Sample pose expansion method for SAR recognition database based on deep conditional diffusion network

Abstract

The present invention discloses a method for expanding the sample attitude of a SAR recognition database based on a deep conditional diffusion network, comprising: constructing a training data set using multiple real SAR images, the training data set including multiple target categories, each category including multiple image samples; constructing a deep conditional diffusion network, the deep conditional diffusion network including a time step encoding module, a category label encoding module, an azimuth encoding module, a coding fusion module, a feature map construction module, multiple downsampling modules, multiple intermediate transition modules, multiple upsampling modules and an image restoration module; training the deep conditional diffusion network using the training data set to obtain a trained deep conditional diffusion network; and generating a SAR image using the trained deep conditional diffusion network. The present invention can generate high-quality SAR images of specified categories and azimuths, thereby expanding the number and attitude of samples in the SAR recognition database.

Description

Technical Field

[0001] The present invention belongs to the field of radar technology, and in particular relates to a method for expanding the sample attitude of a SAR recognition database based on a deep conditional diffusion network. The method can generate high-quality SAR images of specified categories and azimuths in order to solve the problems of insufficient number of database samples and incomplete attitudes in SAR target detection and recognition applications, thereby expanding the number and attitude of SAR recognition database samples. Background Art

[0002] High-resolution SAR images have rich texture and detail information, and have the advantages of being acquired all day and not affected by external weather conditions. They are widely used in remote sensing target recognition and detection. Due to the widespread application of deep neural networks in the field of target detection and recognition, the demand for the number of training sample sets for various deep learning methods is gradually increasing.

[0003] However, obtaining a large number of manually annotated SAR images in real scenes is an expensive, time-consuming and arduous task, so in practice, artificial data expansion is often used. There are three main technical means to expand SAR image datasets: the first is a method based on physical simulation models, which has a relatively complex calculation process and consumes a lot of memory space; the second is a data enhancement method that cannot significantly increase the diversity of samples, such as data translation, shearing, rotation, scaling, adding noise, etc.; the third is an image generation algorithm based on a generative adversarial network, which has problems such as unstable training and gradient disappearance, and it is difficult to generate high-quality SAR images.

[0004] Xidian University proposed an attribute-guided SAR image generation method in its patent application "Attribute-guided SAR image generation method under limited samples" (patent application number: CN202110085648.3). The specific steps of this method are: constructing a training set; constructing a generator network; constructing a discriminator network; constructing an attribute-guided generative adversarial network; training an attribute-guided generative adversarial network; using the trained attribute-guided generative adversarial network to generate SAR simulation images. This method uses a generative adversarial network to generate SAR images of specified categories and azimuth angles, which improves the realism of the generated SAR images. However, this method still has shortcomings: since this method constructs a generator and a discriminator, and uses adversarial training to train the generative adversarial network, the adversarial training method can easily lead to instability in the training process of the generative adversarial network, which is specifically manifested in the poor performance of the generated SAR images, which cannot be improved even after increasing the training time.

[0005] The University of Electronic Science and Technology of China proposed a SAR image directional generation method based on a deep convolutional generative adversarial network in its patent application "A SAR image directional generation method based on a deep convolutional generative adversarial network" (patent application number 2018101674160). The specific steps of this method are: based on SAR target segmentation and edge extraction, the edge circumscribed minimum bounding rectangle method is used to calculate the clockwise angle between the north end in the vertical direction and the longest side of the minimum bounding rectangle as the azimuth of the generated SAR image target. Based on the deep convolutional generative adversarial network method, this method proposes an azimuth-driven deep convolutional generative adversarial network method, which greatly expands the SAR image data set. The shortcomings of this method are: the method is difficult to train, and it is easy to cause network gradient disappearance, mode collapse, etc., which further leads to problems such as poor quality of the generated SAR image, and it is impossible to accurately obtain SAR image samples of the required target scene. Summary of the invention

[0006] In order to solve the above problems existing in the prior art, the present invention provides a method for expanding the sample attitude of a SAR recognition database based on a deep conditional diffusion network. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0007] The present invention provides a method for expanding the sample attitude of a SAR recognition database based on a deep conditional diffusion network, comprising:

[0008] S1: construct a training data set using multiple real SAR images, where the training data set includes multiple target categories, each category includes multiple image samples, and each image sample contains a corresponding target category label and target azimuth information;

[0009] S2: construct a deep conditional diffusion network, the deep conditional diffusion network includes a time step encoding module, a category label encoding module, an azimuth encoding module, a code fusion module, a feature map construction module, multiple downsampling modules, multiple intermediate transition modules, multiple upsampling modules and an image restoration module, wherein the time step encoding module is used to map the time step of the input image to a time step code of a specified dimension; the category label encoding module is used to map the target category label corresponding to the input image to a category label code of a specified dimension; the azimuth encoding module is used to map the azimuth corresponding to the input image to an azimuth code of a specified dimension; the code fusion module is used to sum the time step code, the category label code and the azimuth code to obtain a total code;

[0010] The feature map construction module is used to extract a set of feature maps of equal size from the input image; the multiple downsampling modules are used to encode the feature maps, and each time they pass through a downsampling module, the scale of the output feature map is reduced by half, and a deep feature map is obtained after multiple operations; the multiple intermediate transition modules are used to map the deep feature map to obtain a deep transition feature map; the multiple upsampling modules are used to decode the deep transition feature map, and each time they pass through an upsampling module, the scale of the output feature map is doubled, and a decoded feature map is obtained after multiple operations; the image restoration module is used to perform channel transformation on the decoded feature map to restore it to the original input image size;

[0011] S3: training the deep conditional diffusion network using the training data set to obtain a trained deep conditional diffusion network;

[0012] S4: Generate SAR images using the trained deep conditional diffusion network.

[0013] In one embodiment of the present invention, the time step encoding module includes a first fully connected layer, a first Swish activation layer and a second fully connected layer connected in sequence; the category label encoding module includes a third fully connected layer, a second Swish activation layer and a fourth fully connected layer connected in sequence; the azimuth encoding module includes a fifth fully connected layer, a third Swish activation layer and a sixth fully connected layer connected in sequence; the feature map construction module includes an input convolutional layer.

