SAR ship image generation method and generation device based on improved GAN-GP algorithm

Abstract

The present invention discloses a SAR ship image generation method and a generation device based on an improved GAN-GP algorithm, the generation method comprising: acquiring a target image; obtaining a generated image through a GAN network; training the GAN network to adjust the generated image based on the target image, the generated image, the fused image after the target image and the generated image are fused, and a label indicating the category to which the target image belongs. The present invention realizes the simultaneous acquisition of generated ship targets and their category information in a generative adversarial network, the introduction of a residual network structure in the generator of the GAN network can not only improve the training performance of the deep network, but also better retain the structural information of the ship targets in the SAR image; the introduction of a gradient penalty in the discriminator makes the GAN network training more stable, converges faster, and can generate higher quality sample images.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a SAR ship image generation method and a generation device based on an improved GAN-GP algorithm. Background Art

[0002] Synthetic aperture radar (SAR) can actively emit microwaves, is less affected by weather conditions, and has the ability to image in all weather conditions. It can provide continuous and high-precision observation data for ocean areas with complex meteorological conditions, and has been widely used in marine target detection and classification. At present, SAR image ship classification has received widespread attention and has achieved certain results, mainly based on polarization target scattering characteristics, local statistical features, and deep learning-based classification algorithms.

[0003] At present, among these methods, deep learning algorithms based on convolutional neural networks have been widely studied and have achieved fruitful results. For example, VGG19 with a deeper network convolution structure is used to extract multi-level geometric and texture features of ship targets, and then the fully connected feature descriptions are classified using a softmax function to obtain the category information of ship targets. Training based on a large number of samples can enable convolutional neural networks to effectively identify ship targets of various postures and types in SAR images, but the production of labeled samples often requires a lot of manpower and time costs.

[0004] In recent years, since the Generative Adversarial Network (GAN) was proposed, related image generation network models have been produced one after another and widely used in the generation and classification tasks of remote sensing images. For example, when training the Conditional Generative Adversarial Networks (CGAN), the image category labels are added to both the generator and the discriminator to achieve the directional generation of the image; based on the Auxiliary Classifier Generative Adversarial Networks (ACGAN), the image label information is used for training, and the label information is added to the generator to achieve the directional generation of the image. Although the GAN network has powerful image feature learning and image generation capabilities, such networks (such as ACGAN networks) usually have gradient vanishing and diffusion phenomena due to the complex network structure, making the network model difficult to train. In addition, when the discriminator of the ACGAN network is used for image classification, there are problems such as slow training speed and poor discrimination effect. Summary of the invention

[0005] In order to solve the problems existing in the prior art, the embodiments of the present invention provide a SAR ship image generation method and a generation device based on an improved GAN-GP algorithm, which can at least solve the above-mentioned problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a SAR ship image generation method based on an improved GAN-GP algorithm, comprising:

[0007] Get the target image;

[0008] Get the generated image through the GAN network;

[0009] Based on the target image, the generated image, the fused image after the target image and the generated image are fused, and the label indicating the category to which the target image belongs, the GAN network is trained to adjust the generated image.

[0010] In some embodiments, acquiring the target image includes:

[0011] A SAR image containing a target is acquired, and a slicing process is performed on the SAR image to obtain a target image.

[0012] In some embodiments, it is characterized in that before slicing the acquired SAR image, the method further comprises:

[0013] The acquired SAR image is preprocessed, and the preprocessing includes performing radiation calibration and normalization processing on the acquired SAR image.

[0014] In some embodiments, the GAN network includes a generator and a discriminator connected to the generator,

[0015] The input of the generator includes a random vector and a class label vector, and the output of the generator includes a generated image;

[0016] The input of the discriminator includes a target image, a generated image, a fused image after the target image and the generated image are fused, and a label indicating the category to which the target image belongs. The output of the discriminator includes a true or false judgment of the sample data input into the GAN network and a classification result.

[0017] In some embodiments, the GAN network is a convolutional neural network, and the generator of the GAN network includes a residual connection structure.

