CN117217103A - Satellite-borne SAR sea clutter generation method and system based on multi-scale attention mechanism - Google Patents
Satellite-borne SAR sea clutter generation method and system based on multi-scale attention mechanism Download PDFInfo
- Publication number
- CN117217103A CN117217103A CN202311486202.7A CN202311486202A CN117217103A CN 117217103 A CN117217103 A CN 117217103A CN 202311486202 A CN202311486202 A CN 202311486202A CN 117217103 A CN117217103 A CN 117217103A
- Authority
- CN
- China
- Prior art keywords
- sea
- sea clutter
- image
- data
- clutter
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 55
- 230000007246 mechanism Effects 0.000 title claims abstract description 33
- 238000004088 simulation Methods 0.000 claims abstract description 82
- 238000012545 processing Methods 0.000 claims abstract description 16
- 230000004927 fusion Effects 0.000 claims description 25
- 238000012549 training Methods 0.000 claims description 21
- 238000006243 chemical reaction Methods 0.000 claims description 18
- 230000008569 process Effects 0.000 claims description 18
- 238000005457 optimization Methods 0.000 claims description 14
- 238000001228 spectrum Methods 0.000 claims description 13
- 238000009826 distribution Methods 0.000 claims description 10
- 238000007781 pre-processing Methods 0.000 claims description 10
- 238000012546 transfer Methods 0.000 claims description 8
- 238000004458 analytical method Methods 0.000 claims description 6
- 238000005315 distribution function Methods 0.000 claims description 5
- 238000001914 filtration Methods 0.000 claims description 5
- 238000003709 image segmentation Methods 0.000 claims description 4
- 238000002360 preparation method Methods 0.000 claims description 4
- 230000010354 integration Effects 0.000 claims description 3
- 230000003595 spectral effect Effects 0.000 claims description 3
- 230000006870 function Effects 0.000 description 33
- 238000004422 calculation algorithm Methods 0.000 description 7
- 238000010586 diagram Methods 0.000 description 7
- 238000001514 detection method Methods 0.000 description 4
- 238000012544 monitoring process Methods 0.000 description 3
- 230000007423 decrease Effects 0.000 description 2
- 230000002708 enhancing effect Effects 0.000 description 2
- 238000011176 pooling Methods 0.000 description 2
- 238000007637 random forest analysis Methods 0.000 description 2
- 238000000611 regression analysis Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 230000008485 antagonism Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000005094 computer simulation Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 230000005484 gravity Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005192 partition Methods 0.000 description 1
- 230000000737 periodic effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Landscapes
- Radar Systems Or Details Thereof (AREA)
Abstract
The application discloses a satellite-borne SAR sea clutter generation method and system based on a multi-scale attention mechanism, which are widely applied to the field of radar image processing. The method comprises the following steps: generating a sea surface geometric image, establishing a sea clutter simulation model based on a multi-scale attention mechanism, inputting the sea clutter simulation model to generate a sea clutter image, and extracting sea clutter of a sea area corresponding to the sea surface geometric image; and matching the sea clutter data with the sea area of the meteorological data to form a data pair of the sea area meteorological data and the sea clutter data, inputting the data pair into the sea clutter simulation model, and outputting a sea clutter data set for optimizing the sea clutter simulation model to obtain an optimized model. The application has the characteristics of high precision, strong timeliness and simplified operation.
Description
Technical Field
The application relates to the technical field of remote sensing image processing, in particular to a satellite-borne SAR sea clutter generation method.
Background
The ocean occupies most of the earth's area, and the concealment of the object to be detected in the ocean is far better than that of land and air. As an important means for sea surface target detection, the radar is widely applied because of the advantages of being free from environmental weather influence, being capable of working all weather and all day long, and the like. However, under real conditions radar performance for sea detection is limited by many noise interference signals, and one of the most typical noise is sea clutter. When the radar transmitter transmits a signal to scan the sea surface, the receiver receives a back scattering echo signal from the sea surface, namely sea clutter. The echo received by the radar not only comprises useful target echo, but also comprises sea clutter around the environment where the target is located, the sea clutter has very strong power and large fluctuation, the non-Gaussian and non-stationarity is obvious, and for weak targets, the echo energy of a single sea clutter is far greater than that of the target echo, and the sea clutter must be suppressed. The complex sea clutter environment reduces the observability of sea surface targets, increasing the difficulty of the radar system in sea exploration, which is often faced in real-world applications. Therefore, modeling the sea clutter has important significance for effectively eliminating interference and improving the accuracy of radar target detection.
In the prior art, some traditional modeling methods, such as a statistical method and a filtering method, are mainly adopted, but the calculation steps of the methods are complicated, and the operation is inconvenient. Meanwhile, only the influence of natural wave on the sea surface is considered, and the influence of sea condition parameters such as wind speed, wind direction, effective wave height, wave direction and the like on a scattered wave signal is not considered. Therefore, the precision is not high, the timeliness is poor, and the requirements of real-time monitoring, early warning and the like cannot be met due to certain limitations.
Therefore, a new solution is needed to solve the above-mentioned problems.
Disclosure of Invention
In order to solve the problems generated by the prior art, the application provides a satellite-borne SAR sea clutter generation method and system based on a multi-scale attention mechanism, which have the advantages of high precision, strong timeliness and simplified operation.
In order to achieve the purpose, the satellite-borne SAR sea clutter generation method based on the multi-scale attention mechanism can adopt the following technical scheme: a satellite-borne SAR sea clutter generation method based on a multiscale attention mechanism generates a sea clutter image based on a sea clutter simulation model, wherein the sea clutter simulation model comprises a generator and a discriminator, the generator processes the image based on a network U-net structure and a limiting condition, and the discriminator processes the image based on a discriminating network structure; the generating method comprises the following steps:
1) And acquiring sea area meteorological data and sea clutter data under the meteorological conditions, wherein the sea clutter data are selected from SAR images.