[0014] In one embodiment of the present invention, each downsampling module includes a first GroupNorm layer, a fourth Swish activation layer, a first convolutional layer, a fifth Swish activation layer, a seventh fully connected layer, a second GroupNorm layer, a sixth Swish activation layer, a second convolutional layer, a third convolutional layer, a first fusion unit, and a second fusion unit, wherein:

[0015] The first GroupNorm layer, the fourth Swish activation layer and the first convolutional layer are connected in sequence, the fifth Swish activation layer and the seventh fully connected layer are connected in sequence, the input of the first GroupNorm layer is connected to the output of the feature map construction module, the input of the fifth Swish activation layer is connected to the output of the encoding fusion module, the output of the seventh fully connected layer and the output of the first convolutional layer are both connected to the input of the first fusion unit for addition; the second GroupNorm layer, the sixth Swish activation layer and the second convolutional layer are connected in sequence, the output of the first fusion unit is connected to the input of the second GroupNorm layer, the output of the second convolutional layer and the output of the feature map construction module are added in the second fusion unit and the result is input to the third convolutional layer, and the output of the third convolutional layer is used as the output of the downsampling module.

[0016] In one embodiment of the present invention, each intermediate transition module includes a third GroupNorm layer, a seventh Swish activation layer, a fourth convolutional layer, an eighth Swish activation layer, an eighth fully connected layer, a fourth GroupNorm layer, a ninth Swish activation layer, a fifth convolutional layer, a third fusion unit and a fourth fusion unit, wherein:

[0017] The third GroupNorm layer, the seventh Swish activation layer and the fourth convolutional layer are connected in sequence, the eighth Swish activation layer and the eighth fully connected layer are connected, the input of the third GroupNorm layer is connected to the output of the last downsampling module, the input of the eighth Swish activation layer is connected to the output of the encoding fusion module, the output of the eighth fully connected layer and the output of the fourth convolutional layer are connected to the input of the third fusion unit for addition; the fourth GroupNorm layer, the ninth Swish activation layer and the fifth convolutional layer are connected in sequence, the output of the third fusion unit is connected to the input of the fourth GroupNorm layer, the output of the fifth convolutional layer and the output of the last downsampling module are added in the fourth fusion unit and used as the output of the intermediate transition module.

[0018] In one embodiment of the present invention, each upsampling module includes a fifth GroupNorm layer, a tenth Swish activation layer, a sixth convolutional layer, an eleventh Swish activation layer, a ninth fully connected layer, a sixth GroupNorm layer, a twelfth Swish activation layer, a seventh convolutional layer, a nearest neighbor interpolation layer, an eighth convolutional layer, a fifth fusion unit, and a sixth fusion unit, wherein,

[0019] The fifth GroupNorm layer, the tenth Swish activation layer and the sixth convolutional layer are connected in sequence, the eleventh Swish activation layer and the ninth fully connected layer are connected, the input of the fifth GroupNorm layer is connected to the output of the last intermediate transition module, the input of the eleventh Swish activation layer is connected to the output of the encoding fusion module, the output of the ninth fully connected layer and the output of the sixth convolutional layer are connected to the input of the fifth fusion unit and added, and the sixth GroupNorm layer, the twelfth Swish activation layer and the seventh convolutional layer are connected in sequence, the output of the fifth fusion unit is connected to the input of the sixth GroupNorm layer, the output of the seventh convolutional layer and the output of the last intermediate transition module are added in the sixth fusion unit and the summed results are input to the nearest neighbor interpolation layer and the eighth convolutional layer in sequence, and the output of the eighth convolutional layer is used as the output of the upsampling module.

[0020] In one embodiment of the present invention, the image restoration module includes a seventh GroupNorm layer, a thirteenth Swish activation layer and a ninth convolutional layer connected in sequence.

[0021] In one embodiment of the present invention, S3 includes:

[0022] S3.1: Set the maximum inference time step of the deep conditional diffusion network to T;

[0023] S3.2: using the SAR images in the training sample set to form a small batch sample set, and adding noise to the SAR images in the small batch sample set to obtain a noisy SAR image sample corresponding to each SAR image;

[0024] S3.3: Noise-added SAR image samples corresponding to all SAR images in the small batch sample set Time step t i , category label y i 、Azimuth a i , noise map N i At the same time, it is input into the constructed deep conditional diffusion network, where t i ,y i , a i The feature map construction module, time step encoding module, category label encoding module and azimuth encoding module are input into the network respectively. The deep conditional diffusion network estimates the noise of each input noisy SAR image sample to obtain the network predicted noise map

[0025] S3.4: Using the network to predict the noise map based on the loss function With the input noise map Ni Calculate the network loss value, use the gradient descent method to calculate the gradient of each parameter to be learned in the deep conditional diffusion network and update all parameters to be learned;

[0026] S3.5: Determine whether the loss functions of the deep conditional diffusion network have converged. If so, obtain a trained deep conditional diffusion network and save the parameters of the trained deep conditional diffusion network. Otherwise, continue training.

[0027] In one embodiment of the present invention, the S3.2 includes:

[0028] S3.21: m SAR images are selected from the training data set to form a small batch sample set, where the i-th SAR image is Indicates that i = 1, 2, ..., m;

[0029] S3.22: Use a random integer generator to generate a corresponding time step t for each SAR image in the small batch sample set i , where 1≤t i ≤T,t i is an integer;

[0030] S3.23: Use a Gaussian random number generator to generate a corresponding noise map N for each SAR image in the small batch sample set i ;

[0031] S3.24: According to the noise addition formula, use the time step t corresponding to the SAR image i and noise map N i , for each SAR image Add noise to obtain the noisy SAR image samples corresponding to each SAR image Wherein, the noise addition formula is:

[0032]

[0033] in, α t =1-β t , is the i-th original SAR image, is the i-th SAR image t i The noise result at the moment, ε is the standard Gaussian white noise, β t represents the noise variance at time t.

[0034] In one embodiment of the present invention, the expression of the loss function in S3.4 is:

[0035]

[0036] Among them, L mse represents the total loss of all input samples in the mini-batch sample set, is the predicted noise map output of the deep conditional diffusion network, N i is the noise map corresponding to the i-th noisy SAR image sample.

[0037] In one embodiment of the present invention, the S4 includes:

[0038] S4.1: Set the number of SAR images to be generated to K;

[0039] S4.2: Generate H noise images using a Gaussian random number generator. The data dimension of the noise image is consistent with that of the SAR image. The value of H is equal to the number K of SAR images to be generated.

