Intelligent ship distance Doppler data generation method based on multiple feature constraints

Abstract

The invention provides a ship distance Doppler data intelligent generation method based on multiple feature constraints, which uses a deep learning method in the field of artificial intelligence, solves the problems of limited data, insufficient features and the like in small sample learning, and realizes expansion and balance of real samples. According to the technical scheme, the method comprises a generator module based on a deep convolutional neural network, a discriminator module forming adversarial training and a loss function module based on regularization constraint. The generator module is used for sampling and generating noise, the discriminator module is used for judging whether generated data and real data are true or false, and the loss function module is used for calculating a penalty value according to the quality of the generated data. The whole system needs to train two models at the same time, one is a generator module used for simulating real data distribution, and the other is a discriminator module used for estimating the probability that a sample is from real data instead of the generation module. According to the method, the generative adversarial network is innovatively used for generating the R-D data set of the ship, the simulation degree of the generated image is improved in combination with the multi-feature loss function, and the quality and the number of training samples are effectively improved.

Description

Technical Field

[0001] The invention belongs to the field of intelligent signal generation, and in particular relates to a method for intelligently generating ship Doppler data (Range Doppler, RD) based on multiple feature constraints. Background Art

[0002] In the field of intelligent applications of detectors, ship detection and recognition under the background of sea clutter is one of the current research focuses. With the gradual development of technology, the traditional method of only detecting target location information can no longer meet the needs of multi-faceted target information in complex scenarios. The ability to extract detailed information of the target from the echo signal is an inevitable trend in the future development of artificial intelligence technology. Although deep learning has strong feature extraction capabilities, it requires a large amount of data as a training guarantee for the algorithm, otherwise it will not be able to give full play to its advantages. At present, due to the small number of training data samples and the difficulty in ensuring the completeness of the scene, the training and application of artificial intelligence algorithms are difficult, resulting in problems such as insufficient model training, poor scene adaptability, and weak generalization ability.

[0003] In recent years, with the rapid development of deep learning algorithms, many new frameworks based on neural networks have emerged. Among them, generative adversarial networks can achieve unsupervised learning without labeled training samples, and have been widely used in tasks such as image generation, completion, and style transfer. The basic idea of ​​generative adversarial networks is to learn the probability distribution of training samples and improve the model capabilities in the learning optimization process of the generator and discriminator. Traditional generative adversarial models do not limit the attributes of the data and cannot effectively constrain the generated results. There are often more noise points in the generated images, and the image clarity and resolution are poor. Recently, there have been many works based on generative adversarial networks for image generation, such as visible light images and SAR images. However, there are relatively few related studies on detector echo data. At present, there are few public literatures at home and abroad that apply GAN models to the field of RD image generation.

[0004] RD data is affected by the sea clutter generated by the complex ocean environment and the diversity of ship types, and the information presentation is more complex. The target has an open outline and is scattered at multiple points, and the effective area of ​​the target accounts for a much smaller proportion than the visible light data. In addition, large-scale data collection and sample annotation are relatively difficult or extremely costly. At the same time, with the increase in complex electromagnetic environment scenes, the number of target types in the detector's field of view has increased, which puts higher requirements on the collection quantity and image quality of ship RD training data.

[0005] At present, the sample quantity and feature diversity of RD images are difficult to meet the actual application requirements of deep learning for training data, which seriously limits the application of deep learning in this field. Summary of the invention

[0006] The purpose of the present invention is to provide a method for intelligently generating ship range Doppler data based on multiple feature constraints, which can realize high-quality automatic generation of marine ship RD images and solve the problems of insufficient number of measured data samples and insufficient feature diversity.

[0007] In order to achieve the above objectives, the present invention is implemented by the following technical solutions:

[0008] A method for intelligently generating ship range Doppler data based on multiple feature constraints, comprising:

[0009] Step 1: Collect a set of echo signal data of ships under real sea conditions. The detector antenna works in staring mode. Each data file consists of echo signal data of multiple pulse combinations arranged in sequence.

[0010] Step 2: Process the echo signal data one by one according to the pulse Doppler principle, including two operations: pulse compression and slow-time Fourier transform, to obtain the original image data set;

[0011] Step 3: Perform data annotation, data enhancement and data normalization preprocessing on the original image dataset to form a ship RD image enhancement set as the training dataset of the neural network;

[0012] Step 4: Build a generative adversarial network model, including a generator module, a discriminator module, and a loss function module;

[0013] Step 5: Initialize the network weight parameters, input the training data set generated in step 3 into the generative adversarial network, train the generator module to generate simulated samples from random noise, and realize the augmented generation of limited real sample data, while the discriminator module judges the generated samples and real samples and outputs a judgment result; according to the judgment result, the loss function module is used to calculate the penalty intensity, and the performance of the generator module and the discriminator module is improved by regularization constraints until the number of training times reaches the preset threshold; the loop ends and the final image diversity enhanced data set is generated.

[0014] Optionally, the step 1 specifically includes:

[0015] Step 1.1: The detector uses a combined pulse transmission system to calculate the distance using the time difference between the received signal and the transmitted signal;

[0016] Step 1.2: Set up the detector at the preset test location, start the data logger to record data, and simultaneously record the AIS device data. The staring mode data must last for more than 2 minutes, and stop collecting after meeting the requirements;

[0017] Step 1.3: The collected data file is composed of echo signal data of multiple pulse combinations arranged in sequence. The echo signal data of each pulse combination is composed of "data header + echo data", which contains sea clutter information and ship information under real sea conditions.

[0018] Optionally, the step 2 specifically includes:

[0019] Step 2.1: Input the echo signal data into the matched filter to filter the echo signal so that the amplitude is increased and the pulse width is narrowed to achieve pulse compression;

[0020] Step 2.2: Perform slow-time Fourier spectrum analysis. By calculating the discrete Fourier transform of each slow-time row data, a two-dimensional data matrix is ​​obtained, whose dimensions represent the fast time and Doppler frequency respectively. By imaging the echo signal data, the number of ships and the distance of the ships relative to the detector can be seen.

[0021] Optionally, the step 3 specifically includes:

[0022] Step 3.1: Use the self-developed human-computer interaction annotation program based on Matlab software to annotate the original image data set with ships. The txt file information generated by the annotation includes (x, y, h, w), where (x, y) represents the coordinates of the upper left corner of the target area annotation box, h represents the height of the target area annotation box, and w represents the width of the target area annotation box;

[0023] Step 3.2: Based on the annotation file, randomly intercept the ship area to obtain a data set that removes surrounding noise information and retains only ship features;

[0024] Step 3.3: Use the data preprocessing methods of rotation and folding to enhance the data set generated in step 3.2;

[0025] Step 3.4: Normalize the enhanced data so that the pixel values ​​of the image are at the same order of magnitude and balance the contribution of each feature.

[0026] Optionally, step 4 specifically includes:

[0027] Step 4.1: Construct a generator module, including a six-layer encoding feature extraction structure, which is used to extract features from the input image and generate images by random noise; the first five layers of the encoding feature extraction structure are composed of the encoding convolution layer Conv2DTranspose-pooling layer-batch normalization-ReLu activation function, and the sixth layer of the encoding feature extraction structure is composed of the encoding convolution layer Conv2DTranspose-pooling layer-batch normalization-Tanh activation function;

[0028] Step 4.2: Construct a discriminator module, including a five-layer neural network structure, which is used to identify the authenticity of the generated image output by the generator, that is, to determine whether the generated image is close to the real image; the first five layers of encoding feature extraction structure are composed of encoding convolution layer Conv2D-pooling layer-batch normalization-Leaky ReLu activation function-Dropout layer, and the sixth layer of encoding feature extraction structure is composed of encoding convolution layer Conv2D-Flatten layer-fully connected layer. After integrating all local features extracted by convolution, the discrimination result is output through Sigmoid activation function;

[0029] Step 4.3: Construct a loss function module based on multiple features. The loss function module includes the mean square error based on local pixels and the mean absolute error based on the global image. The two together constitute the feature mismatch penalty term. Add the two in a certain proportion to obtain the loss value, which is returned to the generator module to guide the model's parameter optimization process.

[0030] Optionally, the step S5 specifically includes:

[0031] Step 5.1: Initialize the network weight parameters, set the neural network gradient optimization algorithm, and set the initial learning rate and decay rate of the generator module and the discriminator module;

[0032] Step 5.2: Set the number of training rounds. When the number of training rounds is less than the set value, output the loss value of each round. When the loss value reaches the set value, the training ends and the final image diversity enhanced dataset is generated.

[0033] Compared with the prior art, the present invention has the following technical effects:

[0034] (1) The data generation method is highly automated and intelligent, and can autonomously generate highly realistic ship images. This method does not require tedious manual parameter adjustment and scene modeling, greatly reducing the cost and time consumption of the data preparation stage, and greatly improving the efficiency of data generation;

[0035] (2) The generated data simulation effect is good, far exceeding the degree of realism that can be achieved by traditional data simulation methods. This method clearly and accurately represents the target features, the data distribution is consistent, and the background noise and environmental clutter can be accurately simulated. At the same time, the generated data imaging resolution is high, and the contour boundary between the target and the background is clear;

[0036] (3) The generated data is diverse, ensuring the richness and quality of the data, effectively improving the noise resistance and generalization ability of the subsequent ship target detection and recognition model, and enabling it to better adapt to the complex and changeable marine environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for the description. Obviously, the drawings described below are an embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work:

[0038] Figure 1 It is a flow chart of the method for intelligently generating ship RD data based on multiple feature constraints in the present invention;

[0039] Figure 2 It is a schematic diagram of the working of the generative adversarial network in the present invention;

[0040] Figure 3 It is a schematic diagram of the generation result of the RD image of a small sample of ships according to the present invention. DETAILED DESCRIPTION

[0041] The scheme proposed by the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. According to the following description, the advantages and features of the present invention will be clearer. It should be noted that the accompanying drawings are in a very simplified form and use non-precise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the implementation method of the present invention. In order to make the purposes, features and advantages of the present invention more obvious and easy to understand, please refer to the accompanying drawings. It should be noted that the structures, proportions, sizes, etc. illustrated in the accompanying drawings of this specification are only used to match the contents disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the limiting conditions for the implementation of the present invention, so they have no technical substantive significance. Any structural modification, change in proportional relationship or adjustment of size, without affecting the effects that the present invention can produce and the purposes that can be achieved, should still fall within the scope of the technical content disclosed by the present invention.

[0042] The present invention provides a method for intelligently generating ship range Doppler data based on multiple feature constraints. Figure 1 to Figure 3 As shown, the method specifically comprises the following steps:

[0043] Step 1: Collect a set of echo signal data of ships under real sea conditions. The detector antenna works in staring mode. Each data file consists of echo signal data of multiple pulse combinations arranged in sequence. Specifically include:

[0044] Step 1.1: The detector uses a combined pulse transmission system to calculate the distance using the time difference between the received signal and the transmitted signal;

[0045] Step 1.2: Set up the detector at the preset test site, start the data logger to record data, and simultaneously record the AIS (Automatic Identification System) equipment data. The staring mode data must last for more than 2 minutes, and stop collecting after meeting the requirements;

[0046] Step 1.3: The collected data file is composed of echo signal data of multiple pulse combinations arranged in sequence. The echo signal data of each pulse combination is composed of "data header + echo data", which contains sea clutter information and ship information under real sea conditions.

[0047] Step 2: Process the echo signal data one by one according to the pulse Doppler principle, including pulse compression and slow-time Fourier transform operations, to obtain the original image data set. Specifically include:

[0048] Step 2.1: Input the echo signal data into the matched filter, filter the echo signal, increase the amplitude, narrow the pulse width, and achieve pulse compression; use the pulse compression method to obtain narrow-band pulses, thereby improving the distance resolution of the detector receiving signal.

[0049] The detector antenna and the ship can be equivalent to a linear time-invariant system. The detector echo signal is the result of the impulse response on the input pulse, specifically:

[0050] s τ (t) = s(t) * h(t)

[0051] Where s(t) is the input pulse, h(t) is the impulse response, and s τ (t) is the received echo signal;

[0052] The signal is input into the matched filter to filter the detector echo signal, so that the amplitude is increased and the pulse width is narrowed to achieve pulse compression, specifically:

[0053] s 0 (t) = s τ (t)*h τ (t)

[0054] Among them, s 0 (t) is the demodulated signal after the matched filter, h τ (t) is the matched filter of s(t).

[0055] Step 2.2: Perform slow-time Fourier spectrum analysis. By calculating the discrete Fourier transform of each slow-time row data, a two-dimensional data matrix is ​​obtained, whose dimensions represent the fast time and Doppler frequency respectively. By imaging the echo signal data, the number of ships and the distance of the ships relative to the detector can be seen.