2) And preprocessing SAR images, and matching the SAR images with sea area meteorological data to form a sea clutter data set and a sea area meteorological data set of the same sea area.
3) Establishing a sea surface simulation model, and inputting a meteorological data set into the sea surface simulation model after unifying the format to generate a sea surface geometric image set.
4) And extracting features of the sea surface geometric image set and the sea clutter data set as limiting conditions of a generator to respectively form sea surface simulation data feature parameters and sea surface echo data feature parameters.
5) And establishing an LSGAN least squares loss function.
6) Establishing a sea clutter simulation model based on a multi-scale attention mechanism and a transducer module, and introducing the transducer module and a limiting condition into the generator; and introducing a transducer module into the discrimination network structure, and introducing the LSGAN least squares loss function into the sea clutter simulation model.
7) And inputting the sea clutter data set and the sea surface geometric image set into a sea clutter simulation model, outputting a sea clutter image after the sea surface geometric image set passes through a generator, judging the similarity of the sea clutter image and the sea clutter data set by a discriminator, performing iterative training for a plurality of times until an optimized model of the sea clutter simulation model is obtained, and outputting the sea clutter image corresponding to the optimized model.
Further, the generator network U-net structure sequentially comprises a coding network, a conversion network and a decoding network, wherein the conversion network comprises a residual error module, the coding network, the conversion network and the decoding network are a group of image processing paths, and sea surface geometric image sets output sea clutter images after passing through the image processing paths.
Further, in the generator, the transducer module comprises a fusion transducer jump connection module and a transducer cross-layer fusion module; the image processing paths are provided with a plurality of groups, and the adjacent two groups are added through a trans-former cross-layer fusion module to perform path fusion; each group of convolution kernels are different in size, each group comprises two branch paths, one branch path is formed by adding a fusion converter jump connection module into the conversion network, the other branch path is formed by not adding a fusion converter jump connection module into the conversion network, and the two branch paths are fused in the coding network and the decoding network.
Further, the sea area weather data is selected from ERA5 analysis data; the method comprises wind speed, wind direction, effective wave height of the surging waves, wave direction, wave period, relevant waves of the stormy waves, relevant waves of the mixed waves, temperature and sounding models.
Further, the sea surface simulation model is built on the basis of a linear superposition method by adopting a wind sea spectrum model and a surge spectrum model, and a real sea environment is input into a transfer function in a parameter form; the transfer function is a unilateral cosine direction distribution function.
Further, the sea surface simulation data characteristic parameters comprise wind and wave data; the sea echo data characteristic parameters include amplitude characteristics, spectral characteristics and texture characteristics.
Further, in step 2), the preprocessing is: image segmentation, denoising, filtering and standardization.
Further, the number of the discriminators is two, and the convolution kernel sizes are different, namely a high-resolution discriminator 1D and a low-resolution discriminator 2D.
Further, the LSGAN least squares loss function has the following formulas for the generator and the arbiter, respectively:
；
；
wherein D is a discriminator;is a discriminator loss function; g is a generator; />Is a generator loss function; x is a sea clutter image; z is the input sea surface geometric image and adds a limiting condition; e is a desired value; />Is the probability distribution obeyed by the sea clutter image; />Is the probability distribution obeyed by z; />Representing the result of the generator on the input data z;representing the discrimination result of the discriminator on the sea clutter data set; />The more the output is close to 1, the greater the probability that the input is considered to be real data by the discriminator is; a is a constant, representing a mark of a real image; b is a constant representing a mark for generating an image; c is the generation ofA value that is determined by the arbiter in order for the arbiter to consider the generated image as real data; a=c=1, b=0.
In order to achieve the purpose, the satellite-borne SAR sea clutter generation system based on the multi-scale attention mechanism can adopt the following technical scheme:
a satellite-borne SAR sea clutter generation system based on a multiscale attention mechanism generates a sea clutter image based on a sea clutter simulation model, wherein the sea clutter simulation model comprises a generator and a discriminator, the generator processes the image based on a network U-net structure and a limiting condition, and the discriminator processes the image based on a discriminating network structure; the system comprises a data acquisition module, a sea surface geometric image generation module, a model condition preparation module, a sea clutter simulation model module and an optimization generation module, and specifically comprises the following components:
the data acquisition module is used for acquiring sea area meteorological data and sea clutter data under the meteorological conditions, wherein the sea clutter data are selected from SAR images; the sea surface geometric image generation module is used for preprocessing SAR images and matching the SAR images with sea area meteorological data to form sea clutter data sets and meteorological data sets of the same sea area; the model condition preparation module is used for establishing a sea surface simulation model, and inputting a meteorological data set into the sea surface simulation model after being in a unified format to generate a sea surface geometric image set; the limiting condition generation module is used for extracting the characteristics of the sea surface geometric image set and the sea clutter data set to respectively form sea surface simulation data characteristic parameters and sea surface echo data characteristic parameters which are used as limiting conditions of the generator; the function building module is used for building an LSGAN least squares loss function; the model integration module is used for establishing a sea clutter simulation model based on a multi-scale attention mechanism and the transducer module, and introducing the transducer module and the limiting conditions into the generator; introducing a transducer module into a discrimination network structure, and introducing an LSGAN least squares loss function into a sea clutter simulation model; the model optimization module is used for inputting the sea clutter data set and the sea surface geometric image set into the sea clutter simulation model, outputting a sea clutter image after the sea surface geometric image set passes through the generator, judging the similarity of the sea clutter image and the sea clutter data set by the discriminator, performing iterative training for many times until an optimized model of the sea clutter simulation model is obtained, and outputting the sea clutter image corresponding to the optimized model.