[0040] S4.3: Set the time step t for each SAR image to be generated k , where t k =T, k = 1, 2, ..., K, t k is an integer;

[0041] S4.4: The noise maps corresponding to all SAR images to be generated Time step t k , the expected generated category label y k , the desired azimuth angle a k , is fed into the trained deep conditional diffusion network, where m,y i , a i The feature map construction module, time step encoding module, category label encoding module, and azimuth encoding module are input into the network respectively to obtain the output prediction noise map;

[0042] S4.5: Update the noise map according to the image generation formula and generate the current time step t k The SAR image under the condition of FIG. 1 is shown in FIG. 1 , wherein the expression of the image generation formula is:

[0043]

[0044]

[0045] in, Indicates t k The noise variance at time, is the t corresponding to the kth SAR image to be generated k Noise map at time -1, is the t corresponding to the kth SAR image to be generated kThe noise map at time, To generate the output noise map of the deep conditional diffusion network in the SAR image generation stage, is the current time step t k Generate SAR images under

[0046] S4.6: Determine the current time step t k Is it 1? If t k =1, output and category label y k and azimuth a k Generate the corresponding SAR image, otherwise update the time step t k =t k -1, repeat S4.4 to S4.6.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. The deep conditional diffusion network constructed by the present invention adds a category label encoding module and an azimuth encoding module. During the training process, the category label and azimuth of the SAR image can be used to guide the network generation. The category label encoding module and the azimuth encoding module make the SAR image azimuth finally generated by the deep conditional diffusion network more accurate and controllable, and the distinction between categories is greater, thereby improving the posture completeness and diversity of the generated SAR image.

[0049] 2. The training process of the deep conditional diffusion network constructed by the present invention adopts the forward diffusion process and the reverse generation process, rather than the adversarial training method in the traditional generative adversarial network, which overcomes the problem of unstable training process caused by the adversarial learning of the generator and the discriminator in the prior art. The training method of the present invention is easier to train, more stable, and easier to converge. The generated SAR images are closer to the real SAR images, and the quality of the generated images is improved.

[0050] 3. Since the deep conditional diffusion network constructed by the present invention can learn the information attenuation caused by noise, and then use the learned pattern to generate images, its forward noise addition process and reverse generation process both use probabilistic modeling, so the deep conditional diffusion network constructed by the present invention has better interpretability than the prior art.

[0051] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flow chart of a method for expanding sample attitudes of a SAR recognition database based on a deep conditional diffusion network provided by an embodiment of the present invention;

[0053] Figure 2It is a module schematic diagram of a deep conditional diffusion network provided by an embodiment of the present invention;

[0054] Figure 3 It is a partial structural diagram of a deep conditional diffusion network provided by an embodiment of the present invention;

[0055] Figure 4 It is another structural schematic diagram of a deep conditional diffusion network provided by an embodiment of the present invention;

[0056] Figure 5 SAR images at different azimuth angles generated by the method of an embodiment of the present invention;

[0057] Figure 6 is an experimental result diagram generated by the method of an embodiment of the present invention;

[0058] Figure 7 It is an experimental result diagram generated by using the existing technology method. DETAILED DESCRIPTION

[0059] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, a method for expanding the sample attitude of a SAR recognition database based on a deep conditional diffusion network proposed by the present invention is described in detail below in combination with the accompanying drawings and specific implementation methods.

[0060] The above and other technical contents, features and effects of the present invention are clearly presented in the following detailed description of the specific implementation modes in conjunction with the accompanying drawings. Through the description of the specific implementation modes, the technical means and effects adopted by the present invention to achieve the predetermined purpose can be more deeply and specifically understood. However, the attached drawings are only for reference and explanation purposes and are not used to limit the technical solutions of the present invention.

[0061] 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 "comprise", "include" or any other variants are intended to cover non-exclusive inclusions, so that an article or device including a series of elements includes not only those elements, but also other elements that are not explicitly listed. In the absence of more restrictions, the elements defined by the statement "comprise one..." do not exclude the existence of other identical elements in the article or device including the elements.

[0062] See also Figure 1 , Figure 1This is a flow chart of a method for expanding the sample attitude of a SAR recognition database based on a deep conditional diffusion network provided by an embodiment of the present invention. The method comprises:

[0063] S1: construct a training data set using multiple SAR images, where the training data set includes multiple target categories, each category includes multiple image samples, and each image sample contains a corresponding target category label and target azimuth information.

[0064] Acquire N real SAR images to construct a training data set, wherein the training data set includes C target categories, each category includes S j (1, 2, ..., C) samples, each with an azimuth of a p (p=1,2,…,N);

[0065] The training data set is preprocessed, the size (dimension) of each image in the training data set is cropped to 128×128, and then the pixel value range of each cropped SAR image is converted to between 0 and 255 to obtain the preprocessed training data set.

[0066] S2: Construct a deep conditional diffusion network.

[0067] See also Figure 2 , Figure 2 1 is a module diagram of a deep conditional diffusion network provided by an embodiment of the present invention. The deep conditional diffusion network constructed in this embodiment includes a time step encoding module 101, a category label encoding module 102, an azimuth encoding module 103, a coding fusion module 104, a feature map construction module 105, four downsampling modules 106, two intermediate transition modules 107, four upsampling modules 108, and an image restoration module 109, wherein the time step encoding module 101, the category label encoding module 102 and the azimuth encoding module 103 are connected in parallel and are respectively connected to the input end of the coding fusion module 104, a feature map construction module 105, four downsampling modules 106, two intermediate transition modules 107, four upsampling modules 108 and an image restoration module 109 are connected in series in sequence, and the output end of the coding fusion module 104 is respectively connected to each downsampling module in the four downsampling modules 106, each intermediate transition module in the two intermediate transition modules 107, and each upsampling module in the four upsampling modules 108.

[0068] The time step encoding module 101 is used to map the time step of the input image to the time step encoding of the specified dimension, so as to control the inference time step; the category label encoding module 102 is used to map the target category label corresponding to the input image to the category label encoding of the specified dimension, so as to control the target category of the final generated result; the azimuth encoding module 103 is used to map the azimuth corresponding to the input image to the azimuth encoding of the specified dimension, so as to control the azimuth of the final generated result. The encoding fusion module 104 is used to sum the time step encoding, the category label encoding and the azimuth encoding to obtain the total encoding.

[0069] The feature map construction module 105 is used to extract a set of feature maps of equal size from the input image; the downsampling module 106 is used to encode the feature map, and the scale of the output feature map is reduced by half after each downsampling module, and a deep feature map is obtained after four operations; the intermediate transition module 107 is used to map the deep feature map to obtain a deep transition feature map in preparation for upsampling and restoration; the upsampling module 108 is used to decode the deep transition feature map, and the scale of the output feature map is doubled after each upsampling module, and a decoded feature map is obtained after four operations; the image restoration module 109 is used to perform channel transformation on the decoded feature map to restore it to the original input image size.

