Sea surface radar target classification method based on convolution and channel-by-channel rectification function

Abstract

The invention provides a sea surface radar target classification method based on convolution and a channel-by-channel rectification function. The method comprises the following steps: processing original echoes of a sea detection radar to obtain radar sea surface target detection trace data; radar sea surface target classification marking is completed; constructing a radar-to-sea detection target classification data set; training the training data set by using a sea surface radar target convolutional neural classification network based on convolution and a channel-by-channel rectification function; and according to the recorded recall rate of the classifier test and the average precision mean value, the classifier with the optimal performance index in all iterations is obtained and is used for classifying the sea radar target. According to the invention, the ability of the classification network to adapt to radar targets of different scales / shapes and different clutter backgrounds is enhanced through the channel-by-channel rectification function, so that the classification ability of the classification network to radar sea surface weak targets is improved.

Description

Technical Field

[0001] This invention belongs to the field of radar target classification, specifically a method for classifying sea surface radar targets based on convolution and channel-by-channel rectified functions. Background Technology

[0002] Approximately 70% of the Earth's surface is covered by oceans. As a large country with a long coastline, China bears significant pressure in terms of maritime surveillance and defense. Therefore, building a robust and comprehensive maritime surveillance system is a crucial measure for safeguarding national maritime security. Within this system, radar, as an important remote sensing sensor, plays a vital role in maritime surveillance due to its unique advantages in target detection and classification.

[0003] The excellent classification performance of deep convolutional neural networks has led to their widespread application in radar target classification. However, traditional deep convolutional neural networks use fixed linear rectified functions as activation functions, which are difficult to adapt to complex and ever-changing targets and detection backgrounds. Summary of the Invention

[0004] This invention proposes a sea surface radar target classification method based on convolution and channel-wise rectified functions to solve the problem that traditional deep convolutional neural classification networks, which use fixed linear rectified functions as activation functions, cannot adapt to complex and ever-changing targets and detection backgrounds.

[0005] The technical solution to achieve the objective of this invention is: a sea surface radar target classification method based on convolution and channel-by-channel rectified functions, comprising the following steps:

[0006] Step 1: Process the raw echo of the sea surface detection radar to obtain radar target detection point data;

[0007] Step 2: Based on the radar sea surface target detection point data, target information is matched during the data acquisition process. Targets in the radar video data are labeled with rectangular marker boxes to complete the radar sea surface target classification labeling.

[0008] A local slice of radar target of a set size is extracted from the radar surface target detection point data; the radar target should usually be located in the center of the slice image; a radar sea detection target classification dataset is constructed using target category labeling information and radar target slice images;

[0009] Step 3: Based on the application scenarios and requirements of radar target classification for sea detection, and referring to the principles of completeness and balance in dataset construction, samples are extracted from the radar target classification dataset for sea detection and training datasets and test datasets are constructed.

[0010] Step 4: Train the training dataset using a sea surface radar target convolutional neural classification network based on convolution and channel-wise rectified functions;

[0011] Step 5: Test the trained classification network and network weights using radar target classification test set samples, and record the recall rate and mean precision.

[0012] Step 6: Check if the number of training iterations for the sea surface radar target classification method based on convolution and channel-wise rectification function has reached the maximum value. If the number of iterations has not reached the maximum value, adjust the network weights and jump to step 4; if the number of iterations has reached the set maximum value, proceed to step 7.

[0013] Step 7: Based on the recorded recall and mean precision of the classifier tests, determine the classifier with the best performance across all iterations for classifying marine radar targets.

[0014] Compared with the prior art, the significant advantages of this invention are:

[0015] This invention introduces a channel-wise rectified function to improve the activation function in a deep convolutional neural classification network. It designs a sea surface radar target classification network based on convolution and channel-wise rectified functions. By using the channel-wise rectified function, the network's ability to adapt to radar targets of different sizes and shapes and different clutter backgrounds is enhanced, thereby improving the network's ability to classify weak radar targets on the sea surface.

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the network structure of the present invention.