The application has the following beneficial effects: 1. the method establishes the sea clutter simulation model based on the multiscale and the attention, forms the image processing paths of the coding network, the conversion network and the decoding network, reduces network model parameters, reduces complex operation and simplifies operation under the condition of ensuring the size of the receptive field. 2. According to the application, sea area meteorological data and SAR original images are processed and matched to form data pairs, the influence of sea condition parameters such as wind speed, wind direction, effective wave height, wave direction and the like on scattered wave signals is considered, and simultaneously, a transducer module is introduced into a generator and a discriminator, so that the precision is high. 3. the transducer module captures the characteristics in the previous step of network through a self-attention mechanism and transmits the characteristics to the network of the next stage, so that better associated position information is achieved, the model is facilitated to better understand the structure and sequence of the characteristic diagram, the performance and the robustness of the sea clutter simulation model are improved, and the timeliness is high.
Drawings
FIG. 1 is a schematic diagram of a network architecture of a generator based on a multi-scale attention mechanism of the present application;
FIG. 2 is a schematic diagram of a sea clutter modeling model according to the present application;
fig. 3 is a schematic diagram of a network structure of a multi-scale and attention mechanism-based discriminator according to the application.
Detailed Description
The present application is further illustrated in the accompanying drawings and detailed description which are to be construed as exemplary only and not limiting the scope of the application, since modifications to the application in the art will fall within the scope of the application which is defined by the appended claims after reading the application.
In the embodiment, nineteen ERA5 analysis data in the sea area of the middle China in the last twenty years are obtained, a correlation model of ocean environment parameters and sea clutter characteristics is established based on a regression analysis module of a random forest model, 4-5 longitude and latitude points of the Bohai sea, the yellow sea, the east sea and the south sea are selected, and regression analysis is carried out on a plurality of large data amount tables such as 20 years, four seasons, 0 points, 6 points, 12 points and 18 points, and the like, so that influence factors of the sea clutter characteristics under different ocean parameters in different sea areas and different times are analyzed. Analyzing the physical characteristics of the ocean surface helps to build a more accurate sea clutter model. The application is implemented in a computer having the following configuration: intel i5-12400F 12 core processor, nvidia GeForce RTX3060 graphics processor, main frequency 2.50GHz, memory 16GB, and operating system is windows10. The implementation of satellite-borne sar sea clutter forward modeling is based on Tensorflow2.0 deep learning framework toolkit.
Referring to fig. 2, in the method for generating the sea clutter by the satellite-borne SAR based on the multi-scale attention mechanism, a sea clutter image is generated based on a sea clutter simulation model, the sea clutter simulation model comprises a generator and a discriminator, the generator processes the image based on a network U-net structure and a limiting condition, and the generator generates the image and then evaluates the image through the discriminator. The discriminator processes the image based on discriminating the network structure, specifically: the discriminator receives an input image, the image being a sea clutter dataset and an output image of the generator, representing a real image sample and a generated false image sample, respectively. The discriminators process the image and the convolution layer, transform module converts the input image into a series of feature maps for capturing local and global information of the image. For large-size images generated by the generator, the arbiter gradually reduces the size of the feature map through a convolution and pooling operation. The final goal of the arbiter is to output a probability value that indicates the probability that the generator generated image is a true image. The arbiter maps the feature map to a probability value through the full connection layer.
The generating method comprises the following steps:
1) And acquiring sea area meteorological data and sea clutter data under the meteorological conditions, wherein the sea clutter data are selected from SAR images. Wherein the sea area meteorological data is selected from ERA5 analysis data; the method comprises wind speed, wind direction, effective wave height of the surging waves, wave direction, wave period, relevant waves of the stormy waves, relevant waves of the mixed waves, temperature and sounding models. The SAR image is high-resolution SAR image data in recent years in the sea areas of China (Bohai sea, yellow sea, east sea and south sea), and the data sources are real-1A and high-resolution No. three. Nineteen ERA5 analysis data in the sea area (Bohai sea, yellow sea, east sea and south sea) of the middle-aged and twenty years are obtained, sea area meteorological data with different resolutions are matched based on an interpolation prediction module in a random forest model, space-time difference of the data is unified, and a sea area meteorological data unified format is completed.
2) And preprocessing SAR images, and matching the SAR images with sea area meteorological data to form a sea clutter data set and a sea area meteorological data set of the same sea area. The preprocessing process comprises image segmentation, denoising, filtering and standardization. Firstly, image segmentation is carried out, so that subsequent processing is facilitated. And secondly, denoising processing is carried out, and a complete scattered wave signal is reserved. A filtering operation is then performed to eliminate the uncertain noise and to increase the frequency domain resolution. And finally, carrying out standardization processing, and converting different dimensions of the hydrological parameters and the scattered wave signals into uniform specifications so as to match modeling requirements of the neural network. In this embodiment, professional software such as snap is used for preprocessing. And matching the SAR image with ERA5 analysis data according to the image time and the geocode to obtain a sea clutter data set. And dividing the sea clutter data set into a training set and a verification set, and training and verifying the sea clutter simulation model to obtain an optimized model of the sea clutter simulation model.
3) Establishing a sea surface simulation model, and inputting a meteorological data set into the sea surface simulation model after unifying the format to generate a sea surface geometric image set. The sea surface simulation model is established based on a linear superposition method by adopting a wind sea spectrum model and a surge spectrum model, and a real sea environment is input into a transfer function in a parameter form; the transfer function is a unilateral cosine direction distribution function.