[0070] For further information, see Figure 3 , Figure 3 It is a partial structural diagram of a deep conditional diffusion network provided by an embodiment of the present invention. The time step encoding module 101 includes a first fully connected layer, a first Swish activation layer, and a second fully connected layer connected in sequence. In this embodiment, the number of input units of the first fully connected layer is 128, and the number of output units is 512; the number of input units of the second fully connected layer is 512, and the number of output units is 512; the category label encoding module 102 includes a third fully connected layer, a second Swish activation layer, and a fourth fully connected layer connected in sequence. In this embodiment, the number of input units of the third fully connected layer is 128, and the number of output units is 512, and the number of input units of the fourth fully connected layer is 512, and the number of output units is 512; the azimuth encoding module 103 includes a fifth fully connected layer, a third Swish activation layer, and a sixth fully connected layer connected in sequence. In this embodiment, the number of input units of the fifth fully connected layer is 128, and the number of output units is 512, and the number of input units of the sixth fully connected layer is 512, and the number of output units is 512.

[0071] In practice, the time step of the original input noisy SAR image passes through the first fully connected layer, the first swish activation layer and the second fully connected layer in the time step encoding module 101 in sequence to obtain the time step encoding, the category label of the noisy SAR image passes through the third fully connected layer, the second swish activation layer and the fourth fully connected layer in the category label encoding module 102 in sequence to obtain the category label encoding, and the azimuth of the noisy SAR image passes through the fifth fully connected layer, the third swish activation layer and the sixth fully connected layer in the azimuth encoding module 103 in sequence to obtain the azimuth encoding, and then the encoding fusion module 104 sums the time step encoding, the category label encoding and the azimuth encoding to obtain the total encoding, which will act on the subsequent downsampling module, the intermediate transition module and the upsampling module of the deep conditional diffusion network.

[0072] For further information, see Figure 4 , Figure 4 It is another structural diagram of a deep conditional diffusion network provided by an embodiment of the present invention. The feature map construction module 105 of this embodiment includes an input convolution layer, the input convolution layer has 1 input channel, 128 output channels, a convolution kernel size of 3×3, and a step size of 1.

[0073] The downsampling module 106 includes a first GroupNorm layer, a fourth Swish activation layer, a first convolutional layer, a fifth Swish activation layer, a seventh fully connected layer, a second GroupNorm layer, a sixth Swish activation layer, a second convolutional layer, a third convolutional layer, a first fusion unit and a second fusion unit, wherein the first GroupNorm layer, the fourth Swish activation layer and the first convolutional layer are connected in sequence, the fifth Swish activation layer and the seventh fully connected layer are connected in sequence, the input of the first GroupNorm layer is connected to the output of the feature map construction module 105, the input of the fifth Swish activation layer is connected to the output of the encoding fusion module 104, the output of the seventh fully connected layer and the output of the first convolutional layer are both connected to the input of the first fusion unit for addition; the second GroupNorm layer, the sixth Swish activation layer and the second convolutional layer are connected in sequence, the output of the first fusion unit is connected to the input of the second GroupNorm layer, the output of the second convolutional layer and the output of the feature map construction module 105 are added in the second fusion unit and the result is input to the third convolutional layer, and the output of the third convolutional layer is used as the output of the downsampling module.

[0074] Preferably, in the four downsampling modules 106, the input channels of the first convolution layer are 128, 256, 256, and 256 respectively, the output channels are all 256, the convolution kernel size is 3×3, and the step size is 1; the number of input units of the seventh fully connected layer of the four downsampling modules is 128, and the number of output units is 256; the input channels of the second convolution layers of the four downsampling modules are 128, 256, 256, and 256 respectively, the output channels are all 256, the convolution kernel size is 3×3, and the step size is 1; the input channels of the third convolution layers of the four downsampling modules are all 256, the output channels are all 256, the convolution kernel size is 3×3, and the step size is 2; the number of groups in the first GroupNorm layer and the second GroupNorm layer is 32.

[0075] Further, each intermediate transition module 107 includes a third GroupNorm layer, a seventh Swish activation layer, a fourth convolutional layer, an eighth Swish activation layer, an eighth fully connected layer, a fourth GroupNorm layer, a ninth Swish activation layer, a fifth convolutional layer, a third fusion unit and a fourth fusion unit, wherein the third GroupNorm layer, the seventh Swish activation layer and the fourth convolutional layer are connected in sequence, the eighth Swish activation layer and the eighth fully connected layer are connected, the input of the third GroupNorm layer is connected to the output of the last downsampling module 106, the input of the eighth Swish activation layer is connected to the output of the encoding fusion module 104, the output of the eighth fully connected layer and the output of the fourth convolutional layer are connected to the input of the third fusion unit for addition, and the fourth GroupNorm layer, the ninth Swish activation layer and the fifth convolutional layer are connected in sequence, the output of the third fusion unit is connected to the input of the fourth GroupNorm layer, and the output of the fifth convolutional layer and the output of the last downsampling module 106 are added in the fourth fusion unit and serve as the output of the intermediate transition module 107.

[0076] Preferably, the input channels of the fourth convolutional layer of each intermediate transition module 107 are 256, the output channels are 256, the convolution kernel size is 3×3, and the step size is 1; the number of input units of the eighth fully connected layer is 128, and the number of output units is 256; the input channels of the fifth convolutional layer are 256, the output channels are 256, the convolution kernel size is 3×3, and the step size is 1; the number of groups of the third GroupNorm layer and the fourth GroupNorm layer is 32.

[0077] Each upsampling module 108 includes a fifth GroupNorm layer, a tenth Swish activation layer, a sixth convolution layer, an eleventh Swish activation layer, a ninth fully connected layer, a sixth GroupNorm layer, a twelfth Swish activation layer, a seventh convolution layer, a nearest neighbor interpolation layer, an eighth convolution layer, a fifth fusion unit, and a sixth fusion unit, wherein:

[0078] The fifth GroupNorm layer, the tenth Swish activation layer and the sixth convolutional layer are connected in sequence, the eleventh Swish activation layer and the ninth fully connected layer are connected, the input of the fifth GroupNorm layer is connected to the output of the last intermediate transition module 107, the input of the eleventh Swish activation layer is connected to the output of the encoding fusion module 104, the output of the ninth fully connected layer and the output of the sixth convolutional layer are connected to the input of the fifth fusion unit and added, and the sixth GroupNorm layer, the twelfth Swish activation layer and the seventh convolutional layer are connected in sequence, the output of the fifth fusion unit is connected to the input of the sixth GroupNorm layer, the output of the seventh convolutional layer and the output of the last intermediate transition module 107 are added in the sixth fusion unit and the summed results are input to the nearest neighbor interpolation layer and the eighth convolutional layer in sequence, and the output of the eighth convolutional layer is used as the output of the upsampling module 108.