[0018] In some embodiments, the residual connection structure includes multiple residual blocks, each residual block includes: a deconvolution layer, a convolution layer, a block regularization, an activation function, and a fusion layer.

[0019] In some embodiments, training the GAN network includes:

[0020] Input the random vector and class label vector into the generator to obtain the generated image;

[0021] Input the training sample image, the generated image, the fused image after the training sample image and the generated image are fused, and the class label into the discriminator for training;

[0022] Fix the discriminator parameters, input the random vector and class label vector into the generator again for training, and update the generator parameters;

[0023] Alternately train the generator and discriminator until the GAN network converges.

[0024] In some embodiments, the GAN network model obtained by training the GAN network is further optimized, and the optimization includes:

[0025] Input the random vector and the class label vector into the trained generator and output the generated image to be tested;

[0026] Input the target image to be tested, the generated image to be tested, the fused image to be tested, and the class label into the trained discriminator and compare them with the parameters of the discriminator;

[0027] Based on the test of the GAN network model, the parameters of the GAN network model are optimized.

[0028] A second aspect of an embodiment of the present invention provides an image generating device, including:

[0029] An acquisition module configured to acquire a target image;

[0030] A generation module configured to obtain a generated image through a GAN network;

[0031] A training module is configured to train the GAN network to adjust the generated image based on the target image, the generated image, the fused image after the target image and the generated image are fused, and a label indicating the category to which the target image belongs.

[0032] A third aspect of an embodiment of the present invention provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the above-mentioned SAR ship image generation method based on the improved GAN-GP algorithm are implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of a SAR ship image generation method based on an improved GAN-GP algorithm according to an embodiment of the present invention;

[0034] Figure 2 A schematic diagram of a slice of a ship target image according to an embodiment of the present invention;

[0035] Figure 3 Schematic diagram of the structure of a GAN network according to an embodiment of the present invention;

[0036] Figure 4 Schematic diagram of the structure of the generator and the discriminator in the GAN network model of an embodiment of the present invention;

[0037] Figure 5 It is a schematic diagram of a generated image of a ship target according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The present invention will be described with reference to certain drawings and with respect to specific embodiments, but the present invention is not limited thereto, but only by the limitations of the claims. Any reference numerals in the claims should not be interpreted as limiting the scope. The drawings are merely schematic and non-restrictive. In the drawings, the size of some elements may be enlarged and not drawn in proportion for illustrative purposes. When the term "comprising" is used in the specification and claims, it does not exclude other elements or steps. When an indefinite article or a definite article such as "one (a)" or "(an)", "the" is used to refer to a singular noun, unless otherwise expressly stated, this includes the plural number of the noun. In addition, the terms first, second, third, etc. are used to distinguish similar elements in the specification and claims, and are not necessarily used to describe a sequential order or a temporal order. It should be understood that the terms used in this way can be interchangeable where appropriate, and the embodiments of the present invention described herein can be operated in other orders except for the order described or illustrated herein.

[0039] Unless otherwise defined herein, the scientific and technological terms and phrases used in the present invention shall have the meanings commonly understood by those of ordinary skill in the art. Generally, the molecular and cell biology, genetics, and protein and nucleic acid chemistry and hybridization techniques described herein and the terms used in the techniques are those used in the art that are well known and commonly used. Unless otherwise indicated, the methods are generally performed according to conventional methods that are well known in the art and described in various general and more specific references, which are cited and discussed in this specification.

[0040] In order to more thoroughly understand the features and technical contents of the present invention, embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. The attached drawings and specific embodiments are only for reference and explanation, and are not intended to limit the present invention.

[0041] Figure 1 The flowchart of the SAR ship image generation method based on the improved GAN-GP algorithm of the embodiment of the present invention is as follows. The embodiment of the present invention is described by taking the generation of SAR sample images of ship targets as an example. The ship targets are mainly divided into three categories: cargo ships, container ships and oil tankers.

