Marine radar target detection method based on dynamic activation and FCOS

By introducing dynamic activation functions based on attention mechanism into the FCOS detection network, the problem that the fixed ReLU activation function cannot adapt to different scales and backgrounds is solved, and the detection ability and clutter suppression ability of weak targets on the radar surface are improved.

CN118897971APending Publication Date: 2024-11-05THE 724TH RESEARCH INSTITUTE OF CHINA STATE SHIPBUILDING CORP LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410927826.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-05

AI Technical Summary

Technical Problem

The fixed ReLU activation function used in traditional FCOS detection networks cannot fully adapt to different scales and complex clutter backgrounds, resulting in insufficient detection capabilities of sea radar targets, especially in the detection of weak targets on sea surfaces.

Method used

A dynamic activation function based on attention mechanism is introduced, combined with the ResNet34 network and feature pyramid, a detector for FCOS detection network is built to enhance the network's adaptability to radar targets of different scales and shapes, as well as different clutter backgrounds.

Benefits of technology

It has improved the detection ability of weak targets on the radar sea surface, enhanced the ability to suppress false detection points of sea clutter and cloud and rain clutter, and significantly improved the detection accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118897971A_ABST
    Figure CN118897971A_ABST
Patent Text Reader

Abstract

The invention relates to a sea radar target detection method based on dynamic activation and FCOS. In the application of radar to sea detection target detection, sea clutter, cloud and rain clutter and other complex variable electromagnetic environments can reduce the detection performance of the radar to sea surface weak targets, and the method can dynamically adjust the activation function according to the characteristics of input data based on the FCOS detection network and the dynamic activation function in the deep neural network field. Therefore, the feature extraction and representation capability of the FCOS network is enhanced, the adaptability of the FCOS network to various detection scenes and weak radar target features is improved, and then the detection capability of the FCOS network to sea surface weak targets is enhanced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of radar target detection, and in particular to the field of sea radar target detection based on deep learning. Background Art

[0002] The ocean covers most of the earth's surface, and China has a long coastline. Therefore, the layout and construction of a strict and powerful ocean surveillance system is an important measure to strengthen national defense and maintain national marine security. In the complex ocean surveillance system, radar, as an important remote sensing sensor, plays an important role in fields such as the ocean as the detection background due to its unique advantages in target detection.

[0003] The excellent detection performance of the target detection algorithm based on deep learning has made it widely used in the field of radar target detection. As a typical deep learning target detection network, the activation function in the backbone network of the FCOS detection network is a fixed ReLU function, which cannot fully adapt to the different scales of targets and complex clutter detection backgrounds of radar sea detection. Therefore, it is necessary to introduce a dynamic activation function to enhance the ability of the FCOS network to adapt to different radar video data, so that the network can better adapt to radar targets of different scales and shapes and different clutter backgrounds, thereby improving the detection ability of radar weak targets on the sea surface. Summary of the invention

[0004] In view of the problem that the traditional FCOS detection uses a fixed ReLU activation function and cannot fully adapt to different targets and detection backgrounds, the present invention is based on a dynamic activation function based on an attention mechanism in the field of deep neural networks, and a FCOS (Fully Convolutional One-Stage) detection network, and proposes a sea radar target detection method based on dynamic activation and FCOS. The present invention enhances the ability of the detection network to adapt to radars of different scales and shapes and different clutter backgrounds, thereby improving the detection capability of radar weak targets on the sea surface.

[0005] The technical solution proposed by the present invention includes:

[0006] Step 1: Process the original echo of the sea detection radar; radar echo processing mainly includes IQ demodulation, anomaly removal, pulse compression, modulo operation, constant false alarm detection and point trace condensation. After completing the modulo operation, the radar video data is obtained. By performing constant false alarm detection and point trace condensation on the radar video data, the radar target detection point trace is obtained.

[0007] Step 2: Construction of radar sea target detection dataset. Based on the radar target constant false alarm detection point data and the corresponding multi-frame radar video data, a rectangular marking frame is used to mark the targets in the radar video data to construct the sea target detection data.

[0008] Step 3: According to the specific usage scenarios and requirements of radar sea target detection, and referring to the balance and completeness of the data set, samples in the data set are extracted and divided into training sets and test sets.

[0009] Step 4: Use the ResNet34 network based on dynamic activation function as the backbone feature extractor, and build a detection network in combination with the feature pyramid and the three-branch detection head. Use the detection network based on dynamic activation function and FCOS to train the samples in the training set, and obtain the detector through training. The detector includes the detection network based on dynamic activation function and FCOS, and the corresponding network weights.

[0010] Step 5: Use the radar target detection test set samples to test the detector obtained through training and record the test indicators.

[0011] Step 6: Check whether the number of training iterations has reached the maximum value. If not, adjust the network weights and jump to step 4. If it has reached the maximum value, go to step 7.

[0012] Step 7: Based on the recorded detector test indicators, output the detector with the best performance indicators in all iterations.

[0013] Furthermore, a detection network based on dynamic activation function and FCOS is used to train radar video data samples with labeled information and adjust network weights to obtain the best detector based on the current training set and test set.

[0014] Furthermore, in step 4, a ResNet34 network based on an attention activation function is used as a backbone feature extractor.

[0015] Aiming at the problem that traditional FCOS detection adopts a fixed ReLU activation function and cannot fully adapt to different targets and detection backgrounds, the present invention introduces a dynamic activation function to improve the FCOS backbone network, enhances the ability of the detection network to adapt to radar targets with different signal-to-noise ratios and shapes, and different clutter backgrounds, thereby improving the ability to suppress false detection points of sea clutter and cloud and rain clutter, and enhancing the detection capability of radar weak targets on the sea surface. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the processing flow of an embodiment of the present invention.

[0017] Figure 2It is a schematic diagram of the network structure of the design method of the present invention. DETAILED DESCRIPTION

[0018] The present invention proposes a method for detecting sea radar targets based on dynamic activation and FCOS. The preferred embodiment process is as follows: Figure 1 As shown, including:

[0019] Step 1: Process the original echo of the sea detection radar. Radar echo processing mainly includes IQ demodulation, anomaly removal, pulse compression, modulus operation, constant false alarm detection and point trace condensation. The radar original echo can obtain radar complex data after IQ demodulation and anomaly removal. Then, after pulse compression and modulus operation are performed on the radar complex data, radar video data is obtained. By performing constant false alarm detection and point trace condensation on the radar video data, the radar target detection point trace is obtained.

[0020] Step 2: Construction of radar sea target detection dataset. Based on the radar target constant false alarm detection point data and the corresponding multi-frame radar video data, a rectangular marking frame is used to mark the targets in the radar video data, and sea target detection data is constructed based on different detection scenarios.

[0021] Step 3: According to the specific usage scenarios and requirements of radar detection of sea targets in different weather conditions and detection of complex backgrounds, and referring to the requirements of data set balance and completeness, samples in the data set are extracted and divided into training set and test set.

[0022] Step 4: Use the ResNet34 network based on dynamic activation function as the backbone feature extractor, and build a detection network by combining feature pyramid and three-branch detection head. The network structure diagram is shown in the figure. Figure 2 As shown in Figure 2, the activation function in the C5 layer of the backbone network uses a dynamic activation function (DyReLU). The dynamic activation function expression y c As shown in formula (1):

[0023]

[0024] In expression (1), x is the input vector, x c is the feature of the cth dimension, θ is the hyperparameter function, a k,c and b k,c is the activation function parameter, K is the number of functions. The expression of the hyperparameter function θ is shown in formula (2).

[0025] [a 1,1 ,...a 1,C ,...,a K,1 ,...,a K,C ,b 1,1 ,...,b 1,C,...,b K,1 ,...,b K,C ] T =θ(x) (2)

[0026] In expression (2), C is the total number of dimensions.

[0027] DyReLU can have three forms, with different sharing mechanisms in spatial position and dimension. The first form shares both spatial position and dimension, which is simple to calculate, only needs to output 2K parameters, and has the weakest expressive power; the second form only shares spatial position and outputs 2KC parameters; the third form does not share spatial position and dimension, and each element in each dimension has a corresponding activation function. For input of size C×H×W, the number of output parameters is 2KCHW, which has the strongest expressive power.

[0028] The detection network based on dynamic activation function and FCOS is used to train the samples in the training set, and the detector is obtained through training. The detector includes the detection network based on dynamic activation function and FCOS, as well as the corresponding network weights.

[0029] Step 5: Use the radar target detection test set samples to test the detector obtained through training and record the test indicators. Commonly used test indicators are recall rate (Recall) and mean average precision (mAP, mean Average Precision).

[0030] Step 6: Check whether the number of training iterations has reached the maximum value. If not, adjust the network weights and jump to step 4. If it has reached the maximum value, go to step 7.

[0031] Step 7: Based on the recorded detector test indicators, output the detector with the best performance indicators in all iterations.

Claims

1. A method for sea radar target detection based on dynamic activation and FCOS, characterized in that: Step 1: Process the original echo of the sea detection radar, including IQ demodulation, anomaly rejection, pulse compression, modulo operation, constant false alarm detection and point trace condensation; after completing the modulo operation, obtain the radar video data; obtain the target point trace by performing constant false alarm detection and point trace condensation on the radar video data; Step 2: Based on the radar target constant false alarm detection results and the corresponding multi-frame radar video data, a rectangular marking frame is used to mark the targets in the radar video data to construct sea target detection data; Step 3: According to the specific usage scenarios and requirements of radar sea target detection, and referring to the balance and completeness of the data set, samples in the data set are extracted and divided into training sets and test sets; Step 4: Use the ResNet34 network based on the dynamic activation function as the backbone feature extractor, and build a detection network in combination with the feature pyramid and the three-branch detection head; use the detection network based on the dynamic activation function and FCOS to train the samples in the training set, and obtain the detector through training; the detector includes the detection network based on the dynamic activation function and FCOS, and the corresponding network weights; Step 5: Use the radar target detection test set samples to test the detector obtained through training and record the test indicators; Step 6: Check whether the number of training iterations has reached the maximum value. If not, adjust the network weights and jump to step 4. If it has reached the maximum value, go to step 7. Step 7: Based on the recorded detector test indicators, output the detector with the best performance indicators in all iterations.

2. The method for sea radar target detection based on dynamic activation and FCOS according to claim 1, characterized in that: A detection network based on dynamic activation function and FCOS is used to train radar video data samples with labeled information and adjust network weights to obtain the best detector based on the current training set and test set.

3. The method for sea radar target detection based on dynamic activation and FCOS according to claim 1, characterized in that: In step 4, the ResNet34 network based on the attention activation function is used as the backbone feature extractor.