[0056] Step 3: Perform data annotation, data enhancement and data normalization preprocessing on the original image data set to form a ship RD image enhancement set as the training data set for the neural network. Specifically include:

[0057] Step 3.1: Use the self-developed human-computer interaction annotation program based on Matlab software to annotate the original image data set with ships. The txt file information generated by the annotation includes (x, y, h, w), where (x, y) represents the coordinates of the upper left corner of the target area annotation box, h represents the height of the target area annotation box, and w represents the width of the target area annotation box;

[0058] Step 3.2: Based on the annotation file, randomly intercept the ship area to obtain a data set that removes surrounding noise information and retains only ship features;

[0059] Step 3.3: Use the data preprocessing methods of rotation and folding to enhance the data set generated in step 3.2;

[0060] Image rotation refers to rotating a certain point in the image counterclockwise or clockwise by a certain angle, specifically:

[0061]

[0062] Among them, (x, y) is the coordinate of the image before transformation, (x ′ ,y′) is the coordinate of the transformed image, θ is the angle of image rotation, usually 90°, 180°, 270° counterclockwise around the center point of the image;

[0063] Image folding refers to folding the image with a straight line in the image as the axis of symmetry, specifically:

[0064]

[0065] Among them, x 0 and 0 It is the axis of symmetry of the image fold, usually a straight line through the center of the image;

[0066] Step 3.4: Normalize the enhanced data so that the pixel values ​​of the image are at the same order of magnitude and balance the contribution of each feature. Specifically:

[0067]

[0068] Among them, f(x,y) is the pixel value before the image transformation, f(x ′ ,y ′ ) is the pixel value after normalization of the image, MIN refers to the minimum value of the image sample, and MAX refers to the maximum value of the image sample;

[0069] After the above series of preprocessing operations, the RD image enhancement set of the ship is obtained, which is used as the training set of the neural network.

[0070] Step 4: Build a generative adversarial network model, including the generator module, the discriminator module, and the loss function module.

[0071] like Figure 2 As shown, the generative adversarial network constructed in step 4 completes the image generation task. The network model consists of three parts: generator module, discriminator module, and loss function module. Step 4 specifically includes:

[0072] Step 4.1: Construct a generator module, including a six-layer encoding feature extraction structure, which is used to extract features from the input image and generate images by random noise; the first five layers of encoding feature extraction structure are composed of encoding convolution layer Conv2DTranspose-pooling layer-batch normalization-ReLu activation function, and the sixth layer of encoding feature extraction structure is composed of encoding convolution layer Conv2DTranspose-pooling layer-batch normalization-Tanh activation function. With the increase of the number of convolution layers, the extracted features gradually evolve from low-level visual features such as color, texture, and shape to high-level semantic features such as image abstract information;

[0073] Step 4.2: Construct a discriminator module, including a five-layer neural network structure, which is used to identify the authenticity of the generated image output by the generator, that is, to determine whether the generated image is close to the real image; the first five layers of encoding feature extraction structure are composed of encoding convolution layer Conv2D-pooling layer-batch normalization-Leaky ReLu activation function-Dropout layer, and the sixth layer of encoding feature extraction structure is composed of encoding convolution layer Conv2D-Flatten layer-fully connected layer. After integrating all local features extracted by convolution, the discrimination result is output through Sigmoid activation function;

[0074] Step 4.3: Construct a loss function module. The loss function module based on multiple features includes the mean square error based on local pixels and the mean absolute error based on the global image. The two together constitute the feature mismatch penalty term. Add the two in a certain proportion to obtain the loss value, which is returned to the generator module to guide the model's parameter optimization process.

[0075] Step 5: Initialize the network weight parameters, input the training data set generated in step 3 into the generative adversarial network, train the generator module to generate simulated samples from random noise, and realize the augmented generation of limited real sample data, while the discriminator module judges the generated samples and real samples and outputs a judgment result; according to the judgment result, the loss function module is used to calculate the penalty intensity, and the performance of the generator module and the discriminator module is improved by regularization constraints until the number of training times reaches the preset threshold; the loop ends and the final image diversity enhanced data set is generated.

[0076] Specifically, during the training process, the network weight parameters are first initialized, the neural network gradient optimization algorithm is set, and the initial learning rate and decay rate of the generator and discriminator are set; then, the number of training rounds is set, and when the number of training rounds is less than the set value, the loss value of each round is output; when the loss value reaches the set value, the training ends and the final image diversity enhanced dataset is generated.

[0077] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0078] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present invention. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present invention will occur. Therefore, the protection scope of the present invention should be limited by the appended claims.

Claims

1. A method for intelligently generating ship range Doppler data based on multiple feature constraints, characterized in that: include: Step 1: Collect a set of echo signal data of ships under real sea conditions. The detector antenna works in staring mode. Each data file consists of echo signal data of multiple pulse combinations arranged in sequence. Step 2: Process the echo signal data one by one according to the pulse Doppler principle, including two operations: pulse compression and slow-time Fourier transform, to obtain the original image data set; Step 3: Perform data annotation, data enhancement and data normalization preprocessing on the original image dataset to form a ship RD image enhancement set as the training dataset of the neural network; Step 4: Build a generative adversarial network model, including a generator module, a discriminator module, and a loss function module; Step 5: Initialize the network weight parameters, input the training data set generated in step 3 into the generative adversarial network, train the generator module to generate simulated samples from random noise, and realize the augmented generation of limited real sample data, while the discriminator module judges the generated samples and real samples and outputs a judgment result; according to the judgment result, the loss function module is used to calculate the penalty intensity, and the performance of the generator module and the discriminator module is improved by regularization constraints until the number of training times reaches the preset threshold; the loop ends and the final image diversity enhanced data set is generated.

2. The method for intelligently generating ship range Doppler data based on multiple feature constraints as claimed in claim 1, characterized in that: The step 1 specifically includes: Step 1.1: The detector uses a combined pulse transmission system to calculate the distance using the time difference between the received signal and the transmitted signal; Step 1.2: Set up the detector at the preset test location, start the data logger to record data, and simultaneously record the AIS device data. The staring mode data must last for more than 2 minutes, and stop collecting after meeting the requirements; Step 1.3: The collected data file is composed of echo signal data of multiple pulse combinations arranged in sequence. The echo signal data of each pulse combination is composed of "data header + echo data", which contains sea clutter information and ship information under real sea conditions.

3. The method for intelligently generating ship range Doppler data based on multiple feature constraints as claimed in claim 1, characterized in that: The step 2 specifically includes: Step 2.1: Input the echo signal data into the matched filter to filter the echo signal so that the amplitude is increased and the pulse width is narrowed to achieve pulse compression; Step 2.2: Perform slow-time Fourier spectrum analysis. By calculating the discrete Fourier transform of each slow-time row data, a two-dimensional data matrix is ​​obtained, whose dimensions represent the fast time and Doppler frequency respectively. By imaging the echo signal data, the number of ships and the distance of the ships relative to the detector can be seen.

4. The method for intelligently generating ship range Doppler data based on multiple feature constraints as claimed in claim 1, characterized in that: The step 3 specifically includes: Step 3.1: Use the self-developed human-computer interaction annotation program based on Matlab software to annotate the original image data set with ships. The txt file information generated by the annotation includes (x, y, h, w), where (x, y) represents the coordinates of the upper left corner of the target area annotation box, h represents the height of the target area annotation box, and w represents the width of the target area annotation box; Step 3.2: Based on the annotation file, randomly intercept the ship area to obtain a data set that removes surrounding noise information and retains only ship features; Step 3.3: Use the data preprocessing methods of rotation and folding to enhance the data set generated in step 3.2; Step 3.4: Normalize the enhanced data so that the pixel values ​​of the image are at the same order of magnitude and balance the contribution of each feature.

5. The method for intelligently generating ship range Doppler data based on multiple feature constraints as claimed in claim 1, characterized in that: The step 4 specifically includes: Step 4.1: Construct a generator module, including a six-layer encoding feature extraction structure, which is used to extract features from the input image and generate images by random noise; the first five layers of the encoding feature extraction structure are composed of the encoding convolution layer Conv2DTranspose-pooling layer-batch normalization-ReLu activation function, and the sixth layer of the encoding feature extraction structure is composed of the encoding convolution layer Conv2DTranspose-pooling layer-batch normalization-Tanh activation function; Step 4.2: Construct a discriminator module, including a five-layer neural network structure, which is used to identify the authenticity of the generated image output by the generator, that is, to determine whether the generated image is close to the real image; the first five layers of encoding feature extraction structure are composed of encoding convolution layer Conv2D-pooling layer-batch normalization-Leaky ReLu activation function-Dropout layer, and the sixth layer of encoding feature extraction structure is composed of encoding convolution layer Conv2D-Flatten layer-fully connected layer. After integrating all local features extracted by convolution, the discrimination result is output through Sigmoid activation function; Step 4.3: Construct a loss function module based on multiple features. The loss function module includes the mean square error based on local pixels and the mean absolute error based on the global image. The two together constitute the feature mismatch penalty term. Add the two in a certain proportion to obtain the loss value, which is returned to the generator module to guide the model's parameter optimization process.

6. The method for intelligently generating ship range Doppler data based on multiple feature constraints as claimed in claim 1, characterized in that: The step S5 specifically includes: Step 5.1: Initialize the network weight parameters, set the neural network gradient optimization algorithm, and set the initial learning rate and decay rate of the generator module and the discriminator module; Step 5.2: Set the number of training rounds. When the number of training rounds is less than the set value, output the loss value of each round. When the loss value reaches the set value, the training ends and the final image diversity enhanced dataset is generated.

Images