[0042] like Figure 1 As shown, a first aspect of an embodiment of the present invention provides a SAR ship image generation method based on an improved GAN-GP algorithm, comprising the following steps:

[0043] Step S1: Acquire the target image.

[0044] Acquiring the target image specifically includes: acquiring a SAR image containing the target, and performing slicing processing on the SAR image to obtain the target image.

[0045] In this step, synthetic aperture radar (SAR), such as GF-3, is used to obtain the fully polarized SAR image of the ship as the data source, and the image annotation tool is used to cut the ship sample to obtain the target image with the classification category label. The SAR image is the original image, and the target image is the real sample image, which is used as the input of the real sample of the GAN network.

[0046] The GAN network of the present application is mainly used to generate images that are similar to real sample images. The images generated by the GAN network can be used as sample images to enrich the corresponding sample set, for example, using the generated images to expand the training set of ship target images.

[0047] In an embodiment of the present invention, a certain number of target images to be generated are collected to establish a sample data set, the sample data set includes N sample images, and the N sample images belong to M categories respectively, wherein N and M are positive integers, and N is greater than M, and multiple sample images of multiple classification categories constitute the sample data set. In this embodiment, the ship target images include images of three classification categories: cargo ships, container ships, and tankers, and each category image contains images of multiple different angles and postures.

[0048] The sample data set includes a training set and a test set, which are used for GAN network training and testing respectively.

[0049] For example, the collected sample data set can be divided into a training set and a test set according to a certain ratio, which are used as training samples and test samples of the GAN network respectively. The training samples are used to train the GAN network model, and the test samples are used to evaluate the GAN network model.

[0050] Figure 2 FIG. 4 shows a schematic diagram of a slice of a ship target image (including a cargo ship, a container ship and an oil tanker) according to an embodiment of the present invention; Figure 2 As shown in FIG. 1 , when the acquired ship SAR image is sliced, each slice basically contains only one ship, and there are multiple sample slices of different angles and scales for the same type of ship.

[0051] In the process of making sample slices, the entire acquired SAR image is first sliced ​​using an overlapping sliding window to generate sample slices of 256×256 pixels. Then, the position and circumscribed rectangle of the ship in the sample slice are manually outlined visually in the image annotation tool.

[0052] Before the acquired SAR image is sliced, the method further includes: preprocessing the acquired SAR image, and the preprocessing includes: performing radiation calibration and normalization processing on the acquired SAR image. By normalizing the SAR image, the backscatter value of the image can be normalized to a normal range, and the image contrast error caused by factors such as imaging and sea surface waves can be reduced. After the SAR image is normalized, the image can be stretched according to the histogram statistics to reduce the image contrast error.

[0053] Step S2: Obtain a generated image through the GAN network.

[0054] Specifically, by inputting a random vector and a class label vector into the generator of the GAN network, a generated image can be obtained.

[0055] Figure 3 A schematic diagram of the structure of a GAN network according to an embodiment of the present invention is shown. Figure 3 As shown, the GAN network includes: a generator and a discriminator connected to the generator. The input of the generator includes a random vector and a class label vector (a label representing a classification category), and the generator can generate an image based on the input vector data and output it. The input of the discriminator includes a target image, a generated image, a fused image after the target image and the generated image are fused, and a label (class label) representing the category to which the target image belongs.

[0056] Among them, the target image is the target image obtained by the slicing process in step S1, which is a real sample image; the generated image is generated by the generator, and the fused image is obtained by interpolating and fusing the target image and the generated image. The output of the discriminator includes the true and false judgment of the sample data input to the GAN network and the classification result of the input sample data. The loss function (or objective function) of the discriminator consists of the following two parts:

[0057] L S =E[logP(S=real|x real )]+E[logP(S=fake|x fake )] (1)

[0058] L c =E[logP(C=c|x real )]+E[logP(C=c|x fake )] (2)

[0059] Among them, L S represents the expected likelihood of correctly identifying the target image and the generated image (multi-label loss function), L C represents the expected likelihood probability of the discriminator distinguishing the sample category (multi-category loss function). The training of the generator (G) is to maximize L C -L S , the training of the discriminator (D) is to maximize L C +L S .