[0018] Figure 2 This is a schematic diagram of the principle of the channel-by-channel rectification function in this invention. Detailed Implementation

[0019] A sea surface radar target classification method based on convolution and channel-wise rectified functions, the specific steps of which are as follows:

[0020] Step 1: Process the raw echo of the sea surface detection radar to obtain radar target detection point data;

[0021] In a further embodiment, the method for processing the raw echo of the maritime detection radar mainly includes:

[0022] The process involves IQ demodulation, pulse compression, mode extraction, and target detection. After mode extraction, the raw echo from the maritime surveillance radar is converted into video data. By performing target detection on the radar video data, radar target detection point data can be obtained.

[0023] Step 2: Based on the radar surface target detection point data and the target information acquired during the data acquisition process, targets in the radar video data are labeled using rectangular bounding boxes to classify them, completing the radar surface target classification labeling. The radar surface target classification labeling information includes the target's bounding box size, position, and target category name. Based on the target category name and the bounding box position, a local radar target slice of size J1×J2 is extracted from the 2D radar video data; the radar target should typically be located at the center of the slice image. A radar sea detection target classification dataset is constructed using the target category labeling information and the radar target slice images to support the development of a deep learning target classification network model for sea detection radar.

[0024] Step 3: Based on the application scenarios and requirements of radar target classification for maritime detection, and referring to the principles of completeness and balance in dataset construction, samples are extracted from the radar target classification dataset for maritime detection to construct training and test datasets. The ratio of the number of training and test datasets is usually between 6:4 and 8:2. The completeness principle means that, for the application scenario, the dataset should cover all influencing factors of the object to be processed. The balance principle means that, for the application scenario, the distribution of the number of samples in the dataset should follow a reasonable proportion.

[0025] Step 4: Train the training dataset using a sea surface radar target convolutional neural classification network based on convolution and channel-wise rectified functions. This invention designs a sea surface radar target convolutional neural classification network based on convolution and channel-wise rectified functions; the size of the input sample image is J1×J2, and the specific size of the input sample image is determined according to the actual application; the main structure and parameters of the sea surface radar target convolutional neural classification network based on convolution and channel-wise rectified functions are as follows: Figure 1 As shown. This network consists of 10 layers, and its structure is described below:

[0026] 1) The first layer of the network consists of convolutional layers, a linear rectified function (RCF) for each channel, and a max pooling layer. The parameters of the convolutional layers include a kernel size of N1×N1, 8 channels, a stride of S1, and zero-padding size of P1. The parameters of the linear rectified function include 8 channels and a two-dimensional structure. The parameters of the max pooling layer include a kernel size of M1×M1 and a stride of S2. The values ​​of the main network parameters, such as kernel size, stride, and zero-padding size, for the first layer of the network are determined based on the actual application.

[0027] The schematic diagram of the channel-by-channel rectification function is as follows: Figure 2As shown, the input feature map data x has dimensions C×H×W, where C represents that the feature map data has C channels, H represents that the feature map data has H rows, and W represents that the feature map data has W columns. The input feature map data x is first processed by average pooling, resulting in C parameters. The output of average pooling is then input into a fully connected layer for the first fully connected processing, resulting in C / G parameters. The output of the fully connected layer is then processed by a linear rectified function. The output of the linear rectified function is then input into a fully connected layer for the second fully connected processing, resulting in 2KC parameters, where K is the number of functions. The output of the second fully connected layer is normalized using a Sigmoid parameter, constraining the output parameters to the closed interval [-1, 1]. The 2KC parameters Ω output after normalization are:

[0028] (1)

[0029] Where the superscript T denotes transpose, the values ​​of Δpk,c(x) and Δqk,c(x) depend on the feature map data x, the variable k takes the value of all positive integers in the closed interval [1, K], and the variable c takes the value of all positive integers in the closed interval [1, C]. Based on Δpk,c(x) and Δqk,c(x), the activation function parameters pk,c and qk,c are calculated using the following two expressions:

[0030] (2)

[0031] (3)