[0079] Preferably, the input channel of the sixth convolution layer of each upsampling module 108 is 256, and the output channels are 256, 256, 256, and 128 respectively, the convolution kernel size is 3×3, and the step size is 1; the number of input units of the ninth fully connected layer is 128, and the number of output units is 256, 256, 256, and 128 respectively, the input channel of the seventh convolution layer is 256, and the output channels are 256, 256, 256, and 128 respectively, the convolution kernel size is 3×3, and the step size is 1; the interpolation factor of the nearest neighbor interpolation layer is 2; the input channels of the eighth convolution layer are 256, 256, 256, and 128 respectively, and the output channels are 256, 256, 256, and 128 respectively, the convolution kernel size is 3×3, and the step size is 2; the number of groups in the fifth GroupNorm layer and the sixth GroupNorm layer is 32.

[0080] Furthermore, the image restoration module 109 of this embodiment includes a seventh GroupNorm layer, a thirteenth Swish activation layer and a ninth convolution layer connected in sequence, wherein the ninth convolution layer has an input channel of 128, an output channel of 1, a convolution kernel size of 3×3, and a step size of 1; and the number of groups in the seventh GroupNorm layer is 32.

[0081] S3: Using the training data set to train the deep conditional diffusion network to obtain a trained deep conditional diffusion network.

[0082] S3 of this embodiment specifically includes:

[0083] S3.1: Set the maximum inference time step of the deep conditional diffusion network to T;

[0084] S3.2: Forward diffusion process: SAR images in the training sample set are used to form a small batch sample set, and the SAR images in the small batch sample set are denoised to obtain noisy SAR image samples corresponding to each SAR image.

[0085] Specifically, step S3.2 of this embodiment includes:

[0086] S3.21: Select m SAR images from the training data set to form a small batch sample set, 1≤m≤64, where the i-th SAR image is Indicates that i = 1, 2, ..., m;

[0087] S3.22: Use a random integer generator to generate a corresponding time step t for each SAR image in the small batch sample set i , where 1≤t i ≤T,t i is an integer;

[0088] S3.23: Use a Gaussian random number generator to generate a corresponding noise map N for each SAR image in the small batch sample set i , where the data dimension of the noise map is consistent with that of the SAR image, the number of noise map channels is 1, and the width and height of the noise map are both 128;

[0089] S3.24: According to the noise addition formula, use the time step t corresponding to the SAR image i and noise map N i , for each SAR image Add noise and calculate the noisy SAR image samples corresponding to each SAR image The noise addition formula is as follows:

[0090]

[0091] in, α t =1-β t , is the i-th original SAR image, is the i-th SAR image t i The noise result at the moment, ε is the standard Gaussian white noise, β t It represents the noise variance at time t. Its actual value is the value of the tth number among T numbers sampled at equal intervals in the interval [0.0001, 0.02].

[0092] S3.3: The noisy SAR image samples corresponding to all SAR images in the small batch sample set Time step t i , category label yi 、Azimuth a i , noise map N i At the same time, it is input into the constructed deep conditional diffusion network, where t i ,y i , a i The feature map construction module, time step encoding module, category label encoding module and azimuth encoding module are input into the network respectively. The deep conditional diffusion network estimates the noise of each input noisy SAR image sample to obtain the noise map predicted by the network.

[0093] Specifically, the training samples are first processed by the feature map construction module to obtain a set of feature maps. i After the time step encoding module, the time step encoding and category label y are obtained. i The category label encoding module obtains the category label encoding, and the azimuth angle a i The azimuth coding module obtains the azimuth coding, and then the time step coding, category label coding and azimuth coding are summed as the total coding. The feature map and the total coding are sequentially input into the downsampling module, the intermediate transition module and the upsampling module. Each module is a convolutional layer, a fully connected layer and a convolutional layer structure. The first convolutional layer acts on the feature map to obtain a deep feature map. The fully connected layer acts on the total coding, and the total coding is mapped to a coding vector with the same number of channels as the deep feature map. The vector is copied into a tensor of the same size as the deep feature map and summed with it to obtain a coding-guided deep feature map. The subsequent convolutional layers act on the coding-guided deep feature map to obtain the output feature map, which is used as the input feature map of the next module. The feature map obtained by the above modules is finally sent to the image restoration module to obtain the output of the deep conditional diffusion network, and the output dimension size is consistent with the training sample dimension size;

[0094] According to the loss function, the network is used to predict the noise map vs. Input Noise Graph N i Calculate the network loss value, use the gradient descent method to calculate the gradient of each parameter to be learned in the deep conditional diffusion network, and use the AdamW optimizer to update all parameters to be learned. The expression of the loss function is as follows:

[0095]

[0096] Among them, L mse represents the total loss of all input samples in the mini-batch sample set, is the predicted noise map output of the deep conditional diffusion network, is the i-th noisy SAR image sample in the small batch sample set, t i is the time step corresponding to the i-th noisy SAR image sample, yi is the category label corresponding to the i-th noisy SAR image sample, a i is the azimuth angle corresponding to the i-th noisy SAR image sample, N i is the noise map corresponding to the i-th noisy SAR image sample;

[0097] It is determined whether the loss functions of the deep conditional diffusion network have converged. If so, a trained deep conditional diffusion network is obtained, and the parameters of the trained deep conditional diffusion network are saved. Otherwise, the training is continued.

[0098] S4: Generate SAR images using the trained deep conditional diffusion network.

[0099] Step S4 of this embodiment includes:

[0100] S4.1: Set the number of SAR images to be generated to K;

[0101] S4.2: Generate H noise images using a Gaussian random number generator. The data dimension of the noise image is consistent with that of the SAR image. The value of H is equal to the number K of SAR images to be generated.

