Infrared laser composite signal target identification and positioning method based on YOLO positioning thought

Abstract

The invention discloses an infrared laser composite signal target identification and positioning method based on a YOLO positioning thought, and relates to the technical field of signal identification and positioning. According to the method, the positioning thought of a YOLO series model is innovatively applied to the field of one-dimensional data signals, a network structure model is built according to the thought, a sample data set is built, target recognition and positioning of the one-dimensional data signals can be achieved, and experimental results show that the method can improve the recognition accuracy of the composite signals and improve the positioning accuracy of the composite signals. And the position of the target signal appearing in the segment of data can be determined, and target positioning is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of signal recognition and positioning, and in particular to an infrared laser composite signal target recognition and positioning method based on the YOLO positioning concept. Background Art

[0002] Convolutional neural network is a typical end-to-end deep learning method. It can directly learn the characteristics of the target through the original data and directly output the prediction results. The whole process does not require human intervention. In recent years, this method has also been gradually applied to the research of terminal-sensitive projectile detection signal target recognition. Yan Guangli et al. proposed three information fusion schemes and three network structures with different input methods based on the characteristics of terminal-sensitive projectile composite detection signals, and obtained the best scheme through testing. Wu Junan et al. used linear array laser radar as a detector and built a convolutional neural network model to classify and identify targets for laser radar range images. This method not only has high target recognition accuracy, but also can effectively identify armored targets in complex backgrounds. Zhou Shenghui et al. proposed a composite detection signal recognition method based on multi-channel convolutional neural network and extreme gradient boosting decision tree. This method can reduce the false detection rate and missed detection rate and has a good recognition effect.

[0003] Although the convolutional neural network model can be effectively used to realize the recognition of the composite detection signal of the terminal-sensitive projectile, there is still a problem: for the target signal, it is impossible to determine the location of the target, which makes the control system of the terminal-sensitive projectile unable to calculate the optimal detonation time. YOLO is a popular target detection algorithm that uses a single neural network model to simultaneously predict the location and category of multiple objects in the image, and uses the characteristics of global information and a single scale to achieve efficient and accurate target detection. Although the infrared image or range image of the terminal-sensitive projectile can be used to realize target recognition and target positioning through the YOLO model, there are problems such as many model parameters and large amount of calculation, which cannot meet the requirements of the missile-borne computing system for real-time target recognition. Summary of the invention

[0004] The purpose of the present invention is to provide an infrared laser composite signal target recognition and positioning method based on the YOLO positioning concept, and to use the YOLO model positioning concept to realize one-dimensional composite signal target recognition and positioning. The present invention uses the YOLO model to realize target positioning based on the one-dimensional detection signal of the laser radar and infrared detector to construct a YOLO-1D network model and sample set to realize the recognition and positioning of the target signal, and tests the performance of the model through experiments.

[0005] The technical solution to achieve the purpose of the present invention is: an infrared laser composite signal target recognition and positioning method based on the YOLO positioning concept, comprising the following steps:

[0006] Step 1. Use the infrared laser composite detection system to detect the target, collect n groups of infrared laser data, intercept the infrared laser data through a sliding window of length s, and establish an infrared laser composite signal database. At the same time, according to the position of the infrared laser signal change when the target appears, construct labels (X1, X2, Class) with category information and location information, and divide the infrared laser composite signal database into a training set and a test set in a ratio of 7:3.

[0007] Step 2: Build a SE-YOLO network, which includes 5 CBR modules, an SE attention mechanism module, a feature parameter size integration module, and a feature screening module.

[0008] Among them, the CBR module includes Conv1d, BatchNorm1d and ReLU activation functions; the SE attention mechanism module includes AdaptiveAvgPool1d, Linear, ReLU activation function, Sigmoid activation function; the feature parameter size integration module includes the Conv1d convolution layer; the core of the feature screening module is the non-maximum suppression algorithm.

[0009] Step 3: Use the training set to train the SE-YOLO network to obtain a trained SE-YOLO network model.

[0010] The infrared laser composite signal in the training set is input into the SE-YOLO network, and the features of the infrared laser composite signal are extracted layer by layer through three CBR modules to obtain the first eigenvector P1; P1 is input into the SE attention mechanism module to change the importance of each feature channel, so that the neural network focuses on the channel with large weight value, and the second eigenvector P2 is obtained; the remaining two CBR modules are used to increase the depth of the SE-YOLO network model, and P2 is further extracted to obtain the third eigenvector P3; P3 is information fused through the feature parameter size integration module to obtain the fourth eigenvector V with a size of 8×N; V is screened by the feature screening module to obtain the predicted value with the highest confidence (X'1, X'2, Clss'); the predicted value and its corresponding label are used for error calculation to update the parameters of the SE-YOLO network to obtain the trained SE-YOLO network model.

[0011] Step 4: Input the test set into the SE-YOLO network model to test the accuracy and positioning precision of the SE-YOLO network model.

[0012] Compared with the prior art, the advantages of the present invention are:

[0013] 1) The present invention applies the positioning concept of the YOLO series models to one-dimensional data signals, which can not only realize the recognition of infrared laser composite signals, but also realize the positioning of target signals.

[0014] 2) Introduce the SE attention mechanism module into the overall model, enabling the model to automatically learn which feature channels are more important, thereby enhancing the model's feature expression ability and contributing to improving the classification or recognition performance of the model. Description of the Drawings

[0015] Figure 1 It is a flowchart of the infrared laser composite signal target recognition and positioning method based on the YOLO positioning idea of the present invention.

[0016] Figure 2 It is a framework diagram of the overall network model. Detailed Embodiments

[0017] To make the objectives, technical solutions, and advantages of the present invention clearer, the following further describes the embodiments of the present invention in detail.

[0018] Combined with Figure 1-2 A method for infrared laser composite signal target recognition and positioning based on the YOLO positioning idea described in the present invention includes the following steps:

[0019] Step 1: Use an infrared laser composite detection system to detect the target, collect n groups of infrared laser data, where 1000 < n < 3000, intercept the infrared laser data using a sliding window with a length of s, s = 1024, establish an infrared laser composite signal database, and simultaneously construct labels with presence category information and position information according to the position where the infrared laser signal changes when the target appears. The label format is (X1, X2, Class), where X1 is the position where the infrared laser composite signal target appears, X2 is the position where the infrared laser composite signal target disappears, and Class is the true category of the infrared laser composite signal. Then divide the infrared laser composite signal database into a training set and a test set according to a ratio of 7:3.

[0020] Step 2: Construct an SE-YOLO network, including 5 CBR modules, an SE attention mechanism module, a feature parameter size integration module, and a feature screening module, specifically as follows:

[0021] The CBR module includes Conv1d, BatchNorm1d, and a ReLU activation function, which are used to enhance the network's expression ability, improve the stability and efficiency of training at the same time, and stacking multiple CBR modules can increase the network depth, thereby extracting the features of the infrared laser composite signal layer by layer.

[0022] The SE attention mechanism module includes AdaptiveAvgPool1d, Linear, ReLU activation function, and Sigmoid activation function. The SE attention mechanism module adds an attention mechanism in the channel dimension, changes the importance of each feature channel, and makes the neural network focus on certain channels with large weight values.

[0023] The feature parameter size integration module includes a Conv1d convolution layer, which integrates the features of the extracted infrared laser composite signal to extract higher-level features, and integrates the higher-level feature sizes into 8×N, where N represents the number of columns of the output feature.

[0024] The core of the feature screening module is the non-maximum suppression algorithm, which is used to screen the features extracted by the SE-YOLO network model in the feature extraction stage to obtain the label prediction value with the highest confidence.

[0025] Step 3: Use the training set to train the SE-YOLO network to obtain the trained SE-YOLO network model, as follows:

[0026] Step 3-1: Input the infrared laser composite signal in the training set into the SE-YOLO network, and extract the features of the infrared laser composite signal layer by layer through three CBR modules to obtain the first feature vector P1:

[0027] P1=CBR 32 (CBR 16 (CBR 8 (D))) (1)

[0028] Where D is the input infrared laser composite signal, CBR t (·) is the feature extraction function with output channel size t, t = 8, 16, 32.

[0029] Step 3-2: Input P1 into the SE attention mechanism module to change the importance of each feature channel, so that the neural network focuses on the channels with large weight values, and obtain the second feature vector P2:

[0030] P2=SE(P1) (2)

[0031] Step 3-3, use the remaining two CBR modules to increase the depth of the SE-YOLO network, and further extract P2 to obtain the third feature vector P3:

[0032] P3=CBR 32 (CBR 32 (P2)) (3)

[0033] Step 3-4: Use the feature parameter size integration module to fuse information of P3 and obtain the fourth feature vector V, v with a size of 8×N. ij is the eigenvalue contained in V:

[0034] V=C(P3) (4)

[0035] V=[v 11 , ..., v 1j , ..., v 1N ;

[0036] v 21 , ..., v 2j , ..., v 2N ; ...

[0038] v i1 , ..., v ij , ..., v iN ; ...

[0040] v 81 , ..., v 8j , ..., v 8N ] (5)

[0041] Among them, C(·) represents the Conv1d function, V represents the fourth eigenvector, which contains 8 channels, each of which represents a different meaning: v 1j Represents the position where the first predicted box target appears, v 2j represents the position where the first predicted box target disappears, v 3j Represents the confidence of the first predicted box and the true box, v 4j Represents the position where the second prediction box target appears, v 5j Represents the position where the second prediction box target disappears, v 6j Represents the confidence of the second predicted box and the true box, v 7j represents the probability of no target signal, v 8j Represents the probability of having a target signal, channel number i = 1, 2, 3...8, j = 1, 2,...N.

[0042] Step 3-5: Use the feature screening module to screen the fourth feature vector V to obtain the prediction value (X'1, X'2, Class') with the highest confidence.

[0043] Step 3-6: Calculate the error between the predicted value of the SE-YOLO network model and its corresponding label to update the parameters of the SE-YOLO network and obtain the trained SE-YOLO network model, as follows:

[0044] Step 3-6-1, construct the position loss function L1, category loss function L2 and confidence loss function L3 to approximate the predicted probability distribution and true probability distribution of the SE-YOLO network:

[0045]

[0046] L2=L Cls (Class,Class') (7)

[0047] L3=L Confi (IOU((X'1,X'2),(X1,X2)),1) (8)

[0048] Among them, x m Represents the location tag of the infrared laser composite signal target, x' m represents the predicted position value of the infrared laser composite signal target, m represents the current prediction frame number, L Cls is the cross entropy loss function between the predicted category and the true category, L Confi It is the cross entropy loss function between the IOU value of the predicted box and the true box and 1; (X1, X2) represents the true position interval, (X'1, X'2) represents the predicted position interval, Class represents the true category of the infrared laser signal, Class' represents the predicted category of the infrared laser composite signal, and IOU represents the intersection calculation function.

[0049] Step 3-6-2, construct the final loss function L:

[0050] L=λ1*L1+λ2*L Cls +λ3*L Confi (9)

[0051] Among them, λ1, λ2, λ3 are the balance parameters of the loss function.

[0052] Step 3-6-3: Get the trained SE-YOLO network model.

[0053] Step 4: Input the test set into the trained SE-YOLO network model to test the accuracy and positioning precision of the trained SE-YOLO network model, as follows:

[0054] For the classification problem of infrared laser composite signals, the evaluation indicators use accuracy, precision, recall rate, and F1 value to measure the classification performance of the model; for the target positioning problem of infrared laser composite signals, since positioning is only meaningful on signals with targets, only on correctly classified signals with targets, the mean of the predicted position interval (X'1, X'2) is calculated. Whether it is within the range of the real position interval (X1, X2) is used as the criterion for successful positioning, and the evaluation index is calculated accordingly. The specific principle is shown in the following formula:

[0055]

[0056] Among them, P loc represents the positioning accuracy rate, M represents the number of correctly predicted positive samples, and k represents the number of samples in the test set.

[0057] The meaning of the numerator is as follows: If the predicted mean value of the target signal position is within the range of the true value interval, it is recorded as 1; otherwise, it is recorded as 0. The positioning accuracy rate is calculated by calculating the number of successful positionings.

[0058] Example 1

[0059] An infrared laser composite signal target recognition and positioning method based on the YOLO positioning idea of the present invention includes the following steps:

[0060] Step 1: Use an infrared laser composite detection system to detect the target, collect n groups of infrared laser data, 1000 < n < 3000, intercept the infrared laser data using a sliding window with a length of s, s = 1024, establish an infrared laser composite signal database, which is divided into 2 categories, namely target signal samples and non-target signal samples, with a total of 3500 groups. At the same time, construct labels with category information and position information according to the position where the infrared laser signal changes when the target appears. The label format is (X1, X2, Class), where X1 is the position where the infrared laser composite signal target appears, X2 is the position where the infrared laser composite signal target disappears, and Class is the true category of the infrared laser composite signal. The infrared laser composite signal database is divided into a training set and a test set according to a ratio of 7:3. There are 2450 groups of infrared laser composite signals in the training set and 1050 groups of infrared laser composite signals in the test set.

[0061] Step 2: Construct an SE-YOLO network, including 5 CBR modules, an SE attention mechanism module, a feature parameter size integration module, and a feature screening module, as follows:

[0062] The CBR module includes Conv1d, BatchNorm1d, and ReLU activation functions, which are used to enhance the expression ability of the network, improve the stability and efficiency of training at the same time. Stacking multiple CBR modules can increase the depth of the network, thereby extracting the features of the infrared laser composite signal layer by layer.

[0063] The SE attention mechanism module includes AdaptiveAvgPool1d, Linear, ReLU activation function, and Sigmoid activation function. The SE attention mechanism module adds an attention mechanism in the channel dimension, changes the importance of each feature channel, and makes the neural network focus on certain channels with large weight values.

[0064] The feature parameter size integration module includes a Conv1d convolution layer, which integrates the features of the extracted infrared laser composite signal to extract higher-level features, and integrates the higher-level feature sizes into 8×N, where N represents the number of columns of the output feature.

[0065] The core of the feature screening module is the non-maximum suppression algorithm, which is used to screen the features extracted by the SE-YOLO network model in the feature extraction stage to obtain the label prediction value with the highest confidence.

[0066] Step 3: Use the training set to train the SE-YOLO network to obtain the trained SE-YOLO network model, as follows:

[0067] Step 3-1: Input the infrared laser composite signal in the training set into the SE-YOLO network, and extract the features of the infrared laser composite signal layer by layer through three CBR modules to obtain the first feature vector P1.

[0068] P1=CBR 32 (CBR 16 (CBR 8 (D))) (1)

[0069] Where D is the input infrared laser composite signal, CBR t (·) is the feature extraction function with output channel size t, t = 8, 16, 32.

[0070] Step 3-2: Input P1 into the SE attention mechanism module to change the importance of each feature channel, so that the neural network focuses on the channels with large weight values, and obtain the second feature vector P2.

[0071] P2=SE(P1) (2)

[0072] Step 3-3: Use the remaining two CBR modules to increase the depth of the SE-YOLO network, and further extract P2 to obtain the third feature vector P3.

[0073] P3=CBR 32 (CBR 32 (P2)) (3)

[0074] Step 3-4: Use the feature parameter size integration module to fuse information of P3 and obtain the fourth feature vector V, v with a size of 8×N. ij is the eigenvalue contained in V.

[0075] V=C(P3) (4)

[0076] V=[v 11 , ..., v 1j , ..., v 1N ;

[0077] v 21 , ..., v 2j , ..., v 2N ; ...

[0079] v i1 , ..., v ij , ..., v iN ; ...

[0081] v 81 , ..., v 8j , ..., v 8N ] (5)

[0082] Among them, C(·) represents the Conv1d function, V represents the fourth eigenvector, which contains 8 channels, each of which represents a different meaning: v 1j Represents the position where the first predicted box target appears, v 2j represents the position where the first predicted box target disappears, v 3j Represents the confidence of the first predicted box and the true box, v 4j Represents the position where the second prediction box target appears, v 5j Represents the position where the second prediction box target disappears, v 6j Represents the confidence of the second predicted box and the true box, v 7j represents the probability of no target signal, v 8j Represents the probability of having a target signal, channel number i = 1, 2, 3...8, j = 1, 2,...N.

[0083] Step 3-5: Use the feature screening module to screen the fourth feature vector V to obtain the prediction value (X'1, X'2, Class') with the highest confidence.

[0084] Step 3-6: Calculate the error between the predicted value of the SE-YOLO network model and its corresponding label to update the parameters of the SE-YOLO network and obtain the trained SE-YOLO network model, as follows:

[0085] Step 3-6-1, construct the position loss function L1, category loss function L2 and confidence loss function L3 to approximate the predicted probability distribution and true probability distribution of the SE-YOLO network:

[0086]

[0087] L2=L Cls (Class,Class') (7)

[0088] L3=L Confi (IOU((X'1,X'2),(X1,X2)),1) (8)

[0089] Among them, x m Represents the location tag of the infrared laser composite signal target, x' m represents the predicted position value of the infrared laser composite signal target, m represents the current prediction frame number, L CLs is the cross entropy loss function between the predicted category and the true category, L Confi It is the cross entropy loss function between the IOU value of the predicted box and the true box and 1; (X1, X2) represents the true position interval, (X'1, X'2) represents the predicted position interval, Class represents the true category of the infrared laser signal, Class' represents the predicted category of the infrared laser composite signal, and IOU represents the intersection calculation function.

[0090] Step 3-6-2, construct the final loss function L:

[0091] L=λ1*L1+λ2*L Cls +λ3*L Confi (9)

[0092] Among them, λ1, λ2, λ3 are the balance parameters of the loss function.

[0093] Step 3-6-3: Get the trained SE-YOLO network model.

[0094] Step 4: Input the test set into the trained SE-YOLO network model to test the accuracy and positioning precision of the trained SE-YOLO network model, as follows:

[0095] For the classification problem of infrared laser composite signals, the evaluation indicators use accuracy, precision, recall rate, and F1 value to measure the classification performance of the model; for the target positioning problem of infrared laser composite signals, since positioning is only meaningful on signals with targets, only on correctly classified signals with targets, the mean of the predicted position interval (X'1, X'2) is calculated. Whether it is within the range of the real position interval (X1, X2) is used as the standard for successful positioning, and the evaluation index is calculated based on this. The specific principle is shown in the following formula:

[0096]

[0097] Among them, P loc represents the positioning accuracy, M represents the number of correct positive sample predictions, and k represents the number of test set samples.

[0098] The meaning of the numerator is: if the predicted mean value of the target signal position is within the true value range, it is recorded as 1, otherwise it is recorded as 0; the positioning accuracy is calculated by counting the number of successful positioning.

[0099] The proposed method uses Python programming language and pytorch framework language to build a network framework on an Nvidia 1050Ti GPU host for related experiments, and the learning rate of the training network is set to 1e -4 ,In order to speed up network training, the optimization strategy selects Momentum, uses the Adam optimization algorithm, and the batch-size is set to 16. In the experiment, in order to prevent the ,network overfitting problem, the Dropout value is set to 0.5, the λ1 value is set to 1, the λ2 value is set to 0.5, and the λ3 value is set to 0.5.

[0100] Table 1 Comparative test

[0101]

[0102] In order to test the real situation of infrared laser composite signals as much as possible, the experiments in the present invention did not process the original data signals in any way. In order to demonstrate the superior performance of the algorithm of the present invention, a comparative experiment between the classical algorithm and the algorithm of the present invention was also carried out. The comparative experimental results are shown in Table 1. It can be seen from the experimental results that the algorithm of the present invention has a great improvement in the recognition rate and can achieve target positioning, and the positioning accuracy can reach 85.03%.

Claims

1. A method for infrared laser composite signal target recognition and positioning based on YOLO positioning concept, characterized in that: It includes the following steps: Step 1: Detect the target using an infrared laser composite detection system, collect n groups of infrared laser data, intercept the infrared laser data through a sliding window with a length of s, establish an infrared laser composite signal database, and at the same time construct labels (X1, X2, Class) with presence category information and position information according to the position where the infrared laser signal changes when the target appears, and divide the infrared laser composite signal database into a training set and a test set according to a ratio of 7:3; Step 2: Construct an SE-YOLO network, which includes 5 CBR modules, an SE attention mechanism module, a feature parameter size integration module, and a feature screening module; Among them, the CBR module includes Conv1d, BatchNorm1d, and ReLU activation function; the SE attention mechanism module includes AdaptiveAvgPool1d, Linear, ReLU activation function, Sigmoid activation function; the feature parameter size integration module includes a Conv1d convolutional layer; the core of the feature screening module is the non-maximum suppression algorithm; Step 3: Train the SE-YOLO network using the training set to obtain a trained SE-YOLO network model; Input the infrared laser composite signal in the training set into the SE-YOLO network, extract the features of the infrared laser composite signal layer by layer through 3 CBR modules to obtain the first feature vector P1; input P1 into the SE attention mechanism module to change the importance of each feature channel, so that the neural network focuses on the channels with large weight values to obtain the second feature vector P2; use the remaining 2 CBR modules to increase the depth of the SE-YOLO network model, and further extract P2 to obtain the third feature vector P3; perform information fusion on P3 through the feature parameter size integration module to obtain a fourth feature vector V with a size of 8×N; use the feature screening module to screen V to obtain the prediction value (X’1, X’2, Class’) with the highest confidence; calculate the error between the prediction value and its corresponding label to update the parameters of the SE-YOLO network and obtain a trained SE-YOLO network model; Step 4: Input the test set into the SE-YOLO network model to test the accuracy and positioning accuracy of the SE-YOLO network model.

2. According to the infrared laser composite signal target recognition and positioning method based on the YOLO positioning concept as described in claim 1, it is characterized in that: In Step 1, 1000 < n < 3000, s = 1024, the label format is (X1, X2, Class), X1 is the position where the infrared laser composite signal target appears, X2 is the position where the infrared laser composite signal target disappears, and Class is the true category of the infrared laser composite signal.

3. According to claim 2, a method for infrared laser composite signal target recognition and positioning based on YOLO positioning concept, characterized in that: In Step 2, construct an SE-YOLO network, including 5 CBR modules, an SE attention mechanism module, a feature parameter size integration module, and a feature screening module, specifically as follows: The CBR module includes Conv1d, BatchNorm1d and ReLU activation functions, which are used to enhance the network's expressiveness and improve the stability and efficiency of training. The superposition of multiple CBR modules can increase the network depth, thereby extracting the characteristics of infrared laser composite signals layer by layer. The SE attention mechanism module includes AdaptiveAvgPool1d, Linear, ReLU activation function, and Sigmoid activation function. The SE attention mechanism module adds an attention mechanism in the channel dimension, changes the importance of each feature channel, and makes the neural network focus on certain channels with large weight values; The feature parameter size integration module includes a Conv1d convolution layer, and the feature parameter integration module integrates information of the extracted features of the infrared laser composite signal to extract higher-level features, and integrates the higher-level feature sizes into 8×N, where N represents the number of columns of the output feature; The core of the feature screening module is the non-maximum suppression algorithm, which is used to screen the features extracted by the SE-YOLO network model in the feature extraction stage to obtain the label prediction value with the highest confidence.

