network, has become the backbone of the popular and highly effective YOLO V5 target recognition and detection network. The advantage of deep learning methods lies in their use of multidimensional images as input, avoiding the complex feature extraction and data reconstruction processes required by traditional target recognition algorithms, resulting in significant success in target detection. In the field of SAR image target recognition, targeted target recognition algorithms exist based on the different spatial morphological representations of different types of targets in SAR images. However, existing SAR image detection and recognition methods only utilize the statistical characteristics of local regions in SAR images, effectively achieving pixel-level detection. In other words, current deep learning SAR image target recognition methods require high contrast between the target and the background and distinct discriminative features. Consequently, their detection performance is good in simple scenes, but extremely poor in complex ones. Summary of the Invention

[0005] In response to the problems existing in the prior art, the purpose of the present invention is to provide a SAR image target detection method that combines deep learning and CFAR.

[0006] In order to achieve the above objectives, the present invention adopts the following technical solutions to achieve them.

[0007] A SAR image target detection method combining deep learning and CFAR includes the following steps:

[0008] Step 1: Process N SAR images to obtain a training set and a test set; perform sliding window slicing on the training set and the test set to obtain a training set after sliding window slicing and a test set after sliding window slicing, respectively;

[0009] Step 2: Process the training set and the test set after sliding window slicing; use the processed training set data to train the CNN network to obtain a trained CNN network; then input the processed test set data into the trained CNN network to obtain the classification results of the processed test set data; record the set of all open scenes or slice images containing objects in the test set as the retention set;

[0010] Step 3: Preprocess the training set images and the retained set images after sliding window slicing; use the preprocessed training set data to train the YOLO V5 network to obtain a trained YOLO V5 network; then input the preprocessed retained set data into the trained YOLO V5 network to obtain a target recognition result slice dataset with the least false alarms, which is recorded as dataset A.

[0011] Step 4: Perform the maximum confidence determination of the output box and the total number of detected targets on the data set A. All data in the data set A that meet the determination criteria are used as the primary detection data set. The data in the primary detection data set is input into the CFAR detection algorithm for secondary detection to obtain the secondary detection data set.

[0012] Step 5: Merge the secondary detection dataset data with the primary detection dataset data to obtain the final target detection result.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] First, high target recognition rate: Traditional methods for target recognition in SAR image data generally simply use traditional detectors or classifiers to directly classify and detect the raw data to obtain target recognition results. This does not extract high-dimensional feature data information from SAR images, resulting in low recognition rates for traditional methods. Due to the special nature of SAR image data, deep learning methods for extracting SAR image feature information generally fail to achieve ideal recognition results. To address this, the present invention combines CNN and YOLO V5 with the traditional CFAR detection algorithm to identify targets in SAR images step by step, significantly improving the target recognition rate.

[0015] Second, the detection speed is fast: While not affecting the target recognition rate, the method of the present invention adopts a network model with very few parameters and a fast CFAR detection algorithm. On the other hand, it uses a CNN network to filter out complex scene slices that do not contain targets, so that they are not included in the YOLO V5 and CFAR detection, saving time. The detection process takes less time, thereby improving the detection efficiency of the entire process.

[0016] Third, strong robustness: Since the method of the present invention adopts the YOLO V5 neural network structure and performs normalization, data enhancement and adaptive calculation preprocessing on the data set, it can obtain multi-layer features of SAR images. Compared with traditional recognition methods and target recognition methods with network structures such as Fast RCNN, it has strong robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Figure 1 This is a flowchart of the SAR image target detection method combining deep learning and CFAR in the present invention. DETAILED DESCRIPTION

[0019] The embodiments of the present invention will be described in detail below with reference to examples. However, those skilled in the art will understand that the following examples are only used to illustrate the present invention and should not be construed as limiting the scope of the present invention.

[0020] refer to Figure 1 , a SAR image target detection method combining deep learning and CFAR, comprising the following steps:

[0021] Step 1: Process N SAR images to obtain a training set and a test set; perform sliding window slicing on the training set and the test set to obtain a training set after sliding window slicing and a test set after sliding window slicing, respectively;

[0022] Specifically, the positions and categories of targets in N SAR images are obtained, and the positions and categories of targets in the SAR images are used as labels; a SAR image and the corresponding label are used as a set of data to obtain a sample set; the sample set is randomly divided into a training set and a test set according to a proportion; the SAR images of the training set and the test set are sliced by sliding windows to obtain a fixed-size slice and the label corresponding to each slice for each SAR image, thereby obtaining a training set after sliding window slicing and a test set after sliding window slicing.