[0102] S4.3: Set the time step t for each SAR image to be generated k , where t k =T, k = 1, 2, ..., K, t k is an integer;

[0103] S4.4: The noise maps corresponding to all SAR images to be generated Time step t k , the expected generated category label y k , the desired azimuth angle a k , sent to the trained deep conditional diffusion network, where, m,y i , a i The feature map construction module, time step encoding module, category label encoding module, and azimuth encoding module are input into the network respectively to obtain the output prediction noise map;

[0104] S4.5: Update the noise map according to the image generation formula and generate the current time step t k The SAR image below, wherein the expression of the image generation formula is as follows:

[0105]

[0106]

[0107] in, Indicates t k The noise variance at time, is the t corresponding to the kth SAR image to be generated k Noise map at time -1, is the t corresponding to the kth SAR image to be generated k The noise map at time, To generate the predicted noise map output of the deep conditional diffusion network at the SAR image stage, is the current time step t k Generate SAR images under .

[0108] S4.6: Determine the current time step t k Is it 1? If t k =1, output and category label y k and azimuth a k Generate the corresponding SAR image, otherwise update the time step t k =t k -1, repeat S4.4 to S4.6.

[0109] The effects of the present invention are further described below in conjunction with simulation experiments.

[0110] 1. Simulation experiment conditions:

[0111] The hardware platform of the simulation experiment of the present invention is: the processor is an Intel(R) Xeon(R) Gold 6242R CPU with a main frequency of 3.10GHz, a memory capacity of 256GB, and a graphics card is an Nvidia RTX 3090 with a video memory capacity of 24GB.

[0112] The software platform of the simulation experiment of the present invention is: Ubuntu 21.04 operating system, Python 3.8, Pytorch1.9.

[0113] The experimental data used in the simulation experiment of the present invention is based on the public MSTAR data set of moving and stationary targets published by the MSTAR program supported by the Defense Advanced Research Projects Agency of the United States, with a resolution of 0.3m×0.3m, including SAR images with pitch angles of 15 degrees and 17 degrees, and the pixel size of the image is 128×128.

[0114] The deep conditional diffusion network experiment constructed by the present invention is set up as follows: the batch size during training is 16, the initial learning rate is 2e-4, the optimizer is AdamW, the gradient clipping is set to 1.0, the number of warm-up training iterations is set to 2000, and the total number of iterations is set to 100000.

[0115] 2. Experimental content and results analysis

[0116] (1) Experiment 1

[0117] In the simulation experiment of the present invention, the samples of odd azimuth angles in the training set with a pitch angle of 17 degrees in the MSTAR data set are used as the training data set, and the training data set includes a total of 1166 SAR image samples. The method of the present invention is used to generate 2S1, BMP2, BDRM2, BTR60, BTR70, D7, T62, T72, ZIL131, and ZSU23 / 4 ten categories of omnidirectional SAR images. The example diagram of the generated samples provided by the embodiment of the present invention is as follows: Figure 5 As shown, Figure 5 Each row represents a category, which are 2S1, BMP2, BDRM2, BTR60, BTR70, D7, T62, T72, ZIL131, ZSU23 / 4, and each column represents an azimuth angle, with an azimuth angle of 40° from 0 to 360°.

[0118] from Figure 5 It can be seen that there are differences between the columns, and the orientation of the target reflects its corresponding azimuth angle. Therefore, the generation result of the present invention generates high-quality, full-category, and full-angle SAR images when there is a lack of azimuth angle in the training data. The azimuth angle of the SAR image finally generated is more complete, and the distinction between categories is greater, which improves the diversity of the generated SAR images and realizes the expansion of the number of samples and postures of the SAR recognition database.

[0119] (2) Experiment 2:

[0120] In the simulation experiment of the present invention, a fixed number of SAR images are generated by using the method of the present invention and an existing technology, and the quality of the generated SAR images is evaluated. In this experiment, SAR images with a pitch angle of 17 degrees in the MSTAR dataset are used as a training dataset, and the training dataset includes a total of 2746 SAR image samples.

[0121] The existing technology used is the paper "Unsupervised Representation Learning With Deep Convolutional Generative Adversarial Networks" (DCGAN for short) published by Alec et al., which applies convolutional neural networks to the GAN generative adversarial network model. The discriminator of DCGAN retains the overall architecture of CNN, and the generator replaces the convolution layer with a deconvolution layer. In addition, BN (BatchNormalization) layers are used in the discriminator and generator, ReLU is used as the activation function in the generator network, Tanh is used as the activation function in the last layer, and LeakyReLU is used as the activation function in the discriminator network.

[0122] The FID score (Frechet Inception Distance score) is introduced to judge the quality of the generated image. The FID score measures the similarity of two sets of images from the statistical aspects of the computer vision features of the original image. It is a measure of the distance between the feature vectors of the real image and the generated image. It is calculated by extracting features using the Inception v3 image classification model. The lower the score, the more similar the statistics of the two sets of images, that is, the better the quality of the generated image; conversely, the higher the score, the worse the quality of the generated image.

[0123] The training data set is used to train the generative models of the present invention and the prior art respectively, and the trained generative models are used to generate 2746 SAR image samples respectively. The generated samples constitute the generated sample set, and the FID score between the generated sample set and the training set is calculated. The results are shown in Table 1. The FID score of the conventional method is 43.64, and the FID score of the present method is 24.07. The results show that the SAR image generated by the present invention can obtain a lower FID score, that is, a SAR image with better generated quality can be obtained.

[0124] Table 1 Comparison of image generation quality between the method of the present invention and the prior art

[0125] FID score Prior Art 43.64 The present invention 24.07

[0126] The example diagram of the generated samples provided by the present invention in this experiment is as follows Figure 6 As shown, the example diagram of the generated sample provided by the prior art is as follows Figure 7 As shown in the figure, by comparing the schematic diagram of some results of generating SAR images by the present invention and the prior art, the result generated by the method of the embodiment of the present invention is better than that of the conventional method. Specifically, the present invention can better generate the strong scattering area of ​​the target and the shadow area of ​​the target, and at the same time, the background area can also better correspond to the background area in the real SAR image; while the SAR image generated by the prior art has smear marks in the strong scattering area of ​​the target, and the generated result in the background area is greatly different from the real SAR image.

[0127] (3) Experiment 3

[0128] In the simulation experiment of the present invention, a fixed number of SAR images are generated by using the method of the present invention and an existing technology, and the quality of the generated SAR images is evaluated based on the classification performance. The classification model adopts the CNN model. In this experiment, the samples of odd azimuth angles in the training set with a pitch angle of 17 degrees in the MSTAR data set are used as the generated training data set, and the training data set includes a total of 1166 SAR image samples.

[0129] The prior art is consistent with the prior art in Experiment 2. The generative models of the present invention and the prior art are trained using the training data set, and the trained generative models are used to generate ten types of targets, with the target azimuth angle range of 0 to 360° and the azimuth angle interval of 1°, generating a total of 3610 samples. The generated samples constitute the training set of the classification model, and the SAR image samples with a pitch angle of 15 degrees in the MSTAR data set are used as the test set, and the test set includes a total of 2426 SAR image samples.