[0060] It should be noted that in the GAN network, the class label vector of the input generator and the class label of the input discriminator are essentially the same. Since the input of the generator is in vector form, the class label needs to be integrated with the random vector in the form of a vector and input into the generator together.

[0061] Step S3: Based on the target image, the generated image, the fused image after the target image and the generated image are fused, and the label indicating the category to which the target image belongs, the GAN network is trained to adjust the generated image.

[0062] In step S3, a GAN network model can be obtained by training the GAN network, and the generated image generated by the generator is adjusted to obtain a generated image that meets the requirements.

[0063] In this embodiment, the GAN network is a convolutional neural network, and the structures of the generator and the discriminator may include an input layer, a convolution layer, a deconvolution layer, a pooling layer, and an output layer, without specific limitation.

[0064] Figure 4 FIG. 1 shows a schematic diagram of the structure of the generator and the discriminator in the GAN network model provided by an embodiment of the present invention. Figure 4 As shown, the generator includes a residual connection structure. The residual connection structure in the generator includes multiple residual blocks. Each residual block includes a deconvolution layer (Deconv), a convolution layer (Conv), a block regularization, an activation function, and a fusion layer. The input of the residual block can skip the deconvolution layer (Deconv), the convolution layer (Conv), the block regularization, the activation function and the fusion layer connection to ensure the stability of the network training and the generation of ship sample details.

[0065] Since the ship sample image used in this embodiment is a slice of size 256×256, which is a large-size sample, the generator needs to stack a deep convolution / deconvolution structure to generate it. Therefore, setting a residual connection structure composed of multiple residual blocks in the generator to perform deep convolution / deconvolution operations can not only improve the training performance of the deep network, but also better preserve the structural information of the target image in the SAR image. For example, the detailed expression of the ship texture in the generated image can be enhanced, so that it has a more similar scattering characteristic to the real ship target.

[0066] Block regularization uses batch normalization (BN). Each convolution / deconvolution layer is normalized using a batch normalization layer (BN layer), and then activated using the Leaky Relu activation function. The BN layer speeds up the convergence of the GAN network model, and more importantly, alleviates the problem of "gradient diffusion" in deep networks to a certain extent, making it easier and more stable to train deep network models.

[0067] Each convolutional block structure of the discriminator includes two convolutional layers, and each convolutional layer is followed by a BN layer for normalization, which is then activated using the Leaky Relu activation function. The output layer of the discriminator is connected to the Softmax function, which can be used to classify the input sample data and output the classification results.

[0068] The better the discriminator is trained, the more serious the gradient vanishing of the generator. In this application, by adding a fused image obtained by interpolation between the real sample image (target image) and the generated image to the input layer of the discriminator, the diversity of the generated image is increased, and the problems of difficult and time-consuming calculation of gradient values ​​during back propagation are solved. The introduction of gradient penalty in the discriminator not only improves the convergence speed, but also improves the stability of network training, and can generate higher quality sample images.

[0069] The training of the GAN network includes: inputting random vectors and class label vectors into the generator of the GAN network to train the generator of the GAN network; inputting the generated image, training sample image, and fusion image of the generated image and training sample image generated by the generator into the discriminator of the GAN network to train the discriminator of the GAN network. The training process adopts separate alternating iterative training, first training the discriminator, and then training the generator. The specific steps include the following:

[0070] Step S301: input the random vector and the class label vector into the generator to obtain a generated image.

[0071] In a specific embodiment, the random vector is a multi-dimensional implicit vector (e.g., 100 dimensions) containing a ship target, and the class label vector is a vector representing a ship category. The random vector and the class label vector are sent to the GAN network in batches, and a generated image can be obtained through the generator.