Specifically, a sea surface simulation model is established based on a linear superposition method by adopting an elfouhaily wind sea spectrum model and a jonswap surge spectrum model, and a real sea environment is input into a transfer function in a parameter form (wind field data and wave data); the transfer function is a unilateral cosine direction distribution function. Through PM spectra of different wind elements and JONSWAP gravity spectra, elfouhaity tension spectra of different wind speeds and wind areas, and an elfouhaity spectrum unilateral cosine direction distribution function, linear geometric sea surface simulation modeling is completed.
4) Extracting features of the sea surface geometric image set and the sea clutter data set to form sea surface simulation data feature parameters and sea surface echo data feature parameters respectively, and taking the sea surface simulation data feature parameters and the sea surface echo data feature parameters as limiting conditions of a generator;
by providing the sea surface simulation data characteristic parameter and the sea surface echo data characteristic parameter as constraints to the generator, the properties of the generated image can be accurately controlled. The generator will try to generate images matching these characteristic parameters, ensuring that the generated SAR sea clutter images meet the actual conditions under specific meteorological conditions. The sea surface simulation data characteristic parameters comprise wind and wave data; the sea echo data characteristic parameters include amplitude characteristics, spectral characteristics and texture characteristics. The sea surface simulation data characteristic parameters and the sea surface echo data characteristic parameters are output in the form of a histogram, and the histogram is used for analyzing the reflection intensity, the frequency components and the texture structure of the sea surface, so that important reference information is provided for marine meteorological research and data analysis, and the characteristic distribution situation of sea surface SAR images can be better understood and visualized. Specifically, the amplitude characteristic histogram describes intensity distribution of sea surface reflection, and is used for correcting and denoising SAR images so as to reduce influence of clutter and noise on meteorological parameter estimation. The frequency spectrum characteristic histogram describes the distribution of frequency components reflected by the sea surface, and analyzes the frequency components and the energy distribution of waves on the sea surface, so that meteorological parameters such as wave height, wave length and the like are extracted. The texture feature histogram describes the texture structure of sea surface reflection and is used for identifying and classifying different sea surface phenomena and marine meteorological phenomena, such as sea waves, ripples, storm vortexes and the like. The function of the device is as follows: (1) enhancing the authenticity of the generated image: the introduction of the constraint can make the generated image more realistic and reasonable. The generator needs to generate an image according to the characteristic parameters of the sea surface simulation data and the characteristic parameters of the sea surface echo data, so that the generated image can better simulate the image in the real SAR sea clutter data set. (2) Enhancing the generation of specific attributes: the generated sea clutter image has specific meteorological attributes, such as specific stormy waves, and the fact that the generated image is matched with the attributes can be ensured by introducing sea surface simulation data characteristic parameters as limiting conditions. (3) Avoiding meaningless generation: the constraint may help the generator avoid generating meaningless images. If the generator is limited only to specific sea surface simulation data characteristic parameters and echo data characteristic parameters, it will generate meaningful images within a reasonable range without generating non-realistic images. (4) The consistency of the generation is improved: since the constraints provide additional information during the generation process, the generator is more capable of generating images consistent with these conditions. This may help the generated image better reflect the actual situation under specific weather conditions.
5) And establishing an LSGAN least squares loss function. The built LSGAN least square loss function is used for introducing a sea clutter simulation model. For generator G, its loss function generates a square difference between the generated sample and the real sample, i.e. minimizes the euclidean distance between the generated sample and the real sample. For the arbiter D, its loss function is the square difference between the real sample and the generated sample, which means that the arbiter tries to minimize the distance between the real sample and 1, and the distance between the generated sample and 0. The LSGAN least squares loss function has the following formulas for the generator and the arbiter, respectively:
；
；
wherein D is a discriminator;is a discriminator loss function; g is a generator; />Is a generator loss function; x is a sea clutter image; z is the input sea surface geometric image and adds a limiting condition; e is a desired value; />Is the probability distribution obeyed by the sea clutter image; />Is the probability distribution obeyed by z; />Representing the result of the generator on the input data z;representing the discrimination result of the discriminator on the sea clutter data set; />The more the output is close to 1, the greater the probability that the input is considered to be real data by the discriminator is; a is a constant, representing a mark of a real image; b is a constant representing a mark for generating an image; c is a value that the generator makes the arbiter consider the generated image as real data; a=c=1, b=0.
6) Establishing a sea clutter simulation model based on a multi-scale attention mechanism and a transducer module, and introducing the transducer module and a limiting condition into the generator; and introducing a transducer module into the discrimination network structure, and introducing the LSGAN least squares loss function into the sea clutter simulation model.
The generator processes images based on a network U-net structure and limiting conditions, the generator network U-net structure sequentially comprises a coding network, a conversion network and a decoding network, the conversion network comprises a residual error module, the coding network, the conversion network and the decoding network are a group of image processing paths, and sea surface geometric image sets output sea clutter images after passing through the image processing paths. In the generator, the transducer module comprises a fusion transducer jump connection module and a transducer cross-layer fusion module; the image processing paths are provided with a plurality of groups, and the adjacent two groups are added through a trans-former cross-layer fusion module to perform path fusion; each group of convolution kernels are different in size, each group comprises two branch paths, one branch path is formed by adding a fusion converter jump connection module into the conversion network, the other branch path is formed by not adding a fusion converter jump connection module into the conversion network, and the two branch paths are fused in the coding network and the decoding network. The generator and the arbiter of the optimized sea clutter simulation model are specifically described below.
Referring to fig. 1, in step 3), the coding network is formed by five convolution layers, and the output of the coding network is used as the input of the conversion network; the conversion network is a residual network, the output of the input after two convolution layers is added with the output to obtain the output of the residual network layer, the output is used as the input of the next residual network layer, and five residual network layers form the residual network. The decoding network at the end of the generator consists of four deconvolution layers and one 2D convolution layer, the last convolution layer outputting the generated sea clutter. The network structure is composed of "convolution (C) -instance regularization (IN) -ReLU activation function" units.