[0130] The CNN model was trained using a classification training set constructed using the SAR images generated by the method of the present invention and a classification training set constructed using the SAR images generated by the prior art, and tested using the above test set. The classification experimental results are shown in the following table.

[0131] Table 2 Performance comparison results of the method of the present invention and the prior art

[0132] Classification accuracy Prior Art 82.70％ The present invention 90.60％

[0133] In this experiment, the classification performance of the SAR images generated by the present invention is improved by 8 percentage points compared with the classification performance of the SAR images generated by the prior art. From the perspective of classification, it is further verified that the SAR images generated by the present invention are of higher quality, have greater distinction between categories, are closer to real SAR images, and improve the authenticity of the generated SAR images.

[0134] The deep conditional diffusion network constructed by the present invention adds a category label encoding module and an azimuth encoding module. During the training process, the category label and azimuth of the SAR image can be used to guide the network generation. The category label encoding module and the azimuth encoding module make the SAR image azimuth finally generated by the deep conditional diffusion network more accurate and controllable, and the distinction between categories is greater, thereby improving the posture completeness and diversity of the generated SAR image. The training process of the deep conditional diffusion network constructed by the present invention adopts a forward diffusion process and a reverse generation process, rather than the adversarial training method in the traditional generative adversarial network, which overcomes the problem of instability in the training process caused by the adversarial learning of the generator and the discriminator in the prior art. The training method of the present invention is easier to train, more stable, and easier to converge. The generated SAR image is closer to the real SAR image, and the quality of the generated image is improved.

[0135] In addition, since the deep conditional diffusion network constructed by the present invention can learn the information attenuation caused by noise, and then use the learned pattern to generate images, its forward noisy process and reverse generation process both use probabilistic modeling, so the deep conditional diffusion network constructed by the present invention has better interpretability than the prior art.

[0136] Another embodiment of the present invention provides a storage medium, in which a computer program is stored, and the computer program is used to execute the steps of the SAR recognition database sample posture expansion method based on the deep conditional diffusion network described in the above embodiment. Another aspect of the present invention provides an electronic device, including a memory and a processor, in which a computer program is stored, and when the processor calls the computer program in the memory, the steps of the SAR recognition database sample posture expansion method based on the deep conditional diffusion network described in the above embodiment are implemented. Specifically, the above-mentioned integrated module implemented in the form of a software function module can be stored in a computer-readable storage medium. The above-mentioned software function module is stored in a storage medium, including several instructions for an electronic device (which can be a personal computer, a server, or a network device, etc.) or a processor (processor) to execute some steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), disk or optical disk and other media that can store program codes.

[0137] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A method for expanding the sample attitude of SAR recognition database based on deep conditional diffusion network, characterized in that: include: S1: construct a training data set using multiple real SAR images, where the training data set includes multiple target categories, each category includes multiple image samples, and each image sample contains a corresponding target category label and target azimuth information; S2: construct a deep conditional diffusion network, the deep conditional diffusion network includes a time step encoding module, a category label encoding module, an azimuth encoding module, a code fusion module, a feature map construction module, multiple downsampling modules, multiple intermediate transition modules, multiple upsampling modules and an image restoration module, wherein the time step encoding module is used to map the time step of the input image to a time step code of a specified dimension; the category label encoding module is used to map the target category label corresponding to the input image to a category label code of a specified dimension; the azimuth encoding module is used to map the azimuth corresponding to the input image to an azimuth code of a specified dimension; the code fusion module is used to sum the time step code, the category label code and the azimuth code to obtain a total code; The feature map construction module is used to extract a set of feature maps of equal size from the input image; the multiple downsampling modules are used to encode the feature maps, and each time they pass through a downsampling module, the scale of the output feature map is reduced by half, and a deep feature map is obtained after multiple operations; the multiple intermediate transition modules are used to map the deep feature map to obtain a deep transition feature map; the multiple upsampling modules are used to decode the deep transition feature map, and each time they pass through an upsampling module, the scale of the output feature map is doubled, and a decoded feature map is obtained after multiple operations; the image restoration module is used to perform channel transformation on the decoded feature map to restore it to the original input image size; S3: training the deep conditional diffusion network using the training data set to obtain a trained deep conditional diffusion network; S4: Generate SAR images using the trained deep conditional diffusion network.

2. The SAR recognition database sample attitude expansion method based on deep conditional diffusion network according to claim 1 is characterized in that: The time step encoding module includes a first fully connected layer, a first Swish activation layer, and a second fully connected layer connected in sequence; the category label encoding module includes a third fully connected layer, a second Swish activation layer, and a fourth fully connected layer connected in sequence; the azimuth encoding module includes a fifth fully connected layer, a third Swish activation layer, and a sixth fully connected layer connected in sequence; the feature map construction module includes an input convolutional layer.

3. The SAR recognition database sample attitude expansion method based on deep conditional diffusion network according to claim 1 is characterized in that: Each downsampling module includes a first GroupNorm layer, a fourth Swish activation layer, a first convolutional layer, a fifth Swish activation layer, a seventh fully connected layer, a second GroupNorm layer, a sixth Swish activation layer, a second convolutional layer, a third convolutional layer, a first fusion unit, and a second fusion unit, wherein: The first GroupNorm layer, the fourth Swish activation layer and the first convolutional layer are connected in sequence, the fifth Swish activation layer and the seventh fully connected layer are connected in sequence, the input of the first GroupNorm layer is connected to the output of the feature map construction module, the input of the fifth Swish activation layer is connected to the output of the encoding fusion module, the output of the seventh fully connected layer and the output of the first convolutional layer are both connected to the input of the first fusion unit for addition; the second GroupNorm layer, the sixth Swish activation layer and the second convolutional layer are connected in sequence, the output of the first fusion unit is connected to the input of the second GroupNorm layer, the output of the second convolutional layer and the output of the feature map construction module are added in the second fusion unit and the result is input to the third convolutional layer, and the output of the third convolutional layer is used as the output of the downsampling module.