[0023] The specific process is as follows:

[0024] Determine N SAR images and annotate them using the open source image annotation software LabelImg. Label the target categories and locations on the SAR images, and store this information in an XML file. Generate an XML label file for each SAR image, and generate XML label files corresponding to N SAR images.

[0025] A SAR image and its corresponding label are taken as a set of data to obtain a sample set; the sample set is randomly divided into a training set and a test set according to proportion.

[0026] The SAR image is sliced through a sliding window, and the XML label information is modified as the sliding window slices. After the sliding window slices, a fixed-size slice image with a length and width of I is obtained, as well as the XML label corresponding to the slice image. The XML label file is converted into a txt file and saved. A txt file is generated for each slice image. Each line in the txt file represents a target information of the slice image. This information is the target category and the center coordinates of the target box and the length and width, a total of 5 information, respectively recorded as: class, x-center, y-center, width, height; at this time, the txt label information is recorded as YOLO-Label. The fixed size of the slice is I = 640.

[0027] Step 2: Process the training set and the test set after sliding window slicing; use the processed training set data to train the CNN network to obtain a trained CNN network; then input the processed test set data into the trained CNN network to obtain the classification results of the processed test set data; record the set of all open scenes or slice images containing objects in the test set as the retention set;

[0028] The training set and the test set after sliding window slicing are processed. Specifically, the slice images of the training set and the test set after sliding window slicing are marked: the labels of the slice images of complex scenes without targets are recorded as 0, the labels of the slice images containing targets are recorded as 1, and the labels of other slice images are recorded as 2;

[0029] Then, the slice images and corresponding labels in the training set after a sliding window slicing are used as a set of data to obtain the CNN training set, which is recorded as CNN-Train-Slices; the slice images and corresponding labels in the test set after a sliding window slicing are used as a set of data to obtain the CNN test set, which is recorded as CNN-Test-Slices.

[0030] The CNN network specifically includes four convolutional layers and a fully connected layer connected in sequence; the convolutional layer is used to convolve and downsample the input slice image, and the fully connected layer is used to perform linear transformation processing on the features output by the fourth convolutional layer to obtain the classification result of the input slice image;

[0031] The first convolutional layer is used to convolve and downsample the input slice image to obtain C feature maps x′ after the first convolutional layer downsampling. The process is as follows:

[0032] (1) Normalize the CNN-Train-Slices data;

[0033] (2) The first convolution layer includes C convolution kernels, and the C convolution kernels of the first convolution layer are denoted as K, and the size of K is set to 3×3×1, and the convolution step is set to 1, which is used to convolve the input slice image; the first convolution layer is used to convolve the normalized CNN-Train-Slices with the C convolution kernels of the first convolution layer, and the C convolution results of the first convolution layer are obtained, which are denoted as x:

[0034]

[0035] Wherein, I is the fixed size of the slice in step 1, that is, the input image size; p is the padding size in the convolution; k is the size of the convolution kernel; s is the convolution step; in this embodiment, p = 0, k = 3, and s = 1.

[0036] (3) Input the C convolution results x into the activation function and pooling layer to obtain C feature maps x′ of the first convolution layer:

[0037] x′=p[f(x)]

[0038] Among them, f() represents the Relu activation function; p[] represents the maximum pooling layer, the padding of the maximum pooling layer is 0, and the convolution kernel size is 2; because the activation function does not change the size of the input data, the size of x′ is half of x.

[0039] The second, third, and fourth convolutional layers are used to convolve and downsample the output feature map of the previous layer to obtain the downsampled feature maps y′, z′, q′. The process is as follows:

[0040] (1) Assume that the second convolution layer contains 2C convolution kernels K′, and the convolution kernels of the second convolution layer are the same as the convolution kernels K of the first layer, both of which are 3×3×1, padded with 0, and the convolution step is 1; then use the 2C convolution kernels K′ of the second convolution layer to convolve the C feature maps x′ after downsampling of the first convolution layer, and obtain the result of the 2C convolutions of the second convolution layer, which is recorded as y:

[0041]

[0042] Then, the result y after 2C convolutions is input into the activation function and pooling layer of this layer, resulting in the 2C feature maps y′ of the second convolutional layer: y′-p[f(y)]. Here, f() represents the ReLU activation function; p[] represents the maximum pooling layer, with zero padding and a convolution kernel size of 2.