[0072] The size of the generated image matches the size of the training sample image. The training sample image is the training set in the sample data set obtained after normalization preprocessing and slicing of the acquired SAR image in step S1. Since the size of the target image in step S1 is 256×256, the size of the generated image obtained in this step is also 256×256.

[0073] Step S302: input the training sample image, the generated image, the fused image obtained by fusion of the training sample image and the generated image, and the class label into the discriminator for training, and update the discriminator parameters.

[0074] Specifically, equal batches of training sample images, generated images, and fused images are sent to the discriminator together with the class labels for training. C +L S Minimize and update the discriminator parameters. The class label is the classification label of the training sample image. The training sample image, generated image and fused image are input into the discriminator from different input channels.

[0075] Step S303: fix the discriminator parameters, input the random vector and the class label vector into the generator again for training, and update the generator parameters.

[0076] Specifically, the parameters of the discriminator are fixed, the random vector and the class label vector are input into the generator again to obtain the generated image, and the error of the true and false judgment returned by the discriminator in step S302 is used to train the generator, and the objective function L C -l S Maximize and update the generator parameters.

[0077] Step S304: Repeat steps S302 to S303 to iteratively train the generator and the discriminator alternately until the GAN network converges.

[0078] Specifically, the content generated by the discriminator is back-propagated to the generator for gradient update, and the model parameters of the generator and the discriminator are alternately trained to achieve a dynamic balance. After reaching a dynamic balance, the generator of the GAN network model can generate a generated image with a high degree of similarity to the target image.

[0079] During the training process, the number of iterations can be defined. For example, after 30,000 alternating iterations, the GAN network is close to convergence, and the performance of the generator and discriminator networks reaches an ideal Nash equilibrium state. The sample images generated have more similar scattering characteristics to the real sample images.

[0080] As a preferred implementation scheme, the SAR ship image generation method based on the improved GAN-GP algorithm also includes optimizing the GAN network model obtained by training the GAN network, and the specific optimization steps include: inputting the random vector and the class label vector into the trained generator, and outputting the generated image to be tested; inputting the target image to be tested, the generated image to be tested, the fused image to be tested, and the class label into the trained discriminator, and comparing them with the parameters of the discriminator; optimizing the parameters of the GAN network model according to the test of the GAN network model. Among them, the target image to be tested is the image in the test set obtained by step S1; the fused image to be tested is the image obtained by interpolating and fusion of the sample image to be tested and the generated image to be tested; the class label is the classification label of the target image to be tested.

[0081] By optimizing the trained GAN network model, a training set and a final GAN ​​network model that meet the requirements can be obtained; otherwise, the target image is reacquired and the GAN network is trained until a training set and a final GAN ​​network model that meet the requirements are obtained.

[0082] After training the GAN network to obtain the adjusted generated images, the following steps are also included: post-processing the generated images generated by the generator. Specifically, the generated images are summarized and post-processed using indicators such as MS-SSIM and IS to screen out generated slice images that are close to the real target images to form the final generated sample data set.

[0083] The SAR ship image generation method based on the improved GAN-GP algorithm provided in the embodiment of the present invention is used to generate high-quality sample images on the one hand, and can classify the input target image on the other hand. The target image, the generated image, and the fused image are input into the discriminator together with the class label to classify the target image.

[0084] In order to illustrate the performance of the improved GAN network model in the embodiment of the present invention in terms of SAR image ship target generation and classification, the quality of the generated images and the classification accuracy are evaluated.

[0085] Figure 5 The generated images of three types of ships, cargo ships, container ships and oil tankers, are shown using the SAR ship image generation method based on the improved GAN-GP algorithm provided by the embodiment of the present invention. Figure 5As shown in the figure, the sample generation of cargo ships is better than that of the other two types of ship targets. The structure and texture details of the ship are clearer than those of container ships, and the ocean background clutter is less. The main reason is that the number of samples of container ships and tankers is small, and the distinction between small-scale targets is small, which easily makes it impossible to effectively distinguish these two types of ship targets at small scales during the network learning process, causing the generator to be unable to generate high-quality corresponding targets.