4. The infrared laser composite signal target recognition and positioning method based on the YOLO positioning concept according to claim 3 is characterized in that: In step 3, the SE-YOLO network is trained using the training set to obtain a trained SE-YOLO network model, as follows: Step 3-1, input the infrared laser composite signal in the training set into the SE-YOLO network, and extract the features of the infrared laser composite signal layer by layer through three CBR modules to obtain the first feature vector P1; P1=CBR 32 (CBR 16 (CBR 8 (D))) (1) Where D is the input infrared laser composite signal, CBR t (·) is the feature extraction function with output channel size t, t = 8, 16, 32; Step 3-2, input P1 into the SE attention mechanism module, change the importance of each feature channel, make the neural network focus on the channel with large weight value, and obtain the second feature vector P2; P2=SE(P1) (2) Step 3-3, use the remaining two CBR modules to increase the depth of the SE-YOLO network, and further extract P2 to obtain the third feature vector P3; P3=CBR 32 (CBR 32 (P2)) (3) Step 3-4: Use the feature parameter size integration module to fuse information of P3 and obtain the fourth feature vector V, v with a size of 8×N. ij is the eigenvalue contained in V; V=C(P3) (4) V=[v 11 ,...,v 1j ,...,v 1N ; v 21 ,...,v 2j ,...,v 2N ; ... v i1 ,...,v ij ,...,v iN ; ... v 81 ,...,v 8j ,...,v 8N ] (5) Among them, C(·) represents the Conv1d function, V represents the fourth eigenvector, which contains 8 channels, each of which represents a different meaning: v 1j Represents the position where the first predicted box target appears, v 2j represents the position where the first predicted box target disappears, v 3j Represents the confidence of the first predicted box and the true box, v 4j Represents the position where the second prediction box target appears, v 5j Represents the position where the second prediction box target disappears, v 6j Represents the confidence of the second predicted box and the true box, v 7j represents the probability of no target signal, v 8j Represents the probability of the target signal; channel number i = 1, 2, 3...8, j = 1, 2,...N; Step 3-5: Use the feature screening module to screen the fourth feature vector V to obtain the prediction value (X'1, X'2, Class') with the highest confidence; Step 3-6: perform error calculation between the predicted value of the SE-YOLO network model and its corresponding label to update the parameters of the SE-YOLO network and obtain a trained SE-YOLO network model.

5. The infrared laser composite signal target recognition and positioning method based on the YOLO positioning concept according to claim 3 is characterized in that: In step 3-6, the error between the predicted value of the SE-YOLO network model and its corresponding label is calculated to update the parameters of the SE-YOLO network and obtain the trained SE-YOLO network model, as follows: Step 3-6-1, construct the position loss function L1, category loss function L2 and confidence loss function L3 to approximate the predicted probability distribution and true probability distribution of the SE-YOLO network: L2=L Cls (Class,Class') (7) L3=L Confi (IOU((X'1,X'2),(X1,X2)),1) (8) Among them, x m Represents the location tag of the infrared laser composite signal target, x' m represents the predicted position value of the infrared laser composite signal target, m represents the current prediction frame number, L Cls is the cross entropy loss function between the predicted category and the true category, L Confi is the cross entropy loss function between the IOU value of the predicted box and the real box and 1; (X1, X2) represents the real position interval, (X'1, X'2) represents the predicted position interval, Class represents the real category of the infrared laser signal, Class' represents the predicted category of the infrared laser composite signal, and IOU represents the intersection calculation function; Step 3-6-2, construct the final loss function L: L=λ1*L1+λ2*L Cls +λ3*L Confi (9) Among them, λ1, λ2, λ3 are the balance parameters of the loss function; Step 3-6-3: Get the trained SE-YOLO network model.

6. The infrared laser composite signal target recognition and positioning method based on the YOLO positioning concept according to claim 4 is characterized in that: In step 4, the test set is input into the trained SE-YOLO network model to test the accuracy and positioning precision of the trained SE-YOLO network model, as follows: For the classification problem of infrared laser composite signals, the evaluation indicators use accuracy, precision, recall rate, and F1 value to measure the classification performance of the model; for the target positioning problem of infrared laser composite signals, since positioning is only meaningful on signals with targets, only on correctly classified signals with targets, the mean of the predicted position interval (X'1, X'2) is calculated. Whether it is within the range of the real position interval (X1, X2) is used as the standard for successful positioning, and the evaluation index is calculated based on this. The specific principle is shown in the following formula: Among them, P loc represents the positioning accuracy, M represents the number of correct positive sample predictions, and k represents the number of test set samples; The meaning of the numerator is: if the predicted mean value of the target signal position is within the true value range, it is recorded as 1, otherwise it is recorded as 0; the positioning accuracy is calculated by counting the number of successful positioning.

Images