[0043] (2) Assume that the third convolution layer contains 4C convolution kernels K″, and the convolution kernel of the third convolution layer has the same size as the convolution kernel K of the first layer, both 3×3×1, padded with 0, and the convolution step is 1; then use the 4C convolution kernels K″ of the third convolution layer to convolve with the 2C feature maps y′ after downsampling of the second convolution layer, and obtain the result of the 4C convolutions of the third convolution layer, which is recorded as z:

[0044]

[0045] Then, the result z after 4C convolutions is input into the activation function and pooling layer of this layer, resulting in 4C feature maps z′ of the third convolutional layer: z′ = p[f(z)]. Here, f() represents the ReLU activation function; p[] represents the max pooling layer, with zero padding and a convolution kernel size of 2.

[0046] (3) Assume that the fourth convolutional layer contains 8C convolution kernels K″′, and the convolution kernel of the fourth convolutional layer has the same size as the convolution kernel K of the first layer, both 3×3×1, padded with 0, and the convolution step is 1; then use the 8C convolution kernels K″′ of the fourth convolutional layer to convolve with the 4C feature maps z′ after downsampling in the third convolutional layer, and obtain the result of the 8C convolutions in the fourth convolutional layer, which is recorded as q:

[0047]

[0048] The 8C convolution results q are then fed into the activation function and pooling layer of this layer, yielding the 8C feature maps q′ of the fourth convolutional layer: q′ = p[f(q)], where f() represents the ReLU activation function and p[] represents the max pooling layer, with zero padding and a kernel size of 2.

[0049] The fully connected layer is used to perform linear transformation on the 8C feature maps q′ after the downsampling process of the fourth convolutional layer, and obtain the data result after the linear transformation process of the fifth fully connected layer. Its expression is:

[0050]

[0051] Wherein, q′ is B×U, B is the number of samples Batchsize input for training in CNN-Train-Slices, B is a positive integer greater than 0, and U is the number of sample features; W represents the randomly initialized weight matrix of the fifth fully connected layer, W is U×O dimension, and O is the number of output classification categories; b represents the vector bias of the fifth fully connected layer, b is B×O dimension; in this embodiment, B is 5 and O is 3.

[0052] The data result after linear transformation of the fifth fully connected layer The data is B×O dimensional, and the fifth layer is fully connected after linear transformation. There is only one maximum value in .

[0053] Input CNN-Test-Slices into the trained CNN classification network, and perform linear transformation on the fifth fully connected layer. Slice classification is performed to determine whether to retain or filter the slices. The process is as follows:

[0054] (1) Determine the data result after linear transformation of the fifth fully connected layer corresponding to the slice S in the input CNN-Test-Slices The position label of the maximum value is j, 0≤j≤2.

[0055] (2) According to CNN-Label, the label of slice data of complex scenes without targets is 0, the label of slice data with targets is 1, and the label of other slice data is 2.

[0056] (3) If j = 0, the result of the slice S classification is considered to be a complex scene and does not contain target slice data, which will be filtered out; if j = 1, the result of the slice S classification is considered to be a target slice data, which will be retained; if j = 2, the result of the slice S classification is considered to be other slice data, which will be retained; the filtered slices will no longer participate in the subsequent steps, and the retained slice data will be recorded as Reserved Test-Slices.

[0057] Step 3: Preprocess the training set images and the retained set images after sliding window slicing; use the preprocessed training set data to train the YOLO V5 network to obtain a trained YOLO V5 network; then input the preprocessed retained set data into the trained YOLO V5 network to obtain a target recognition result slice dataset with the least false alarms, which is recorded as dataset A.

[0058] Specifically, in sub-step 3a, the training set images and the retention set images after sliding window slicing are preprocessed as follows:

[0059] Sub-step 3a.1, amplify and data augment the training set images and the holdout set images after sliding window slicing to obtain a sufficiently large dataset with high image quality; the amplified training set is denoted as dataset a, and the amplified holdout set is denoted as dataset b;

[0060] The training set and the holdout set are expanded and enhanced through rotation transformation, affine transformation, and mosaic methods to generate a sufficiently large dataset with high image quality to improve the speed and accuracy of network training in subsequent steps. Here, the training set and test set ratio is expanded to 8:1.