[0032] Where the parameters λp and λq are hyperparameters, u k and v k These are the initial values ​​for the activation function parameters; Figure 2 In the expression p1:K,1:C, it represents KC parameters pk,c (where the variable k takes the value of all positive integers within the closed interval [1, K], and the variable c takes the value of all positive integers within the closed interval [1, C]), and q 1:K,1:C Let KC be the parameters qk,c (where k takes the value of all positive integers within the closed interval [1, K], and c takes the value of all positive integers within the closed interval [1, C]); the final output of the hyperparameter function η in the channel-by-channel rectification function is:

[0033] (4)

[0034] Where x is the input feature data, xc is the feature data of the c-th dimension, pk,c and qk,c are the activation function parameters; the final output of the channel-wise rectified function y is:

[0035] (5)

[0036] Where max represents taking the maximum value among all variables;

[0037] 2) The second layer of the network consists of convolutional layers, a linear rectified function (RCF) for each channel, and a max pooling layer. The parameters of the convolutional layers include a kernel size of N²×N², 16 channels, a stride of S³, and zero-padding size of P². The parameters of the linear rectified function include 16 channels and a two-dimensional structure. The parameters of the max pooling layer include a kernel size of M¹×M¹ and a stride of S⁴. The values ​​of the main network parameters, such as kernel size, stride, and zero-padding size, for the second layer are determined based on the specific application.

[0038] 3) The third layer of the network consists of convolutional layers, a linear rectified function (RCF) for each channel, and a max pooling layer. The parameters of the convolutional layers include a kernel size of N³×N³, 32 channels, a stride of S⁵, and zero-padding size of P³. The parameters of the linear rectified function include 32 channels and a two-dimensional structure. The parameters of the max pooling layer include a kernel size of M³×M³ and a stride of S⁶. The values ​​of the main network parameters, such as kernel size, stride, and zero-padding size, for the third layer are determined based on the specific application.

[0039] 4) The fourth layer of the network consists of convolutional layers and a channel-wise rectified function. The parameters of the convolutional layers include a kernel size of N4×N4, 64 channels, a stride of S7, and zero-padding size of P4. The parameters of the linear rectified function include 64 channels and a two-dimensional dimension. The values ​​of the main network parameters, such as the kernel size, stride, and zero-padding size, are determined based on the actual application.

[0040] 5) The fifth layer of the network consists of convolutional layers and a channel-wise rectified function. The parameters of the convolutional layers include a kernel size of N5×N5, 128 channels, a stride of S8, and zero-padding size of P5. The parameters of the linear rectified function include 128 channels and a two-dimensional structure. The values ​​of the main network parameters, such as the kernel size, stride, and zero-padding size, are determined based on the actual application.

[0041] 6) The 6th layer of the network contains convolutional layers. The parameters of the convolutional layers are: kernel size N6×N6, number of channels 256, stride S9, and zero-padding size P6. The values ​​of the main network parameters such as kernel size, stride, and zero-padding size of the 6th layer of the network are determined according to the actual application.

[0042] 7) The 7th layer of the network contains convolutional layers. The parameters of the convolutional layers are: kernel size N7×N7, number of channels 256, stride S10, and zero-padding size P7. The values ​​of the main network parameters such as kernel size, stride, and zero-padding size of the 7th layer are determined according to the actual application.

[0043] 8) The 8th layer group of the network includes a dropout layer. In the parameters of the dropout layer, the dropout rate is Dr, and the value range of Dr is 0 < Dr < 1; the dropout layer randomly deactivates neurons with a probability of Dr×100%;

[0044] 9) The 9th layer group of the network includes a convolutional layer. In the parameters of the convolutional layer, the convolutional kernel size is N8×N8, the number of channels is 256, the stride is S11, and the zero-padding size is P8; The values of the main network parameters such as the convolutional kernel size, stride, and zero-padding size of the 9th layer group of the network are determined according to the actual application situation.

[0045] 10) The 10th layer group of the network includes a flattening layer. After the output result of the 9th layer group is processed by this layer group, it is flattened into one-dimensional vector data for determining the target category.