[0086] Therefore, in the further study of ship targets, the above-mentioned SAR ship image generation method based on the improved GAN-GP algorithm can be used to generate a sufficient number of labeled ship target images, providing a large number of automatic labeled and diverse effective training sets for subsequent SAR image ship target detection and classification; at the same time, the diverse data set can indirectly improve the generalization ability of the classification network. In addition, the above-mentioned SAR image generation method can reduce the time and manpower costs of manually labeling ship samples.

[0087] Table 1 shows the ship classification accuracy obtained by training the VGG16 classification network model, the ACGAN classification network model and the improved GAN network model in the embodiment of the present invention using the same training data. As shown in Table 1, the ship classification accuracy obtained by training using the improved GAN network model in the embodiment of the present invention is the highest.

[0088] Table 1 Comparison of ship classification accuracy

[0089] VGG16 ACGAN Improved GAN network model cargo ship 0.892 0.876 0.975 container ship 0.952 0.861 0.96 Tanker 0.937 0.872 0.951

[0090] In the embodiment of the present invention, the MS-SSIM (Multi-Scale Structural Similarity Index) index is used to evaluate the quality of the generated ship samples, and the similarity between the ship image generated by the GAN network and the ship target in the real SAR image is evaluated. This index is a full-reference image quality evaluation method, which measures the image similarity from three aspects: brightness, contrast, and structure. The MS-SSIM value range is [0, 1]. The larger the MS-SSIM value (quantized value of the degree of truth), the closer the image is to the reference target.

[0091] Table 2 shows the MS-SSIM values ​​of the generated images generated by the improved GAN network model and the ACGAN network model provided by the embodiments of the present invention. It can be seen from Table 2 that the generated image obtained by using the improved GAN network model of the embodiment of the present invention is closer to the real SAR ship target image and has more realistic texture and structural features. And among the three types of ship targets, the generated image of the cargo ship has the best effect. From the above, it can be seen that the improved GAN network model obtained by training the GAN network in the embodiment of the present invention is better than the other two types of ship targets in feature learning of cargo ships. It can also be seen from the classification accuracy of the discriminator in Table 1 that the feature learning of cargo ships is better than the other two types of ship targets.

[0092] Table 2 MS-SSIM values ​​of ship generated images

[0093]

[0094]

[0095] The classification accuracy of the generated images of ship targets generated by the above-mentioned SAR ship image generation method based on the improved GAN-GP algorithm exceeds 80%, and the similarity evaluation index with the real ship exceeds 90%. The generated images have extremely high real distribution characteristics and texture structure characteristics.

[0096] In order to obtain generated sample images with higher similarity to real ship target images and higher classification accuracy, it is necessary to continue to annotate high-quality cargo ships, container ships, tankers and other ship targets, obtain a large number of training samples, and train a more accurate GAN network model.

[0097] The SAR ship image generation method based on the improved GAN-GP algorithm provided by the embodiment of the present invention can simulate different types of ship targets at different angles and postures, and generate trainable samples that can be used for classification and identification tasks. At the same time, the classification result of the ship target can also be output at the output end of the discriminator part, so that the generated ship target and its category information can be obtained simultaneously in a generative adversarial network. The introduction of the residual network structure in the generator can not only improve the training performance of the deep network, but also better retain the structural information of the ship target in the SAR image. The optimization strategy of gradient penalty is introduced into the conditional generative adversarial network, and the image after the real ship and the generated ship are fused as input to the discriminator of the GAN network to classify different real ship samples, thereby improving the training convergence ability of the GAN network model and the learning ability of different ship target features.

[0098] A second aspect of an embodiment of the present invention provides a SAR ship image generation device based on an improved GAN-GP algorithm, comprising: an acquisition module, configured to acquire a target image; a generation module, configured to obtain a generated image through a GAN network; and a training module, configured to train the GAN network to adjust the generated image based on the target image, the generated image, a fused image after the target image and the generated image are fused, and a label indicating the category to which the target image belongs.