[0061] Sub-step 3a.2: Verify the annotation information of the data in dataset a and dataset b: Calculate the best recall rate of the default anchor boxes of the data in dataset a and dataset b; if the best recall rate is greater than or equal to 0.98, there is no need to update the anchor boxes; if the best recall rate is less than 0.98, it is necessary to recalculate the anchor boxes that match this dataset.

[0062] Using Auto Learning Bounding Box Anchors, which automatically analyzes input data using k-means clustering and genetic learning algorithms to obtain preset anchor boxes suitable for predicting the target bounding box in the input data. This box is automatically learned based on the training data.

[0063] Create a configuration file in YAML format. Its content includes the paths to datasets A and B, as well as a dictionary to save the number of target categories and corresponding category names of the SAR data detected in datasets A and B, and store them in the corresponding order.

[0064] Sub-step 3b, build the YOLO V5 network:

[0065] The YOLO V5 detection network consists of a baseline network, a neck network, and a head network. The baseline network and neck network models read SAR image pixel information, extract features of the SAR image at different network depths, and then input them into the head network. The head network outputs the position information, category, and confidence of the output box in the SAR image, and then performs maximum suppression through NMS to remove redundant output boxes.

[0066] This example uses the CSPDarknet53 and Focus structures as the baseline networks for the YOLO V5 network. The Focus structure primarily crops the input image through slicing operations: the original input image size is I*I*3, and after slicing and concatenation operations, it outputs a feature map of I / 2*I / 2*12. This is followed by a convolutional layer with C channels, which outputs a feature map of I / 2*I / 2*C. The CSPDarknet53 structure has two structures: the CSP1_X structure used in the Backbone network, and the CSP2_X structure used in the Neck network. The network uses a convolution size of 3*3 or 1*1, and the activation function uses the Sigmoid function.

[0067] The FPN+PAN structure is used as the Neck network of YOLO V5 to further improve feature diversity and robustness. The CSP2_X structure designed by CSPDarknet53 is also adopted to enhance the network's feature fusion capabilities. The entire YOLO V5 detection network structure does not have pooling layers or fully connected layers. The network downsampling is achieved by setting the convolution stride to 3.

[0068] During YOLO V5 detection network training, the loss function consists of three components: output box loss, confidence loss, and classification loss. The loss function measures the distance between YOLO V5's predictions and the expected information (i.e., labels). The closer the predictions are to the expected information, the smaller the loss function value. Training primarily involves three types of losses: rectangular box loss (i.e., output box, clc_loss), confidence loss (obj_loss), and classification loss (clc_loss).

[0069] Therefore, the loss function of the YOLO V5 detection network is defined as:

[0070] Loss=a*obj_loss+b*rect_loss+c*clc_loss

[0071] Among them, a, b, and c are the weights of the three losses, here a=0.4, b=0.3, c=0.3. CIOUloss is used to calculate the rectangular box loss, and BCE loss is used to calculate the confidence loss and classification loss. Let the predicted rectangular box be A, and the coordinates of its upper left corner and lower right corner are (x p1 ,y p1 )、(x p2 ,y p2 ), the label rectangle is B, and its upper left corner and lower right corner coordinates are (x l1 ,y l1 )、(x l2 ,y l2 );

[0072] CIOU loss adds the overlapping area, center point distance, and aspect ratio of the two rectangular boxes A and B into the calculation at the same time. The formula is:

[0073]

[0074]

[0075]

[0076] Where ρ is the distance between the centers of A and B, c is the diagonal length of the smallest enclosing rectangle of A and B, v is the similarity of the aspect ratios of A and B, and α is the influencing factor of v. Since the value range of the arctan function is 0-π / 2, the value range of v is 0-1. When the aspect ratios of A and B are equal, v is 0, and when the aspect ratios of A and B differ infinitely, v is 1. When the distance between A and B is infinite and the aspect ratio difference is infinite, Take -1, v take 1, α take 0.5, then CIOU loss takes -1-0.5=-1.5; when A and B completely overlap, If α is 0, v is 1, and α is 0, then CIOU loss is 1. Therefore, the value range of CIOU loss is -1.5-1.

[0077] The larger the IOU, that is, the larger the overlap area between A and B, the larger α is, and thus the greater the influence of v. Conversely, the smaller the IOU, that is, the smaller the overlap area between A and B, the smaller α is, and thus the smaller the influence of v. Therefore, during the optimization process: if the overlap area between A and B is small, the aspect ratio v has a smaller impact on the loss function, and in this case, the focus is on optimizing the distance between A and B. If the overlap area between A and B is large, the aspect ratio v has a greater impact on the loss function, and in this case, the focus is on optimizing the aspect ratio between A and B. Then, the calculation formula for CIOU loss is:

[0078] CIU loss=1-CIOU

[0079] An input image is divided into g*g grids. The network predicts three rectangular boxes near each grid. The prediction information of each output box includes the center coordinates, width, height, confidence, and classification probability. Therefore, the network outputs a total of 3*g*g prediction confidences ranging from 0 to 1, corresponding one to one with 3*g*g output boxes. The confidence of each output box represents the reliability of the output box. The larger the value, the more reliable the output box is, that is, the closer it is to the true minimum bounding box of the target. The single BCE loss formula for calculating the confidence loss of a single output box and the total confidence loss BCE loss object formula are as follows:

[0080] single BCE loss(z,x,y)

[0081] =-B(z,x,y)*logA(z,x,y)-(1-B(z,x,y))*log(1-A(z,x,y))

[0082]

[0083]

[0084] BCE loss object=α*lobj+(1-α)*lnoobj

[0085] Among them, x and y represent the center position of the rectangular box in the row and column grid of the image, and z represents the three nearby rectangular boxes predicted by the grid. Therefore, the value ranges of x, y, and z are 0≤x<g, 0≤y<g, and 0≤z<3 respectively;

[0086] The prediction information for each output frame contains N 0-1 classification probabilities, where N is the total number of categories. The network then predicts a total of 3*g*g*N classification probabilities, forming a prediction probability matrix. To reduce overfitting and increase training stability, a smoothing operation is usually performed on the one-hot coded labels. This is as follows:

[0087] label smooth =label*(1-β)+β / N

[0088] label is all the values in the one-hot code, β is the smoothing coefficient, ranging from 0 to 1, usually 0.1; let the label probability matrix be Bsmooth, the prediction probability matrix be A, and n represent the predicted category. Then the signal BCE loss and the total classification loss BCE loss class for each value in the matrix are calculated as follows:

[0089] signal BCE loss(z,x,y,n)=-Bsmooth(z,x,y,n)*logA(z,x,y,n)-(1-Bsmooth(z,x,y,n))*log(1-A(z,x,y,n))

[0090]

[0091] In sub-step 3b, the YOLO V5 network is trained for multiple rounds using dataset a to obtain a trained YOLO V5 network.

[0092] In sub-step 3c, dataset b is input into the trained YOLO V5 detection network, and the output box position information, category, and confidence of the SAR image are output. Then, NMS is performed to perform maximum suppression to remove redundant output boxes. The classification of the identified output box is determined to be consistent with the true target category. The IOU ratio between the output box and the labeled true box is calculated to determine whether the YOLO V5 detection network has detected the target. The confidence level is used to determine the detection accuracy of the output box for a single target. The specific process is as follows:

[0093] Sub-step 3c.1, first determine whether the prediction confidence of the output box is greater than the set confidence threshold, and whether the calculated IOU is greater than the IOU threshold. If both are exceeded, it is considered that there is a target in the output box, thereby obtaining the approximate location of the target; then, the prediction box with the target is filtered according to the non-maximum suppression algorithm, and the repeated output boxes corresponding to the same target are eliminated; finally, based on the classification probability of the filtered prediction box, the index corresponding to the maximum probability is taken as the classification index number of the target, thereby obtaining the category of the target. For output boxes that are determined to have no detected targets, it is a false alarm; for targets that are not detected in the real box, it is a missed detection;

[0094] In sub-step 3c.2, the location information, category, and confidence information of the output box of the target detected by the network are saved, and the slice images containing this type of output box are counted and saved to obtain the target recognition result slice dataset with the least false alarms, which is recorded as First Test-Slices.

[0095] Step 4: Perform the maximum confidence Max-score judgment on the output box and the total number of detected objects on the data set A. All the data in the data set A that meet the judgment conditions are used as the primary detection data set. The data in the primary detection data set is input into the CFAR detection algorithm for secondary detection to obtain the secondary detection data set.

[0096] The specific process is as follows:

[0097] Sub-step 4.1, counting the number of output boxes OBS in each slice image in the statistical data set A, and the maximum confidence MS in the output box of each slice image;

[0098] Sub-step 4.2: When the number of output boxes OBN of a slice image is greater than or equal to the output box number threshold OBN_threshold, and the maximum confidence MS in the output box is greater than or equal to the confidence threshold MS_threshold, then the slice image meets the judgment condition; all slice images that meet the judgment condition are used as a detection data set;

[0099] In sub-step 4.3, the data of the primary detection dataset is input into the CFAR detection algorithm for secondary detection to obtain a secondary detection dataset.

[0100] Step 5: Merge the secondary detection dataset data with the primary detection dataset data to obtain the final target detection result.

[0101] The effects of the present invention are further verified and illustrated through the following simulation experiments.

[0102] (1) Experimental conditions

[0103] The data used in the experiment consists of 15 large-scale (over 10,000 x 10,000) SAR images of vehicles. The SAR image data is divided into a training set and a test set, and then all vehicle target data in the training set and the test set are labeled with corresponding categories.

[0104] After sliding window slicing of the training set and test set respectively, the training sample set contains 6728 training samples and the test sample set contains 1163 test samples; among them, the training samples contain 183 vehicle target samples; and the test samples contain 128 vehicle target samples.

[0105] (2) Experimental content and results:

[0106] Use different methods to perform object detection on the test set.

[0107] Experiment 1: First, the YOLO V5 network is trained using the training set, and then the test samples of complex scenes are input into the trained YOLO V5 detection network;

[0108] The test samples of complex scenes and open scenes were tested respectively, and the results are as follows:

[0109] The accuracy of target detection in complex scenes is 90.0%, the recall rate is 13.7%, and the F1 is 23.8%. It takes 18.2161 seconds to detect one image.

[0110] The accuracy of target detection in open scenes is 84.8%, the recall rate is 62.2%, and the F1 is 71.8%. It takes 13.2258 seconds to detect one image.

[0111] Experiment 2: First, we train the YOLO V5 network using the training set. Then, we feed the complex scene test samples into the trained YOLO V5 detection network, and then feed them into the CFAR algorithm for detection.

[0112] The test samples of complex scenes and open scenes were tested respectively, and the results are as follows:

[0113] The accuracy of target detection in complex scenes is 35.8%, the recall rate is 75.6%, and the F1 is 48.6%. It takes 20.8815 seconds to detect one image.

[0114] The accuracy of target detection in open scenes is 67.9%, the recall rate is 84.4%, and the F1 is 75.2%. The time taken to detect one image is 15.5691s.

[0115] Experiment 3: We used a combined target detection method using the YOLO V5 detection network and the CFAR detection algorithm, and used average brightness pixel information and the aspect ratio and area information of the detection box to filter out some false alarms.

[0116] The test samples of complex scenes and open scenes were tested respectively, and the results are as follows:

[0117] The accuracy of target detection in complex scenes is 58.8%, the recall rate is 72.6%, and the F1 is 65%. The time taken to detect one image is 20.8075s.

[0118] The accuracy of target detection in open scenes is 84.4%, the recall rate is 84.4%, and the F1 is 84.4%. The time taken to detect one image is 15.3425s.

[0119] Experiment 4: Using the proposed method for SAR image target detection using combined deep learning and CFAR, complex scene test samples were first fed into a trained CNN classification network, then into a trained YOLO V5 detection network, and then into the CFAR algorithm for detection. False alarms were filtered using average pixel brightness information and the detection box aspect ratio and area information.

[0120] The test samples of complex scenes and open scenes were tested respectively, and the results are as follows:

[0121] The accuracy of target detection in complex scenes is 59.1%, the recall rate is 72.6%, and the F1 is 65.1%. The time taken to detect one image is 18.6976s.

[0122] The accuracy of target detection in open scenes is 92.7%, the recall rate is 84.4%, and the F1 is 88.4%. It takes 11.0404s to detect one image.

[0123] By comparing the complex scene detection results of Experiments 1 to 4 with the open scene detection results of Experiments 1 to 4, it can be seen that the detection accuracy and detection capability of the SAR image target detection method of the present invention combining deep learning and CFAR are far superior to other target detection methods in both complex and open scenes.