4. The SAR recognition database sample attitude expansion method based on deep conditional diffusion network according to claim 1 is characterized in that: Each intermediate transition module includes a third GroupNorm layer, a seventh Swish activation layer, a fourth convolutional layer, an eighth Swish activation layer, an eighth fully connected layer, a fourth GroupNorm layer, a ninth Swish activation layer, a fifth convolutional layer, a third fusion unit, and a fourth fusion unit, wherein: The third GroupNorm layer, the seventh Swish activation layer and the fourth convolutional layer are connected in sequence, the eighth Swish activation layer and the eighth fully connected layer are connected, the input of the third GroupNorm layer is connected to the output of the last downsampling module, the input of the eighth Swish activation layer is connected to the output of the encoding fusion module, the output of the eighth fully connected layer and the output of the fourth convolutional layer are connected to the input of the third fusion unit for addition; the fourth GroupNorm layer, the ninth Swish activation layer and the fifth convolutional layer are connected in sequence, the output of the third fusion unit is connected to the input of the fourth GroupNorm layer, the output of the fifth convolutional layer and the output of the last downsampling module are added in the fourth fusion unit and used as the output of the intermediate transition module.

5. The method for expanding the sample attitude of the SAR recognition database based on the deep conditional diffusion network according to claim 1 is characterized in that: Each upsampling module includes a fifth GroupNorm layer, a tenth Swish activation layer, a sixth convolution layer, an eleventh Swish activation layer, a ninth fully connected layer, a sixth GroupNorm layer, a twelfth Swish activation layer, a seventh convolution layer, a nearest neighbor interpolation layer, an eighth convolution layer, a fifth fusion unit, and a sixth fusion unit, wherein, The fifth GroupNorm layer, the tenth Swish activation layer and the sixth convolutional layer are connected in sequence, the eleventh Swish activation layer and the ninth fully connected layer are connected, the input of the fifth GroupNorm layer is connected to the output of the last intermediate transition module, the input of the eleventh Swish activation layer is connected to the output of the encoding fusion module, the output of the ninth fully connected layer and the output of the sixth convolutional layer are connected to the input of the fifth fusion unit and added, and the sixth GroupNorm layer, the twelfth Swish activation layer and the seventh convolutional layer are connected in sequence, the output of the fifth fusion unit is connected to the input of the sixth GroupNorm layer, the output of the seventh convolutional layer and the output of the last intermediate transition module are added in the sixth fusion unit and the summed results are input to the nearest neighbor interpolation layer and the eighth convolutional layer in sequence, and the output of the eighth convolutional layer is used as the output of the upsampling module.

6. The method for expanding the sample attitude of the SAR recognition database based on the deep conditional diffusion network according to claim 1 is characterized in that: The image restoration module includes a seventh GroupNorm layer, a thirteenth Swish activation layer and a ninth convolutional layer which are connected in sequence.

7. The method for expanding the sample attitude of the SAR recognition database based on the deep conditional diffusion network according to claim 1 is characterized in that: The S3 includes: S3.1: Set the maximum inference time step of the deep conditional diffusion network to T; S3.2: using the SAR images in the training sample set to form a small batch sample set, and adding noise to the SAR images in the small batch sample set to obtain a noisy SAR image sample corresponding to each SAR image; S3.3: Noise-added SAR image samples corresponding to all SAR images in the small batch sample set Time step t i , category label y i 、Azimuth a i , noise map N i At the same time, it is input into the constructed deep conditional diffusion network, where t i ,y i , a i The feature map construction module, time step encoding module, category label encoding module and azimuth encoding module are input into the network respectively. The deep conditional diffusion network estimates the noise of each input noisy SAR image sample to obtain the network predicted noise map S3.4: Using the network to predict the noise map based on the loss function With the input noise map N i Calculate the network loss value, use the gradient descent method to calculate the gradient of each parameter to be learned in the deep conditional diffusion network and update all parameters to be learned; S3.5: Determine whether the loss functions of the deep conditional diffusion network have converged. If so, obtain a trained deep conditional diffusion network and save the parameters of the trained deep conditional diffusion network. Otherwise, continue training.

8. The method for expanding the sample attitude of the SAR recognition database based on the deep conditional diffusion network according to claim 7 is characterized in that: The S3.2 includes: S3.21: m SAR images are selected from the training data set to form a small batch sample set, where the i-th SAR image is Indicates that i = 1, 2, ..., m; S3.22: Use a random integer generator to generate a corresponding time step t for each SAR image in the small batch sample set i , where 1≤t i ≤T,t i is an integer; S3.23: Use a Gaussian random number generator to generate a corresponding noise map N for each SAR image in the small batch sample set i ; S3.24: According to the noise addition formula, use the time step t corresponding to the SAR image i and noise map N i , for each SAR image Add noise to obtain the noisy SAR image samples corresponding to each SAR image Wherein, the noise addition formula is: in, α t =1-β t , is the i-th original SAR image, is the i-th SAR image t i The noise result at the moment, ε is the standard Gaussian white noise, β t represents the noise variance at time t.

9. The method for expanding the sample attitude of the SAR recognition database based on the deep conditional diffusion network according to claim 7 is characterized in that: The expression of the loss function in S3.4 is: Among them, L mse represents the total loss of all input samples in the mini-batch sample set, is the predicted noise map output of the deep conditional diffusion network, N i is the noise map corresponding to the i-th noisy SAR image sample.

10. The method for expanding the sample attitude of a SAR recognition database based on a deep conditional diffusion network according to any one of claims 7 to 9, characterized in that: The S4 includes: S4.1: Set the number of SAR images to be generated to K; S4.2: Generate H noise images using a Gaussian random number generator. The data dimension of the noise image is consistent with that of the SAR image. The value of H is equal to the number K of SAR images to be generated. S4.3: Set the time step t for each SAR image to be generated k , where t k =T, k = 1, 2,, K, t k is an integer; S4.4: The noise maps corresponding to all SAR images to be generated Time step t k , the expected generated category label y k , the desired azimuth angle a k , sent to the trained deep conditional diffusion network, where, m,y i , a i The feature map construction module, time step encoding module, category label encoding module, and azimuth encoding module are input into the network respectively to obtain the output prediction noise map; S4.5: Update the noise map according to the image generation formula and generate the current time step t k The SAR image under the condition of FIG. 1 is shown in FIG. 1 , wherein the expression of the image generation formula is: in, Indicates t k The noise variance at time, is the t corresponding to the kth SAR image to be generated k Noise map at time -1, is the t corresponding to the kth SAR image to be generated k The noise map at time, To generate the output noise map of the deep conditional diffusion network in the SAR image generation stage, is the current time step t k Generate SAR images under S4.6: Determine the current time step t k Is it 1? If t k =1, output and category label y k and azimuth a k Generate the corresponding SAR image, otherwise update the time step t k =t k -1, repeat S4.4 to S4.6.

Images