[0046] In the training of the classification network, the multi-label classification task cross-entropy loss function is used to evaluate the prediction loss for the sample distribution. The calculation method of the multi-label classification task cross-entropy loss function LossC is shown in formula (6):

[0047] (6)

[0048] In expression (6), B_size is the number of samples in each batch for training, the total number of classification categories is n, the true distribution of the samples is w, and the output distribution of the classification network is w , 。

[0049] The network classifier is obtained through the training of the sea radar target convolutional neural classification network based on convolution and per-channel rectification function. The classifier includes the sea radar target convolutional neural classification network based on convolution and per-channel rectification function, and the corresponding network weights.

[0050] Step 5: Use the samples of the radar target classification test set to test the trained classification network and network weights, and record the two test metrics of recall rate (Recall) and mean average precision (mAP, mean Average Precision).

[0051] Step 6: Check whether the number of training iterations of the sea radar target classification method based on convolution and per-channel rectification function has reached the maximum value. If the number of iterations has not reached the maximum value, adjust the network weights and jump to Step 4; if the number of iterations has reached the set maximum value, go to Step 7.

[0052] Step 7: Based on the two metrics of recall rate and mean average precision recorded for the classifier test, obtain the classifier with the optimal performance metrics in all iterations for classifying sea radar targets.

[0053] This invention improves deep convolutional neural network classification by using a channel-wise rectified function based on an attention mechanism. This enhances the network's ability to adapt to radar targets of different sizes and shapes, as well as different clutter backgrounds, thereby improving its classification capability for weak radar targets on the sea surface.

[0054] As one embodiment, a sea surface radar target classification method based on convolution and channel-by-channel rectified functions is implemented as follows:

[0055] Step 1: Obtain radar target detection point data by processing the raw echo of the maritime surveillance radar. The processing of the raw echo mainly includes IQ demodulation, pulse compression, mode extraction, and target detection processing. After mode extraction, the raw echo of the maritime surveillance radar is converted into maritime surveillance radar video data. By performing target detection on the radar video data, radar target detection point data can be obtained.

[0056] Step 2: Based on the radar surface target detection point data and the target information acquired during the data acquisition process, targets in the radar video data are labeled using rectangular bounding boxes to classify them, completing the radar surface target classification labeling. The radar surface target classification labeling information includes the target's bounding box size, position, and target category name. Based on the target category name and the bounding box position, a local radar target slice of size J1×J2 is extracted from the 2D radar video data. Typically, J1×J2 is ≥32, and J2≥32. The radar target should usually be located at the center of the slice image. A radar sea detection target classification dataset is constructed using the target category labeling information and the radar target slice image to support the development of a deep learning target classification network model for sea detection radar.

[0057] Step 3: Based on the application scenarios and requirements of radar maritime detection target classification, and referring to the principles of completeness and balance in dataset construction, samples are extracted from the radar maritime detection target classification dataset to construct training and test datasets. The ratio of the number of training and test datasets is usually between 6:4 and 8:2. The completeness principle means that, for the application scenario, the dataset should cover all influencing factors of the object to be processed. The balance principle means that, for the application scenario, the distribution of the number of samples in the dataset should follow a reasonable proportion.

[0058] Step 4: Train the training dataset using a sea surface radar target convolutional neural network based on convolution and channel-wise rectified functions. This invention designs a sea surface radar target convolutional neural network based on convolution and channel-wise rectified functions; the main structure and parameters of the sea surface radar target convolutional neural network based on convolution and channel-wise rectified functions are as follows: Figure 1As shown. The size of the input sample image for the classification network is J1×J2. Taking J1=88 and J2=88 as an example, the network structure and its parameters are described below. The network consists of 10 layers, and the network structure is described as follows:

[0059] 1) The first layer of the network consists of a convolutional layer, a channel-wise rectified function, and a max pooling layer. The parameters of the convolutional layer include a kernel size of N1×N1, 8 channels, a stride of S1, and zero-padding size of P1. The parameters of the rectified function include 8 channels and a two-dimensional dimension. The parameters of the max pooling layer include a kernel size of M1×M1 and a stride of S2. The values ​​of the main network parameters, such as kernel size, stride, and zero-padding size, are determined according to the actual application. In this embodiment, the main parameters of the first layer are N1=3, S1=1, P1=0, M1=2, and S2=2.