And introducing a fusion transformer jump connection module, and constructing jump connection on the current group of image processing paths. And splicing one characteristic diagram obtained in the encoding network stage with the current decoding network through two parallel circuits, and splicing the other characteristic diagram with the current decoding network after entering a trans-former cross-layer fusion module. The coding network uses four fusion transformation former jump connection modules for capturing shallow, low-level and fine characteristic images in a plurality of groups of coding networks and associated characteristic images captured by corresponding groups of transformation former modules and cross-layer fusion modules, and fusion of characteristic information with more scales and depth supervision are carried out, so that the performance and robustness of the model are further improved. And introducing a transform cross-layer fusion module into the previous group of the current group, downsampling the output of the transform cross-layer fusion module of the previous group to the same size of the output characteristic diagram of the current group of the transformers through convolution and pooling, adding the two, and splicing the two with a decoding network and a coding network corresponding to the current layer.
Referring to fig. 3, a discrimination network based on a multi-scale and attention mechanism is established, and a transducer module is added to the discriminator network to extract global features. Through feature methods such as multi-scale sensing, contextual information introduction, local feature learning, computational efficiency and the like, a more comprehensive, fine and efficient image discrimination capability is provided, making it advantageous in applications in an antagonism generation network. By introducing two discriminators with the same structure but different input and output sizes, the two discriminators are made to be a high-resolution discriminator 1D and a low-resolution discriminator 2D. Assessing the authenticity of the generated image from different angles and scales contributes to the stability of the resistance training. One of the discriminators focuses on the authenticity of the overall image, while the other discriminators focuses on the authenticity of the local detail. The method can capture the characteristics and details of different scales, so that the detection precision and capability of the discriminator on the generated image can be improved, and the authenticity of the image can be accurately judged. The design of the discrimination network incorporates a PatchGAN structure and sets Patch to 64. For the Patchgan arbiter network, for 256x256 images and 32x32 Patches, the partition is 8x8 Patches, for a total of 64 Patches. The use of the multi-scale judgment network can reduce network model parameters, optimize memory and remarkably promote model training under the condition of ensuring the size of the receptive field.
7) And inputting the sea clutter data set and the sea surface geometric image set into a sea clutter simulation model, outputting a sea clutter image after the sea surface geometric image set passes through a generator, judging the similarity of the sea clutter image and the sea clutter data set by a discriminator, performing iterative training for a plurality of times until an optimized model of the sea clutter simulation model is obtained, and outputting the sea clutter image corresponding to the optimized model.
In training, the arbiter receives input from the real image and the false image generated by the generator and outputs a probability value between 0 and 1 representing the probability that the input image is a real image. For a real image, the objective of the arbiter is to output a probability value close to 1, while for a generated image, the objective of the arbiter is to output a probability value close to 0. Meanwhile, the generation process of the generator is opposite to the output of the discriminator. The generator and the arbiter play games with each other, and the generator generates images and then evaluates the images through the arbiter. The goal of the generator is that the generated image can fool the arbiter so that it can be determined as a true image by the arbiter. The countering loss causes the discriminator to strive to bring the discrimination probability of the generated image close to 0, while the discrimination probability of the real image approaches 1. The generator will then strive to generate an image that can spoof the arbiter. Iterative training: the arbiter and the generator play games with each other, and iterative training is continuously performed. The generator tries to generate a realistic image, while the arbiter strives to improve the accuracy of distinguishing between the realistic and the generated image.
The optimization process comprises the following steps:
(1) firstly training the discriminators, and judging whether the two discriminator networks are real image data or the image subjected to style conversion, and optimizing the weight of the discriminator networks through countermeasures. The setting of the weights affects the classification decision boundary of the arbiter and thus its discrimination capability between the generated data and the real data. The weights of the discriminators are optimally updated by back propagation and optimization algorithms. Forward propagation: given a real data sample and a sample generated by the generator, the discriminator transmits the two samples to the forward propagation respectively, and calculates the corresponding discrimination score. The discrimination score represents the likelihood that the sample is discriminated as real data. Calculating loss: using the discrimination score and the corresponding tag (tag 1 for real data and tag 0 for generated data), the loss of the discriminator is calculated. Back propagation: and calculating the gradient of each parameter in the discriminator through a back propagation algorithm according to the loss value in the calculated loss. These gradients represent the rate of change of the loss function with respect to the parameter, i.e. the direction of adjustment of the parameter. Optimization algorithm: and updating the weight of the discriminator through a random gradient descent optimization algorithm according to the calculated gradient. The value of each parameter is updated according to the direction of the gradient and the learning rate to reduce the loss function.
(2) And training the generator, and optimizing the weight of the generator according to the characteristic matching loss and the countering loss. The generator weight optimization updating steps are as follows. Back propagation: gradients of the various parameters in the discriminators are calculated by a back propagation algorithm based on the generator loss function. How to adjust the weight of the generator can reduce the loss. Optimization algorithm: and updating the weight of the discriminator through a random gradient descent optimization algorithm according to the calculated gradient. The value of each parameter is updated according to the direction of the gradient and the learning rate to reduce the loss function. Iterative training: repeating the steps, and gradually adjusting the weight of the generator through repeated iterative training, so that the quality of the generated image is continuously improved. The effect of the generator weights continuously optimizing updates is as follows: the generator network parameters are adjusted to make the generated image more realistic, thereby spoofing the arbiter.
(3) And (3) continuously repeating the step (1) and the step (2) by the network model, and finishing model training until the maximum iteration times are reached. (4) And outputting the generated sea clutter image by inputting the sea surface geometric image.