[0099] The SAR ship image generation device based on the improved GAN-GP algorithm provided in the embodiment of the present invention belongs to the same concept as the embodiment of the SAR ship image generation method based on the improved GAN-GP algorithm in the above embodiment. The specific implementation process is detailed in the method embodiment in the above embodiment and will not be repeated here.

[0100] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-mentioned embodiment of the SAR ship image generation method based on the improved GAN-GP algorithm.

[0101] The computer-readable storage medium of the embodiment of the present invention may include any entity or device or recording medium capable of carrying computer program code, for example, ROM / RAM, magnetic disk, optical disk, flash memory and other memories.

[0102] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A SAR ship image generation method based on an improved GAN-GP algorithm is characterized in that: include: Get the target image; Generate images through the GAN network; Based on the target image, the generated image, the fused image after the target image and the generated image are fused, and the label indicating the category to which the target image belongs, the GAN network is trained to adjust the generated image; wherein: The GAN network includes a discriminator and a discriminator loss function associated therewith. The input of the discriminator includes a target image, a generated image, a fused image after interpolation and fusion of the target image and the generated image, and a label indicating the category to which the target image belongs. The output of the discriminator includes a true or false judgment and a classification result of the sample data input into the GAN network; the discriminator loss function includes the loss of the target image and the generated image.

2. The SAR ship image generation method based on the improved GAN-GP algorithm according to claim 1 is characterized in that: The acquiring of the target image comprises: A SAR image containing a target is acquired, and a slicing process is performed on the SAR image to obtain a target image.

3. The SAR ship image generation method based on the improved GAN-GP algorithm according to claim 2 is characterized in that: Before slicing the acquired SAR image, the following steps are also included: The acquired SAR image is preprocessed, and the preprocessing includes performing radiation calibration and normalization processing on the acquired SAR image.

4. The SAR ship image generation method based on the improved GAN-GP algorithm according to claim 1 is characterized in that: The GAN network also includes a generator, the input of the generator includes a random vector and a class label vector, and the output of the generator includes a generated image.

5. The SAR ship image generation method based on the improved GAN-GP algorithm according to claim 1 is characterized in that: The GAN network is a convolutional neural network, and the generator of the GAN network includes a residual connection structure.

6. The SAR ship image generation method based on the improved GAN-GP algorithm according to claim 5 is characterized in that: The residual connection structure includes multiple residual blocks, and each residual block includes: a deconvolution layer, a convolution layer, a block regularization, an activation function, and a fusion layer.

7. The SAR ship image generation method based on the improved GAN-GP algorithm according to claim 4 is characterized in that: The training of the GAN network includes: Input the random vector and class label vector into the generator to obtain the generated image; Input the training sample image, the generated image, the fused image after the training sample image and the generated image are fused, and the class label into the discriminator for training, and update the discriminator parameters; Fix the discriminator parameters, input the random vector and class label vector into the generator again for training, and update the generator parameters; Alternately train the generator and discriminator until the GAN network converges.

8. The SAR ship image generation method based on the improved GAN-GP algorithm according to claim 1 is characterized in that: The method also includes optimizing the GAN network model obtained by training the GAN network, wherein the optimization includes: Input the random vector and the class label vector into the trained generator and output the generated image to be tested; Input the target image to be tested, the generated image to be tested, the fused image to be tested, and the class label into the trained discriminator and compare them with the parameters of the discriminator; Based on the test of the GAN network model, the parameters of the GAN network model are optimized.

9. An image generating device, characterized in that: include: An acquisition module configured to acquire a target image; A generation module, configured to obtain a generated image through the GAN network in the method according to claim 1; A training module is configured to train the GAN network to adjust the generated image based on the target image, the generated image, the fused image after the target image and the generated image are fused, and a label indicating the category to which the target image belongs.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for generating SAR ship images based on the improved GAN-GP algorithm according to any one of claims 1 to 8 are implemented.

Images