[0060] The schematic diagram of the channel-by-channel rectification function is as follows: Figure 2 As shown, the input feature map data x has dimensions C×H×W, where C represents that the feature map data has C channels, H represents that the feature map data has H rows, and W represents that the feature map data has W columns. The input feature map data x is first processed by average pooling, resulting in C parameters. The output of average pooling is then input into a fully connected layer for the first fully connected processing, resulting in C / G parameters. The output of the fully connected layer is then processed by a linear rectified function. The output of the linear rectified function is then input into a fully connected layer for the second fully connected processing, resulting in 2KC parameters, where K is the number of functions. The output of the second fully connected layer is normalized using a Sigmoid parameter, constraining the output parameters to the closed interval [-1, 1]. The 2KC parameters Ω output after normalization are:

[0061] (1)

[0062] Where the superscript T denotes transpose, the values ​​of Δpk,c(x) and Δqk,c(x) depend on the feature map data x, the variable k takes the value of all positive integers in the closed interval [1, K], and the variable c takes the value of all positive integers in the closed interval [1, C]. Based on Δpk,c(x) and Δqk,c(x), the activation function parameters pk,c and qk,c are calculated using the following two expressions:

[0063] (2)

[0064] (3)

[0065] Where λp and λq are hyperparameters, and uk and vk are the initial values ​​of the activation function parameters; Figure 2In the expression p1:K,1:C, we have KC parameters pk,c (where the variable k takes the value of all positive integers within the closed interval [1, K], and the variable c takes the value of all positive integers within the closed interval [1, C]), and q1:K,1:C, we have KC parameters qk,c (where the variable k takes the value of all positive integers within the closed interval [1, K], and the variable c takes the value of all positive integers within the closed interval [1, C]). In this embodiment, the parameters λp and λq take the values ​​λp=1 and λq=0.5, and the parameters uk and vk take the values ​​1 and 0, respectively. Figure 2 In this context, u1:K represents a K×1 vector consisting of 1s, and v1:K represents a K×1 vector consisting of 0s; the final output of the hyperparameter function η in the channel-by-channel rectified function is:

[0066] (4)

[0067] Where x is the input feature data, xc is the feature data of the c-th dimension, pk,c and qk,c are the activation function parameters; the final output of the channel-wise rectified function y is:

[0068] (5)

[0069] Where max represents taking the maximum value among all variables;

[0070] 2) The second layer of the network consists of a convolutional layer, a channel-wise rectified function, and a max pooling layer. The parameters of the convolutional layer include a kernel size of N2×N2, 16 channels, a stride of S3, and zero-padding size of P2. The parameters of the rectified function include 16 channels and a two-dimensional dimension. The parameters of the max pooling layer include a kernel size of M1×M1 and a stride of S4. The values ​​of the main network parameters, such as kernel size, stride, and zero-padding size, are determined according to the actual application. In this embodiment, the main parameters of the second layer are N2=4, S3=1, P2=0, M2=2, and S4=2.

[0071] 3) The third layer of the network consists of a convolutional layer, a channel-wise rectified function, and a max pooling layer. The parameters of the convolutional layer include a kernel size of N3×N3, 32 channels, a stride of S5, and zero-padding size of P3. The parameters of the rectified function include 32 channels and a two-dimensional dimension. The parameters of the max pooling layer include a kernel size of M3×M3 and a stride of S6. The values ​​of the main network parameters, such as kernel size, stride, and zero-padding size, are determined according to the actual application. In this embodiment, the main parameters of the third layer are N3=3, S5=1, P3=0, M3=2, and S6=2.

[0072] 4) The fourth network layer group sequentially includes a convolutional layer and a per-channel rectification function. Among them, in the parameters of the convolutional layer, the convolutional kernel size is N4×N4, the number of channels is 64, the stride is S7, and the zero-padding size is P4; in the parameters of the rectified linear unit function, the number of channels is 64 and the dimension is two-dimensional; the values of the main network parameters such as the convolutional kernel size, stride, and zero-padding size of the fourth network layer group are determined according to the actual application situation; in this embodiment, the main parameter values of the fourth network layer group are N4 = 3, S7 = 1, and P4 = 0.