Wherein, the weight of the discriminator and the weight of the generator are further described. Along with the optimization updating of the discriminator weight and the generator weight in an iterative mode, the corresponding generated network loss and the discrimination network loss show different trends. Generating network loss: the goal of generating a network is to produce a high quality image that can fool the discrimination network, making it difficult to distinguish between the generated image and the real image. Therefore, the loss function of the generated network is higher in the early stage of training, and the loss of the generated network gradually decreases as the generator gradually learns better image conversion. Judging network loss: the goal of the discrimination network is to accurately distinguish between the generated image and the actual image. In the early stage of training, the discrimination network can easily discriminate the generated image, so the loss of the discrimination network can be relatively low. As generators continue to improve, it becomes more difficult to identify networks, and therefore the loss of identification networks gradually increases until an equilibrium state is reached. In the comprehensive view, as training proceeds, the generated network loss gradually decreases, and the network loss gradually increases. This reflects the dynamic balance between the generator and the arbiter, and the evolution of the sea clutter simulation model into an optimization model when learning a stable image conversion strategy.
The application also discloses a satellite-borne SAR sea clutter generation system based on the multi-scale attention mechanism, which is used for establishing a sea clutter simulation model based on the multi-scale attention mechanism and the transducer module, wherein the sea clutter simulation model comprises a generator and a discriminator, the generator processes images based on a network U-net structure and a limiting condition, and the discriminator processes images based on a discriminating network structure; the system comprises: the data acquisition module is used for acquiring sea area meteorological data and sea clutter data under the meteorological conditions, wherein the sea clutter data are selected from SAR images; the sea surface geometric image generation module is used for preprocessing SAR images and matching the SAR images with sea area meteorological data to form sea clutter data sets and meteorological data sets of the same sea area; the model condition preparation module is used for establishing a sea surface simulation model, and inputting a meteorological data set into the sea surface simulation model after being in a unified format to generate a sea surface geometric image set; the limiting condition generation module is used for extracting the characteristics of the sea surface geometric image set and the sea clutter data set to respectively form sea surface simulation data characteristic parameters and sea surface echo data characteristic parameters which are used as limiting conditions of the generator; the function building module is used for building an LSGAN least squares loss function; the model integration module is used for establishing a sea clutter simulation model based on a multi-scale attention mechanism and the transducer module, and introducing the transducer module and the limiting conditions into the generator; introducing a transducer module into a discrimination network structure, and introducing an LSGAN least squares loss function into a sea clutter simulation model; the model optimization module is used for inputting the sea clutter data set and the sea surface geometric image set into the sea clutter simulation model, outputting a sea clutter image after the sea surface geometric image set passes through the generator, judging the similarity of the sea clutter image and the sea clutter data set by the discriminator, performing iterative training for many times until an optimized model of the sea clutter simulation model is obtained, and outputting the sea clutter image corresponding to the optimized model.
In addition, in the model application, it may be found that the performance of the model still cannot meet the actual requirements. The present application thus employs model deployment and model monitoring strategies.
The model application comprises the following specific steps: after model optimization is completed, the model may be deployed into a production environment, such as an embedded device or cloud server. The method comprises the following specific steps: converting the trained sea surface simulation model and the sea clutter simulation model into executable codes, such as a TensorFlow Lite format; and deploying the converted model on target equipment or a server so as to realize functions of modeling, removing and the like of sea clutter in real time. Model monitoring, after model deployment, the model needs to be monitored to ensure the performance and accuracy of the model. The method comprises the following specific steps: and carrying out periodic performance and accuracy tests on the model to detect whether the model has the problems of over fitting or under fitting and the like. And according to the test result, adjusting the model parameters and the structure to improve the performance and the accuracy of the model. The model is updated in time to accommodate new data and scene requirements.
In summary, the satellite-borne SAR sea clutter generation method and system based on the multi-scale attention mechanism have the characteristics of high precision, strong timeliness and simplified operation.
Claims (10)
1. A satellite-borne SAR sea clutter generation method based on a multiscale attention mechanism is characterized in that a sea clutter image is generated based on a sea clutter simulation model, the sea clutter simulation model comprises a generator and a discriminator, the generator processes the image based on a network U-net structure and a limiting condition, and the discriminator processes the image based on a discrimination network structure; the generating method comprises the following steps:
1) Acquiring sea area meteorological data and sea clutter data under meteorological conditions, wherein the sea clutter data are selected from SAR images;
2) Preprocessing SAR images, and matching the SAR images with sea area meteorological data to form a sea clutter dataset and a sea area meteorological dataset of the same sea area;
3) Establishing a sea surface simulation model, and inputting a meteorological data set into the sea surface simulation model after unifying the format to generate a sea surface geometric image set;
4) Extracting features of a sea surface geometric image set and a sea clutter data set to respectively form sea surface simulation data feature parameters and sea surface echo data feature parameters, and taking the sea surface simulation data feature parameters and the sea surface echo data feature parameters as limiting conditions of a generator;
5) Establishing an LSGAN least squares loss function;
6) Establishing a sea clutter simulation model based on a multi-scale attention mechanism and a transducer module, and introducing the transducer module and a limiting condition into the generator; introducing a transducer module into a discrimination network structure, and introducing an LSGAN least squares loss function into a sea clutter simulation model;
7) And inputting the sea clutter data set and the sea surface geometric image set into a sea clutter simulation model, outputting a sea clutter image after the sea surface geometric image set passes through a generator, judging the similarity of the sea clutter image and the sea clutter data set by a discriminator, performing iterative training for a plurality of times until an optimized model of the sea clutter simulation model is obtained, and outputting the sea clutter image corresponding to the optimized model.
