Sea clutter filtering method based on lightweight convolutional neural network

Abstract

The invention discloses a sea clutter filtering method based on a lightweight convolutional neural network. The method comprises the following steps: firstly, taking a constant false alarm rate detection model as a convolution kernel of a convolutional neural network, and establishing the convolutional neural network based on the CFAR convolution kernel; inputting the calibrated sea clutter and target data into the network for training, and naming the trained model as a teacher network model; establishing a lightweight convolutional neural network which is called a student network; based on a knowledge distillation algorithm, a teacher network and a student network are put in the same network for training, and after being trained, the student network is used as a lightweight convolutional neural network for sea clutter filtering. Experiments prove that after the network model is subjected to lightweight processing, the parameter scale is reduced to 2.32% of the original parameter scale, and edge measurement deployment is facilitated.

Description

Technical Field

[0001] The invention belongs to the field of radar target detection and recognition, and is suitable for solving the problem of sea clutter identification and filtering in radar detection. Background Art

[0002] Constant False Alarm Rate (CFAR) is a classic method for radar target detection in sea clutter. However, the detection threshold setting of this method requires the knowledge of the distribution of sea clutter or manual experience, which is not conducive to the radar's adaptability to sea clutter in changing environments. Deep learning provides a new idea for radar target detection and clutter suppression in sea clutter. Combining CFAR with deep learning and identifying sea clutter based on data-driven has gradually become a hot topic in the field of radar detection in sea clutter in recent years.

[0003] Because it is based on deep learning methods, such as convolutional neural networks, it is generally based on large network models and large amounts of data training, making it difficult to deploy on the edge of radar and difficult to apply to actual devices. Reducing the model size means low accuracy, while increasing the model size means the algorithm is too large and not convenient for edge deployment.

[0004] In 2014, Hinton et al. first proposed the concept of knowledge distillation and verified its effectiveness and feasibility in convolutional neural network compression through experiments. The knowledge distillation strategy has played a role in many computer vision applications such as object detection, semantic segmentation, object recognition, video classification, and image dehazing. The essence of knowledge distillation is that teachers will pass on their thinking process and summary as a kind of knowledge essence to students, and students will obtain abstract and refined knowledge through understanding and learning, so as to reach a level close to that of teachers.

[0005] The basic idea of ​​knowledge distillation is to let the convolutional neural network imitate human learning behavior, refine the knowledge learned by the large network (teacher network) and pass it to the small network (student network), and guide the training of the small network, thereby achieving the purpose of compressing the large network into a small network. It has become a trend to use the knowledge distillation method to lightweight deep learning models. If the complex deep learning neural network model can be successfully compressed into a small neural network through this method, the possibility of deploying the model to edge devices will be greatly improved.

[0006] The invention discloses a sea clutter identification method based on a lightweight convolutional neural network. First, a constant false alarm detection model is used as the convolution kernel of a convolutional neural network to establish a convolutional neural network based on a CFAR convolution kernel, referred to as DSW-CNN; then calibrated sea clutter and target data are input into the network for training, and the trained model is called a teacher network model; then a lightweight convolutional neural network is established, referred to as LDSW-CNN, called a student network; through a target distillation algorithm, the teacher network and the student network are placed in the same network for training, and after the student network is trained, it is used as a lightweight convolutional neural network for sea clutter filtering. Experimental verification shows that after the network model is lightweighted, the parameter scale is reduced to 2.32% of the original, which is conducive to edge measurement deployment. Summary of the invention

[0007] The problem to be solved by the present invention is how to identify and filter out sea clutter in radar target detection under sea clutter background, and a sea clutter identification method based on a lightweight convolutional neural network is provided.

[0008] A sea clutter filtering method based on a lightweight convolutional neural network comprises the following steps:

[0009] Step 1: Use the constant false alarm model as the convolution kernel of the convolutional neural network, and establish a convolutional neural network based on the CFAR convolution kernel, including convolution layer, maximum pooling layer, convolution layer, maximum pooling layer, convolution layer, maximum pooling layer, transposed convolution layer, transposed convolution layer, transposed convolution layer, convolution layer, and convolution layer;

[0010] Step 2: Input the calibrated sea clutter and target data into the convolutional neural network based on the CFAR convolution kernel for training, and use the trained model as the teacher network model;

[0011] Step 3: Build a lightweight convolutional neural network, including convolution layer, maximum pooling layer, transposed convolution layer, and convolution layer, and use this network as the student network model;

[0012] Step 4: Through the target distillation algorithm, the input data is simultaneously input into the trained teacher network model and the student network model to be trained. A soft target can be obtained through the teacher network output and the distillation temperature parameter. This value is combined with the student network model output value for loss processing to obtain the distillation loss. The prediction result of the student network and the true expected value are processed for loss to obtain the student loss. The distillation loss and the student loss are weighted summed to obtain the loss function of a training. Through this method, the student network is trained, and finally a trained lightweight convolutional neural network model is obtained;

[0013] Step 5: Input the radar signal into the lightweight convolutional neural network model to obtain the target signal after sea clutter suppression and filter out the sea clutter in the signal.

[0014] Furthermore, the convolutional neural network based on the CFAR convolution kernel and the lightweight convolutional neural network in step 1 and step 3 include

[0015] The convolutional neural network based on the CFAR convolution kernel mentioned in step 1 includes a 3-layer convolutional network. In practical applications, the number of convolutional layers can be expanded as needed. The maximum pooling method is used, and the average pooling method is used according to the actual situation.

[0016] The lightweight convolutional neural network mentioned in step 3 builds a single-layer convolutional neural network. In actual applications, convolutional neural networks with other layers can be selected as needed.

[0017] The beneficial effect of the present invention is that the proposed sea clutter identification method based on lightweight convolutional neural network is superior to the traditional CFAR algorithm in low observable target detection, and its performance is slightly lower than that of the network before lightweighting. However, after lightweighting, the parameter scale of the network is reduced to less than 10% of the original. The network can run at a speed of tens of milliseconds when processing sea clutter, and the running time is reduced by more than half, which can meet the needs of online real-time sea clutter filtering. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Attached Figure 1 is a flow chart of the method steps of the present invention;

[0019] Attached Figure 2 This is the simulation data when the signal-to-noise ratio is 8dB;

[0020] Attached Figure 3 This is the CFAR detection result when the signal-to-noise ratio is 8dB;

[0021] Attached Figure 4 This is the DSW-CNN detection result when the signal-to-noise ratio is 8dB;

[0022] Attached Figure 5 This is the LDSW-CNN detection result when the signal-to-noise ratio is 8dB. DETAILED DESCRIPTION

[0023] The present invention will be described in further detail below with reference to the accompanying drawings.

[0024] See attached Figure 1 A sea clutter filtering method based on a lightweight convolutional neural network of the present invention comprises the following steps:

[0025] Step 1: Use the constant false alarm model as the convolution kernel of the convolutional neural network, and establish a convolutional neural network based on the CFAR convolution kernel, including convolution layer, maximum pooling layer, convolution layer, maximum pooling layer, convolution layer, maximum pooling layer, transposed convolution layer, transposed convolution layer, transposed convolution layer, convolution layer, and convolution layer;

[0026] Step 2: Input the calibrated sea clutter and target data into the convolutional neural network based on the CFAR convolution kernel for training, and use the trained model as the teacher network model;

[0027] Step 3: Build a lightweight convolutional neural network, including convolution layer, maximum pooling layer, transposed convolution layer, and convolution layer, and use this network as the student network model;

[0028] Step 4: Through the target distillation algorithm, the input data is simultaneously input into the trained teacher network model and the student network model to be trained, and the distillation temperature parameter T=5 is set. A soft target can be obtained through the teacher network output and the distillation temperature parameter. This value is combined with the student network model output value for loss processing to obtain the distillation loss. The prediction result of the student network and the true expected value are processed for loss to obtain the student loss. The distillation loss and the student loss are weighted summed to obtain the loss function of a training. Through this method, the student network is trained, and finally a trained lightweight convolutional neural network model is obtained;

[0029] Step 5: Input the radar signal into the lightweight convolutional neural network model to obtain the target signal after sea clutter suppression and filter out the sea clutter in the signal.

[0030] The effect of the present invention can be further illustrated by the following simulation experiment:

[0031] Simulation environment: When SCR = 8dB, the simulation data is as shown in the attached Figure 2 .

[0032] Simulation results and analysis: Appendix Figure 3 , Attachment Figure 4 , Attachment Figure 5 The CFAR algorithm, DSW-CNN network, and LDSW-CNN network are compared and demonstrated respectively. The experimental results show that this method has better detection performance under low SCR conditions, which is better than the traditional CFAR method, but is slightly slower than the CFAR algorithm in running speed. The trainable parameter size of the DSW-CNN network is 31.7k, and the trainable parameter size of the LDSW-CNN network is 737, which is 2.32% of the model. The provided is still at the millisecond level, which can meet the real-time requirements of practical applications.

Claims

1. A sea clutter filtering method based on a lightweight convolutional neural network, characterized in that: The following steps are involved: Step 1: Use the constant false alarm model as the convolution kernel of the convolutional neural network, and establish a convolutional neural network based on the CFAR convolution kernel, including convolution layer, maximum pooling layer, convolution layer, maximum pooling layer, convolution layer, maximum pooling layer, transposed convolution layer, transposed convolution layer, transposed convolution layer, convolution layer, and convolution layer; Step 2: Input the calibrated sea clutter and target data into the convolutional neural network based on the CFAR convolution kernel for training, and use the trained model as the teacher network model; Step 3: Build a lightweight convolutional neural network, including convolution layer, maximum pooling layer, transposed convolution layer, and convolution layer, and use this network as the student network model; Step 4: Through the target distillation algorithm, the input data is simultaneously input into the trained teacher network model and the student network model to be trained. A soft target is obtained through the teacher network output and the distillation temperature parameter. This value is combined with the student network model output value for loss processing to obtain the distillation loss. The prediction result of the student network and the true expected value are processed for loss to obtain the student loss. The distillation loss and the student loss are weighted summed to obtain the loss function of a training. Through this method, the student network is trained, and finally a trained lightweight convolutional neural network model is obtained; Step 5: Input the radar signal into the lightweight convolutional neural network model to obtain the target signal after sea clutter suppression and filter out the sea clutter in the signal.

2. The sea clutter filtering method based on a lightweight convolutional neural network according to claim 1 is characterized in that: The convolutional neural network based on the CFAR convolution kernel and the lightweight convolutional neural network to which steps 1 and 3 belong include: The convolutional neural network based on the CFAR convolution kernel mentioned in step 1 includes a 3-layer convolutional network. In practical applications, the number of convolutional layers can be expanded as needed. The maximum pooling method is used, and the average pooling method is used according to the actual situation. The lightweight convolutional neural network mentioned in step 3 builds a single-layer convolutional neural network. In actual applications, convolutional neural networks with other layers can be selected as needed.

Images