[0073] 5) The fifth network layer group sequentially includes a convolutional layer and a per-channel rectification function. Among them, in the parameters of the convolutional layer, the convolutional kernel size is N5×N5, the number of channels is 128, the stride is S8, and the zero-padding size is P5; in the parameters of the rectified linear unit function, the number of channels is 128 and the dimension is two-dimensional; the values of the main network parameters such as the convolutional kernel size, stride, and zero-padding size of the fifth network layer group are determined according to the actual application situation; in this embodiment, the main parameter values of the fifth network layer group are N5 = 3, S8 = 1, and P5 = 0.

[0074] 6) The sixth network layer group includes a convolutional layer. In the parameters of the convolutional layer, the convolutional kernel size is N6×N6, the number of channels is 256, the stride is S9, and the zero-padding size is P6; the values of the main network parameters such as the convolutional kernel size, stride, and zero-padding size of the sixth network layer group are determined according to the actual application situation; in this embodiment, the main parameter values of the sixth network layer group are N6 = 3, S9 = 1, and P6 = 0.

[0075] 7) The seventh network layer group includes a convolutional layer. In the parameters of the convolutional layer, the convolutional kernel size is N7×N7, the number of channels is 256, the stride is S10, and the zero-padding size is P7; the values of the main network parameters such as the convolutional kernel size, stride, and zero-padding size of the seventh network layer group are determined according to the actual application situation; in this embodiment, the main parameter values of the seventh network layer group are N7 = 3, S10 = 1, and P7 = 0.

[0076] 8) The eighth network layer group includes a dropout layer. In the parameters of the dropout layer, the dropout rate is Dr, and the value range of Dr is 0 < Dr < 1; the dropout layer randomly deactivates neurons with a probability of Dr×100%.

[0077] 9) The ninth network layer group includes a convolutional layer. In the parameters of the convolutional layer, the convolutional kernel size is N8×N8, the number of channels is 256, the stride is S11, and the zero-padding size is P8; the values of the main network parameters such as the convolutional kernel size, stride, and zero-padding size of the ninth network layer group are determined according to the actual application situation; in this embodiment, the main parameter values of the ninth network layer group are N8 = 2, S11 = 1, and P8 = 0.

[0078] 10) The 10th layer of the network contains a flattening layer. The output of the 9th layer is processed by this layer and flattened into one-dimensional vector data, which is used to determine the target category.

[0079] In the training of the classification network, the cross-entropy loss function for multi-label classification tasks is used to evaluate the prediction loss for the sample distribution. The calculation method of the cross-entropy loss function LossC for multi-label classification tasks is shown in formula (6):

[0080] (6)

[0081] In expression (6), B_size is the number of samples trained in each batch, the total number of classification categories is n, the true distribution of the samples is w, and the output distribution of the classification network is w. , .

[0082] The network classifier is obtained by training a sea surface radar target convolutional neural network based on convolution and channel-wise rectified functions. The classifier includes the sea surface radar target convolutional neural network based on convolution and channel-wise rectified functions, as well as the corresponding network weights.

[0083] Step 5: Use radar target classification test set samples to test the trained classification network and network weights, and record the two test metrics: recall and mean average precision (mAP).

[0084] Step 6: Check if the number of training iterations for the sea surface radar target classification method based on convolution and channel-wise rectification function has reached the maximum value. If the number of iterations has not reached the maximum value, adjust the network weights and jump to step 4; if the number of iterations has reached the set maximum value, proceed to step 7; in this embodiment, the maximum number of iterations is set to 300.

[0085] Step 7: Based on the recorded recall and mean precision of the classifier tests, determine the classifier with the best performance across all iterations for classifying marine radar targets.

[0086] As one embodiment, a control server is manifested in the form of a general-purpose computing device, including but not limited to: one or more processors or processing units, system memory, and a bus connecting the system memory and the processing unit.

[0087] A bus refers to one or more of several bus architectures, including a memory bus or memory controller peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. Examples of these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0088] The control server includes a variety of computer system-readable media. These media can be any available media that can be accessed by the control server, including volatile and non-volatile media, and removable and non-removable media.

[0089] The system memory may include computer system-readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The control server may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the storage system may be used to read and write non-removable, non-volatile magnetic media (commonly referred to as "hard disk drives"). Disk drives for reading and writing to removable non-volatile disks (e.g., "floppy disks") and optical disk drives for reading and writing to removable non-volatile optical discs (e.g., CD-ROMs, DVD-ROMs, or other optical media) may be provided. In these cases, each drive may be connected to a bus via one or more data media interfaces. The system memory may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of the present invention.

[0090] A program / utility having a set (at least one) of program modules may be stored, for example, in system memory. Such program modules include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. The program modules typically perform the functions and / or methods described in the embodiments of this invention.

[0091] The control server can also communicate with one or more external devices, such as keyboards, pointing devices, displays, etc., and with one or more devices that enable users to interact with them, and / or with any device that enables the control server to communicate with one or more other computing devices (e.g., network cards, modems, etc.). This communication can be performed via input / output (I / O) interfaces. Furthermore, the server can communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via a network adapter. The network adapter communicates with other modules of the control server via a bus. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with the control server, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems. The processing unit executes various functional applications and data processing by running programs stored in the system memory, such as implementing the sea surface radar target classification method based on convolution and channel-by-channel rectification functions provided in the embodiments of this invention.

[0092] As one embodiment, a computer-readable storage medium stores a computer program that, when executed by a processor, can implement any of the sea surface radar target classification methods based on convolution and channel-by-channel rectification functions described in the above embodiments.

[0093] The computer storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. Computer-readable storage media can be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0094] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0095] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to:

[0096] Wireless, wires, fiber optics, RF, etc., or any suitable combination of the above.

[0097] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

Claims

1. A method for classifying sea surface radar targets based on convolution and channel-by-channel rectified functions, characterized in that, Includes the following steps: Step 1: Process the raw echo of the sea surface detection radar to obtain radar target detection point data; Step 2: Based on the radar sea surface target detection point data, target information is matched during the data acquisition process. Targets in the radar video data are labeled with rectangular marker boxes to complete the radar sea surface target classification labeling. Extract a local slice of radar target of a set size from radar surface target detection point data; the radar target should usually be located at the center of the slice image. A radar target classification dataset for maritime detection was constructed using target category annotation information and radar target slice images. Step 3: Based on the application scenarios and requirements of radar target classification for sea detection, and referring to the principles of completeness and balance in dataset construction, samples are extracted from the radar target classification dataset for sea detection and training datasets and test datasets are constructed. Step 4: Train the training dataset using a sea surface radar target convolutional neural classification network based on convolution and channel-wise rectified functions; Step 5: Test the trained classification network and network weights using radar target classification test set samples, and record the recall rate and mean precision. Step 6: Check if the number of training iterations for the sea surface radar target classification method based on convolution and channel-wise rectification function has reached the maximum value. If the number of iterations has not reached the maximum value, adjust the network weights and jump to step 4; if the number of iterations has reached the set maximum value, proceed to step 7. Step 7: Based on the recorded recall and mean precision of the classifier tests, determine the classifier with the best performance across all iterations for classifying marine radar targets.

2. The sea surface radar target classification method based on convolution and channel-by-channel rectified functions according to claim 1, characterized in that, The specific process of processing the raw echo of the maritime detection radar to obtain radar target detection point data on the sea surface is as follows: The raw radar echo is sequentially processed by IQ demodulation, pulse compression, mode extraction, and target detection. After mode extraction, the raw radar echo is converted into radar video data. Target detection is then performed on the radar video data to obtain radar target detection point data.

3. The sea surface radar target classification method based on convolution and channel-by-channel rectified functions according to claim 1, characterized in that, Radar sea surface target classification and labeling information includes the target's bounding box size, location, and target category name.

4. The sea surface radar target classification method based on convolution and channel-by-channel rectified functions according to claim 3, characterized in that, Based on the target category name information and the location of the classification label box, a local slice of radar target with size J1×J2 is extracted from the two-dimensional radar video data, where J1≥32 and J2≥32.

5. The sea surface radar target classification method based on convolution and channel-by-channel rectified functions according to claim 1, characterized in that, The convolutional neural network for classifying sea surface radar targets comprises 10 layers connected in sequence. The first layer consists of a convolutional layer, a channel-wise rectified function, and a max-pooling layer. The second layer consists of the same convolutional layer, channel-wise rectified function, and max-pooling layer. The third layer consists of the same convolutional layer, channel-wise rectified function, and max-pooling layer. The fourth layer consists of a convolutional layer and a channel-wise rectified function. The fifth layer consists of the same convolutional layer and a channel-wise rectified function. The sixth and seventh layers consist of convolutional layers. The eighth layer consists of a random deactivation layer. The ninth layer consists of a convolutional layer. The tenth layer consists of a flattened layer.

6. The sea surface radar target classification method based on convolution and channel-by-channel rectified functions according to claim 1, characterized in that, The channel-wise rectified function is the same in each layer group. The input of the channel-wise rectified function is the feature map data x with dimensions C×H×W, where C represents that the feature map data has C channels, H represents that the feature map data has H rows, and W represents that the feature map data has W columns. The input feature map data x is first subjected to average pooling, resulting in C parameters. Then, the output of average pooling is input into a fully connected layer for the first fully connected processing, resulting in C / G parameters. Subsequently, the output of the fully connected layer is input into a linear rectified function for processing. The output of the linear rectified function is input into a fully connected layer for the second fully connected processing, resulting in 2KC parameters, where K is the number of functions. The output of the second fully connected layer is normalized using a Sigmoid parameter, constraining the output parameters to the closed interval [-1,1]. After normalization, 2KC parameters are output, specifically: Wherein, the superscript T indicates transpose. The slope parameter residual of the k-th linear function. Let x be the slope parameter residual of the k-th linear function, x be the feature map data, the value range of variable k is all positive integers in the closed interval [1, K], and the value range of variable c is all positive integers in the closed interval [1, C]. based on and The slope parameter of the k-th linear function is calculated. and the slope parameter of the k-th linear function : in, and For hyperparameters, and The initial values ​​for the activation function parameters are: k = 1, K; the variable k takes the value of all positive integers in the closed interval [1, K]; and the variable c takes the value of all positive integers in the closed interval [1, C]. The final output of the hyperparameter function η in the channel-by-channel rectification function is: Where x represents the input feature data; The final output of the channel-by-channel rectification function is: Here, max represents taking the maximum value among all variables.

7. The sea surface radar target classification method based on convolution and channel-by-channel rectified functions according to claim 1, characterized in that, When training the training dataset using a sea surface radar target convolutional neural classification network based on convolution and channel-wise rectified functions, the cross-entropy loss function for multi-label classification tasks is used to evaluate the prediction loss for the sample distribution. The network classifier is obtained by training the sea surface radar target convolutional neural classification network based on convolution and channel-wise rectified functions. The classifier includes the sea surface radar target convolutional neural classification network based on convolution and channel-wise rectified functions, as well as the corresponding network weights.

8. The sea surface radar target classification method based on convolution and channel-by-channel rectified functions according to claim 1 or 7, characterized in that, In step 4, the sea surface radar target convolutional neural classification network based on convolution and channel-wise rectified functions uses the cross-entropy loss function for multi-label classification tasks to evaluate the prediction loss for the sample distribution. The specific calculation formula is as follows: In the formula, B_size is the number of samples in each batch for training, and n is the total number of classification categories. For the true distribution of the sample, This represents the output distribution of the classification network.

9. A control server, characterized in that, include: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the sea surface radar target classification method based on convolution and channel-by-channel rectification function as described in any one of claims 1-8.

10. A storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the sea surface radar target classification method based on convolution and channel-by-channel rectification function as described in any of claims 1-8.

Images