2. The method for generating sea clutter based on the multiscale attention mechanism according to claim 1, wherein the generator network U-net structure sequentially comprises an encoding network, a converting network and a decoding network, the converting network comprises a residual module, the encoding network, the converting network and the decoding network are a group of image processing paths, and sea surface geometric image sets output sea clutter images after passing through the image processing paths.
3. The method for generating the sea clutter based on the multiscale attention mechanism according to claim 2, wherein in the generator, a transducer module comprises a fusion transducer jump connection module and a transducer cross-layer fusion module; the image processing paths are provided with a plurality of groups, and the adjacent two groups are added through a trans-former cross-layer fusion module to perform path fusion; each group of convolution kernels are different in size, each group comprises two branch paths, one branch path is formed by adding a fusion converter jump connection module into the conversion network, the other branch path is formed by not adding a fusion converter jump connection module into the conversion network, and the two branch paths are fused in the coding network and the decoding network.
4. The method for generating sea clutter based on the multiscale attentiveness mechanism according to claim 1, wherein the sea area meteorological data is selected from ERA5 analysis data; the method comprises wind speed, wind direction, effective wave height of the surging waves, wave direction, wave period, relevant waves of the stormy waves, relevant waves of the mixed waves, temperature and sounding models.
5. The method for generating the sea clutter of the spaceborne SAR based on the multi-scale attention mechanism according to claim 1, wherein the sea surface simulation model is established based on a linear superposition method by adopting a wind sea spectrum model and a surge spectrum model, and a real sea environment is input into a transfer function in a parameter form; the transfer function is a unilateral cosine direction distribution function.
6. The method for generating sea clutter based on the multiscale attentiveness mechanism according to claim 5, wherein the sea surface simulation data characteristic parameters comprise wind and wave data; the sea echo data characteristic parameters include amplitude characteristics, spectral characteristics and texture characteristics.
7. The method for generating sea clutter in a satellite SAR based on a multi-scale attention mechanism according to claim 1, wherein in step 2), the preprocessing is: image segmentation, denoising, filtering and standardization.
8. The method for generating sea clutter based on the multiscale attention mechanism according to claim 1, wherein the number of the discriminators is two, and the convolution kernels are different in size and are respectively a high-resolution discriminator 1D and a low-resolution discriminator 2D.
9. The method for generating sea clutter in a satellite-borne SAR based on a multi-scale attention mechanism according to claim 1, wherein the LSGAN least squares loss function has the following formulas for the generator and the arbiter, respectively:
；
；
wherein D is a discriminator;is a discriminator loss function; g is a generator; />Is a generator loss function; x is a sea clutter image; z is the input sea surface tableAdding a limiting condition to the image; e is a desired value; />Is the probability distribution obeyed by the sea clutter image; />Is the probability distribution obeyed by z; />Representing the result of the generator on the input data z; />Representing the discrimination result of the discriminator on the sea clutter data set; />The more the output is close to 1, the greater the probability that the input is considered to be real data by the discriminator is; a is a constant, representing a mark of a real image; b is a constant representing a mark for generating an image; c is a value that the generator makes the arbiter consider the generated image as real data; a=c=1, b=0.
10. A satellite-borne SAR sea clutter generation system based on a multiscale attention mechanism is characterized in that a sea clutter image is generated based on a sea clutter simulation model, the sea clutter simulation model comprises a generator and a discriminator, the generator processes the image based on a network U-net structure and a limiting condition, and the discriminator processes the image based on a discrimination network structure; the system comprises: the data acquisition module is used for acquiring sea area meteorological data and sea clutter data under the meteorological conditions, wherein the sea clutter data are selected from SAR images; the sea surface geometric image generation module is used for preprocessing SAR images and matching the SAR images with sea area meteorological data to form sea clutter data sets and meteorological data sets of the same sea area; the model condition preparation module is used for establishing a sea surface simulation model, and inputting a meteorological data set into the sea surface simulation model after being in a unified format to generate a sea surface geometric image set; the limiting condition generation module is used for extracting the characteristics of the sea surface geometric image set and the sea clutter data set to respectively form sea surface simulation data characteristic parameters and sea surface echo data characteristic parameters which are used as limiting conditions of the generator; the function building module is used for building an LSGAN least squares loss function; the model integration module is used for establishing a sea clutter simulation model based on a multi-scale attention mechanism and the transducer module, and introducing the transducer module and the limiting conditions into the generator; introducing a transducer module into a discrimination network structure, and introducing an LSGAN least squares loss function into a sea clutter simulation model; the model optimization module is used for inputting the sea clutter data set and the sea surface geometric image set into the sea clutter simulation model, outputting a sea clutter image after the sea surface geometric image set passes through the generator, judging the similarity of the sea clutter image and the sea clutter data set by the discriminator, performing iterative training for many times until an optimized model of the sea clutter simulation model is obtained, and outputting the sea clutter image corresponding to the optimized model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311486202.7A CN117217103B (en) | 2023-11-09 | 2023-11-09 | Satellite-borne SAR sea clutter generation method and system based on multi-scale attention mechanism |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311486202.7A CN117217103B (en) | 2023-11-09 | 2023-11-09 | Satellite-borne SAR sea clutter generation method and system based on multi-scale attention mechanism |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117217103A true CN117217103A (en) | 2023-12-12 |