[0124] Although this specification has provided a detailed description of the present invention using general descriptions and specific embodiments, it will be apparent to those skilled in the art that modifications and improvements may be made based on the present invention. Therefore, such modifications and improvements, which do not depart from the spirit of the present invention, are intended to be within the scope of protection claimed herein.

Claims

1. A SAR image target detection method combining deep learning and CFAR, characterized in that: The following steps are involved: Step 1: Process N SAR images to obtain training sets and test sets; Perform sliding window slicing on the training set and the test set to obtain the training set after sliding window slicing and the test set after sliding window slicing respectively; Step 2: Process the training set and the test set after sliding window slicing; Use the processed training set data to train the CNN network to obtain a trained CNN network; then input the processed test set data into the trained CNN network to obtain the classification results of the processed test set data; The set of all empty scenes or slice images containing objects in the test set is recorded as the retention set; Step 3: Preprocess the training set images and the retained set images after sliding window slicing; Use the preprocessed training set data to train the YOLO V5 network to obtain a trained YOLO V5 network; Then, the pre-processed reserved set data is input into the trained YOLO V5 network to obtain the target recognition result slice dataset with the least false alarms, which is recorded as dataset A. Step 4: Perform the maximum confidence determination of the output box and the total number of detected targets on the data set A, and use all the data of the data set A that meet the determination conditions as a detection data set; The data in the primary detection data set is input into the CFAR detection algorithm for secondary detection to obtain a secondary detection data set; Step 5: Merge the secondary detection dataset data with the primary detection dataset data to obtain the final target detection result.

2. The SAR image target detection method based on combined deep learning and CFAR according to claim 1, characterized in that: Step 1, specifically, obtain the location and category of the target in N SAR images, and use the location and category of the target in the SAR image as labels; use a SAR image and the corresponding label as a set of data to obtain a sample set; randomly divide the sample set into a training set and a test set according to proportion; perform sliding window slicing on the SAR images of the training set and the test set to obtain a fixed-size slice and the label corresponding to the slice for each SAR image, and obtain the training set after sliding window slicing and the test set after sliding window slicing.

3. The SAR image target detection method based on combined deep learning and CFAR according to claim 1, characterized in that: In step 2, the training set and the test set after sliding window slicing are processed. Specifically, the slice images of the training set and the test set after sliding window slicing are marked: the labels of the slice images of complex scenes and without targets are recorded as 0, the labels of the slice images containing targets are recorded as 1, and the labels of other slice images are recorded as 2; Then, the slice images and corresponding labels in the training set after a sliding window slice are used as a set of data to obtain the CNN training set; The sliced images and corresponding labels in the test set after a sliding window slice are taken as a set of data to obtain the CNN test set.

4. The SAR image target detection method combining deep learning and CFAR according to claim 1, characterized in that: The CNN network specifically includes four convolutional layers and a fully connected layer connected in sequence; the convolutional layer is used to convolve and downsample the input slice image, and the fully connected layer is used to perform linear transformation processing on the features output by the fourth convolutional layer to obtain the classification result of the input slice image.

5. The SAR image target detection method combining deep learning and CFAR according to claim 1, characterized in that: Preprocess the training set images and the retention set images after sliding window slicing. Specifically, the sub-steps are as follows: Sub-step 3a.1, amplify and data augment the training set images and the holdout set images after sliding window slicing to obtain a sufficiently large dataset with high image quality; the amplified training set is denoted as dataset a, and the amplified holdout set is denoted as dataset b; In sub-step 3a.2, calculate the best recall of the default anchor boxes for dataset a and dataset b. If the best recall is greater than or equal to 0.98, there is no need to update the anchor boxes. If the best recall is less than 0.98, recalculate the anchor boxes that match this dataset.

6. The SAR image target detection method combining deep learning and CFAR according to claim 1, characterized in that: The sub-steps of step 4 are as follows: Sub-step 4.1, counting the number of output boxes OBS in each slice image in the statistical data set A, and the maximum confidence MS in the output box of each slice image; Sub-step 4.2: When the number of output boxes OBN of a slice image is greater than or equal to the output box number threshold OBN_threshold, and the maximum confidence MS in the output box is greater than or equal to the confidence threshold MS_threshold, then the slice image meets the judgment condition; all slice images that meet the judgment condition are used as a detection data set; In sub-step 4.3, the data of the primary detection dataset is input into the CFAR detection algorithm for secondary detection to obtain a secondary detection dataset.

Images