| CN117217103B CN117217103B (en) | 2024-03-15 |
Family
ID=89039393
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311486202.7A Active CN117217103B (en) | 2023-11-09 | 2023-11-09 | Satellite-borne SAR sea clutter generation method and system based on multi-scale attention mechanism |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117217103B (en) |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104318593A (en) * | 2014-09-30 | 2015-01-28 | 北京环境特性研究所 | Simulation method and system of radar sea clusters |
| WO2020136258A1 (en) * | 2018-12-27 | 2020-07-02 | Thales | Device for generating a simulated sea-clutter data set, and associated method and computer program |
| CN111626961A (en) * | 2020-05-29 | 2020-09-04 | 中国人民解放军海军航空大学 | Radar image clutter suppression method and system based on generation countermeasure network |
| CN113449850A (en) * | 2021-07-05 | 2021-09-28 | 电子科技大学 | Intelligent inhibition method for clutter of sea surface monitoring radar |
| CN114117912A (en) * | 2021-11-26 | 2022-03-01 | 东南大学 | Sea clutter modeling and inhibiting method under data model dual drive |
| CN114492744A (en) * | 2022-01-18 | 2022-05-13 | 西北工业大学 | Method for generating ground-sea clutter spectrum data sample based on confrontation generation network |
| CN114970378A (en) * | 2022-08-01 | 2022-08-30 | 青岛国数信息科技有限公司 | Sea clutter sample library construction method based on GAN network |
| CN115859167A (en) * | 2022-11-29 | 2023-03-28 | 西北工业大学 | Sky wave radar ground-sea clutter semi-supervised classification model construction method for improving generation of countermeasure network |
| CN116755090A (en) * | 2023-06-12 | 2023-09-15 | 南京工业大学 | SAR ship detection method based on novel pyramid structure and mixed pooling channel attention mechanism |
-
2023
- 2023-11-09 CN CN202311486202.7A patent/CN117217103B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104318593A (en) * | 2014-09-30 | 2015-01-28 | 北京环境特性研究所 | Simulation method and system of radar sea clusters |
| WO2020136258A1 (en) * | 2018-12-27 | 2020-07-02 | Thales | Device for generating a simulated sea-clutter data set, and associated method and computer program |
| CN111626961A (en) * | 2020-05-29 | 2020-09-04 | 中国人民解放军海军航空大学 | Radar image clutter suppression method and system based on generation countermeasure network |
| CN113449850A (en) * | 2021-07-05 | 2021-09-28 | 电子科技大学 | Intelligent inhibition method for clutter of sea surface monitoring radar |
| CN114117912A (en) * | 2021-11-26 | 2022-03-01 | 东南大学 | Sea clutter modeling and inhibiting method under data model dual drive |
| CN114492744A (en) * | 2022-01-18 | 2022-05-13 | 西北工业大学 | Method for generating ground-sea clutter spectrum data sample based on confrontation generation network |
| CN114970378A (en) * | 2022-08-01 | 2022-08-30 | 青岛国数信息科技有限公司 | Sea clutter sample library construction method based on GAN network |
| CN115859167A (en) * | 2022-11-29 | 2023-03-28 | 西北工业大学 | Sky wave radar ground-sea clutter semi-supervised classification model construction method for improving generation of countermeasure network |
| CN116755090A (en) * | 2023-06-12 | 2023-09-15 | 南京工业大学 | SAR ship detection method based on novel pyramid structure and mixed pooling channel attention mechanism |
Non-Patent Citations (2)
| Title |
|---|
| SHANSHAN TIAN 等: "Target Detection in Sea Clutter with Transformer Neural Network", 2021 CIE INTERNATIONAL CONFERENCE ON RADAR, pages 1436 - 1440 * |
| 李云龙 等: "基于K分布雷达海杂波建模与仿真", 制导与引信, vol. 32, no. 02, pages 14 - 19 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117217103B (en) | 2024-03-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110555841B (en) | SAR image change detection method based on self-attention image fusion and DEC | |
| CN106355151A (en) | Recognition method, based on deep belief network, of three-dimensional SAR images | |
| Chen et al. | Spatial–temporal convolutional gated recurrent unit network for significant wave height estimation from shipborne marine radar data | |
| Zhang et al. | A GANs-based deep learning framework for automatic subsurface object recognition from ground penetrating radar data | |
| CN113361367B (en) | Underground target electromagnetic inversion method and system based on deep learning | |
| CN112069955B (en) | Typhoon intensity remote sensing inversion method based on deep learning | |
| CN109446894A (en) | The multispectral image change detecting method clustered based on probabilistic segmentation and Gaussian Mixture | |
| Chen et al. | Rain-contaminated region segmentation of X-band marine radar images with an ensemble of SegNets | |
| Huang et al. | Wave height estimation from X-band nautical radar images using temporal convolutional network | |
| CN113920255B (en) | High-efficient mapping system based on point cloud data | |
| Wong et al. | Automatic tropical cyclone eye fix using genetic algorithm | |
| Guo et al. | Sea surface wind direction retrieval based on convolution neural network and wavelet analysis | |
| Hu et al. | A sample enhancement method based on simple linear iterative clustering superpixel segmentation applied to multibeam seabed classification | |
| CN116338628B (en) | Laser radar sounding method and device based on learning architecture and electronic equipment | |
| CN112766381A (en) | Attribute-guided SAR image generation method under limited sample | |
| CN117217103B (en) | Satellite-borne SAR sea clutter generation method and system based on multi-scale attention mechanism | |
| CN112669201A (en) | Infrared light-based visible light cloud image conversion method, system and terminal | |
| CN114511787A (en) | Neural network-based remote sensing image ground feature information generation method and system | |
| CN110991507A (en) | Road underground cavity identification method, device and system based on classifier | |
| Wang et al. | Global ocean mesoscale vortex recognition based on DeeplabV3plus model | |
| CN115861824B (en) | Remote sensing image recognition method based on improved transducer | |
| CN110135280A (en) | A kind of multiple view SAR automatic target recognition method based on sparse representation classification | |
| CN113971755B (en) | All-weather sea surface target detection method based on improved YOLOV model | |
| CN116596198B (en) | Urban green land biomass calculation method, device, electronic equipment and storage medium | |
| CN113378766B (en) | Marine large-scale wind power station monitoring system based on synthetic aperture radar |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |