SAR (Synthetic Aperture Radar)-to-optical image conversion technology based on preliminary mapping and post-processing fine tuning

Abstract

The invention discloses an SAR-to-optical image conversion technology based on preliminary mapping and post-processing fine tuning, and relates to the technical field of remote sensing image processing, and the technology comprises the following steps: constructing a mapping path based on a generative adversarial network, and obtaining a preliminary mapping network; constructing a feature extraction model to obtain a post-processing fine tuning network; training the preliminary mapping network and the post-processing fine tuning network through the real optical image and the SAR image; performing test reasoning on the preliminary mapping network and the post-processing fine tuning network through the SAR image to complete conversion from the SAR image to an optical image; compared with the prior art, the method has the advantages that the inter-domain difference between the SAR image and the optical image is reduced, the generated pseudo-optical image has better optical characteristics, the detail quality of the optical image obtained through final conversion is improved, the interpretability of the SAR image to non-professionals is improved, a data user is assisted in interpretation and analysis, and the interpretation efficiency and the application efficiency of the SAR image are improved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a SAR-to-optical image conversion technology based on preliminary mapping and post-processing fine-tuning. Background Technology

[0002] SAR imaging has significant advantages, but its unique imaging mechanism poses many challenges to image interpretation. Furthermore, with the continuous increase in data volume, the need for interpretation is becoming increasingly urgent. Therefore, it is imperative to carry out research on the mapping from SAR to optical images.

[0003] Existing technologies for converting SAR images to optical images face a series of problems. First, they severely lack the ability to adapt to and express SAR image features, failing to effectively capture and present the complex features inherent in SAR images due to their unique imaging mechanisms. Second, they suffer from poor information sensitivity; shallow networks are prone to information loss, making it difficult to perceive subtle feature changes during processing, resulting in the loss of a large amount of image details during the shallow network processing stage. Furthermore, the significant differences between SAR and optical image features make feature mining extremely difficult, hindering the exploration of effective mapping relationships between the two. Finally, the limited number of paired data, especially aligned data for specific land cover categories, severely restricts the sufficiency and diversity of model training, leading to insufficient model generalization ability and seriously affecting the effectiveness and application expansion of SAR image-to-optical image conversion technology.

[0004] Therefore, a SAR-to-optical image conversion technique based on preliminary mapping and post-processing fine-tuning is provided to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a SAR-to-optical image conversion technology based on preliminary mapping and post-processing fine-tuning, which improves the interpretability of SAR images for non-professionals, assists data users in interpretation and analysis, improves interpretation efficiency, and thus enhances the application efficiency of SAR images.

[0006] To achieve the above objectives, this invention provides a SAR-to-optical image conversion technique based on preliminary mapping and post-processing fine-tuning, comprising the following steps:

[0007] S1: Construct a mapping path based on a generative adversarial network to obtain a preliminary mapping network;

[0008] S2: Construct a feature extraction model to obtain a post-processing fine-tuning network;

[0009] S3: Train the initial mapping network and the post-processing fine-tuning network using real optical images and SAR images;

[0010] S4: Test and infer the initial mapping network and the post-processing fine-tuning network using SAR images to complete the conversion from SAR images to optical images.

[0011] Preferably, step S1 specifically includes the following steps:

[0012] S11: Set the first-layer fusion module in the encoder of the generator, and set the dilated convolution module in the middle layer of the generator;

[0013] S12: A domain adaptation module and a feature loss module are set between the generator and the optical image reconstruction network. The domain adaptation module, the feature loss module, and the optical image reconstruction network together constitute the optical image reconstruction module.

[0014] S13: Design loss constraints based on generative adversarial networks.

[0015] Preferably, step S13 specifically includes the following steps:

[0016] Step 1: Design the loss function L of the discriminator cGAN (D), the discriminant's loss function L cGAN (D) Specifically set as follows:

[0017] L cGAN (D)=-E x,y [logD(x,y)]-E x [log(1-D(x,G(x)))]

[0018] Among them, E x,y E represents the expected value of the corresponding calculated result after inputting a SAR image and a real optical image. x This represents the expected result of the calculation after inputting the SAR image, where x represents the SAR image, y represents the real optical image, D(·) represents the output of the generator, and G(·) represents the output of the discriminator.

[0019] Step 2: Design the generator's adversarial loss L cGAN (G), the generator's adversarial loss L cGAN (G) is specifically set as follows:

[0020] L cGAN (G)=E x [log(D(x,G(x)))];

[0021] Step 3: Design the L1 loss L1(G) between the real optical image and the SAR image. The L1 loss L1(G) is specifically set as follows:

[0022] L1(G)=E x,y [||yG(x)||1];

[0023] Step 4: Obtain the total loss function L(G), which is specifically set as follows:

[0024] L(G)=L cGAN (G)+λ L1 L1(G)+λ MMD L MMD (G)+λ features L features (G)

[0025] Among them, L MMD (G) represents the MMD constraint of the domain adaptation module, L features (G) represents the hierarchical feature loss of the decoder for both the generator and the discriminator, λ L1 , λ MMD and λ features Both represent hyperparameters.

[0026] Preferably, step S2 specifically includes the following steps:

[0027] S21: Construct a feature extraction model based on self-attention and long jump connections. The similarity calculation function Attention(Q,K,V) in the self-attention mechanism is specifically set as follows:

[0028]

[0029] Where Q represents the query matrix, K represents the key matrix, V represents the value matrix, and K... T d represents the transpose of the key matrix. k The dimension representing the key;

[0030] S22: Design of noise prediction loss function L based on conditional diffusion model simple (θ), thus obtaining the post-processing fine-tuning network, and the noise prediction loss function L. simple (θ) is specifically set as follows:

[0031]

[0032] in, This represents the expected loss function value obtained through calculation, where ∈ represents the actual noise added at step t. θ Represents a noise prediction network. Indicates noise level. denoted as pseudo-optical image, t represents the number of noise addition steps, and x represents SAR image.

[0033] Preferably, step S3 specifically includes the following steps:

[0034] S31: Input the real optical image and SAR image into the initial mapping network;

[0035] S32: The preliminary mapping network enhances the image information of the SAR image through the first-layer fusion module. The preliminary mapping network uses optical prior knowledge and combines the dilated convolution module to suppress speckle noise in the SAR image and generate a pseudo-optical image.

[0036] S33: The pseudo-optical image, real optical image, and SAR image are transmitted to the post-processing fine-tuning network. The post-processing fine-tuning network improves the image quality of the pseudo-optical image through self-attention mechanism, long jump connection mechanism, and data augmentation strategy to generate the final mapping result.

[0037] Preferably, in step S33, the data augmentation strategy is set to a probabilistic scheduling strategy (w,h)~p for image patch cropping based on PatchDiffusion. (w,h) Probabilistic scheduling strategy (w,h) ~ p (w,h) Specifically set as follows:

[0038]

[0039] Where p represents the ratio of the training images that retain their original size, W represents the width of the original pseudo-optical image-real optical image pair, H represents the height of the original pseudo-optical image-real optical image pair, w represents the width of the original pseudo-optical image-real optical image pair after being cropped into image blocks, and h represents the height of the original pseudo-optical image-real optical image pair after being cropped into image blocks.

[0040] Preferably, step S4 specifically includes the following steps:

[0041] S41: Input the SAR image into the preliminary mapping network;

[0042] S42: The initial mapping network converts the SAR image into a pseudo-optical image with optical features;

[0043] S43: The pseudo-optical image and SAR image are transmitted to the post-processing fine-tuning network. The post-processing fine-tuning network optimizes the pseudo-optical image and generates an optical image corresponding to the SAR image, thus completing the conversion from SAR image to optical image.

[0044] Therefore, the present invention employs the above-mentioned SAR-to-optical image conversion technology based on preliminary mapping and post-processing fine-tuning, which has the following beneficial effects:

[0045] (1) This scheme proposes a feature extraction model based on feature adaptation and speckle suppression, which effectively solves the problem of insufficient adaptation and expressive ability of SAR image features;

[0046] (2) This solution proposes a feature extraction model based on attention and long jump connections, which effectively solves the problems of insufficient information sensitivity and easy loss of information in shallow networks;

[0047] (3) This scheme proposes a sub-network for SAR image and optical image domain adaptation and optical image reconstruction, which effectively solves the problem of large feature differences between SAR images and optical images and the difficulty of feature mining;

[0048] (4) This scheme proposes a data augmentation application mode under the generative model and a data augmentation strategy based on the diffusion model, which effectively solves the problems of limited paired data and scarce aligned data for specific land cover categories.

[0049] The method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0050] Figure 1 This is a flowchart of a SAR-to-optical image conversion technology based on preliminary mapping and post-processing fine-tuning according to the present invention;

[0051] Figure 2 This is a framework diagram of the SAR to optical image mapping training stage of the present invention;

[0052] Figure 3 This is a framework diagram of the SAR to optical image testing and inference stage of this invention;

[0053] Figure 4 This is a schematic diagram of the network framework of the preliminary mapping network of the present invention;

[0054] Figure 5 This is a schematic diagram of the network framework of the post-processing fine-tuning network of the present invention;

[0055] Figure 6 This is a schematic diagram of the input end of the feature extraction model of the present invention;

[0056] Figure 7 This is a schematic diagram of the dilated convolution module of the present invention;

[0057] Figure 8 This is a schematic diagram of the feature extraction model based on attention and long jump connections of the present invention;

[0058] Figure 9 This is a schematic diagram of the long-hop connection mechanism of the present invention;

[0059] Figure 10 This is a schematic diagram of the multi-domain adaptation architecture for SAR and optical images of the present invention;

[0060] Figure 11 This is a schematic diagram of the architecture of the optical image reconstruction subnetwork of the present invention;

[0061] Figure 12 This is a schematic diagram illustrating the applicable modes of the data augmentation strategy of the present invention;

[0062] Figure 13 This is a schematic diagram illustrating the framework of the data augmentation strategy of this invention. Detailed Implementation

[0063] The method of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0064] Unless otherwise defined, the methodological or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0065] The terms "comprising" or "including" as used in this invention mean that the element preceding the term encompasses the element listed after the term, and do not exclude the possibility of encompassing other elements. Terms such as "inner," "outer," "upper," and "lower" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. When the absolute position of the described object changes, the relative positional relationship may also change accordingly. In this invention, unless otherwise explicitly specified and limited, the term "attached" and similar terms should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can refer to a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication of two elements or the interaction relationship between two elements. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0066] Example

[0067] SAR image to optical image translation is a conditional image generation technique that converts SAR images into optical images. It aims to learn the mapping relationship between the two. This embodiment achieves the conversion by combining generative adversarial networks and diffusion models. This technique is of great significance because SAR images are difficult to interpret due to grayscale rendering, geometric distortion, and speckle noise. After being converted into optical images, they can assist non-professionals in interpretation. At the same time, in fields such as military, disaster monitoring, and agricultural detection, it can provide intuitive image information under complex conditions and improve the efficiency of SAR image application.

[0068] like Figures 1-13 As shown, this embodiment provides a SAR-to-optical image conversion technique based on preliminary mapping and post-processing fine-tuning, including the following steps:

[0069] S1: Construct a mapping path based on a generative adversarial network to obtain a preliminary mapping network;

[0070] Step S1 specifically includes the following steps:

[0071] S11: A first-layer fusion module is set in the encoder of the generator. The first-layer fusion module embeds a U-Net network in the first layer of the feature extraction network to enhance the image information of the SAR image and obtain high-quality feature maps. A dilated convolution module is set in the middle layer of the generator. The dilated convolution module suppresses speckle noise by using different dilated convolution kernel sizes, suppresses the speckle noise inherent in the SAR image, and achieves smooth mapping from SAR image to optical image.

[0072] S12: A domain adaptation module and a feature loss module are set between the generator and the optical image reconstruction network. The domain adaptation module, the feature loss module and the optical image reconstruction network together constitute the optical image reconstruction module. The optical image reconstruction module plays a role in guiding the SAR to the optical image mapping in the network and assists in generating pseudo-optical images that are closer to the features of real optical images.

[0073] The optical image reconstruction subnetwork has a similar structure to the SAR-to-optical mapping network. By introducing feature similarity loss, it guides the mapping from SAR to optical images. During training, this subnetwork optimizes the mapping network based on optical prior information and feature similarity loss, making the generated optical images closer to the features of real optical images.

[0074] S13: Design loss constraints based on generative adversarial networks.

[0075] Step S13 specifically includes the following steps:

[0076] Step 1: Design the loss function L of the discriminator cGAN (D), the discriminant's loss function L cGAN (D) Specifically set as follows:

[0077] L cGAN (D)=-E x,y [logD(x,y)]-E x [log(1-D(x,G(x)))]

[0078] Among them, E x,y E represents the expected value of the corresponding calculated result after inputting a SAR image and a real optical image. x This represents the expected result of the calculation after inputting the SAR image, where x represents the SAR image, y represents the real optical image, D(·) represents the output of the generator, and G(·) represents the output of the discriminator.

[0079] Step 2: Design the generator's adversarial loss L cGAN (G), the generator's adversarial loss L cGAN (G) is specifically set as follows:

[0080] L cGAN (G)=Ex [log(D(x,G(x)))];

[0081] Step 3: Design the L1 loss L1(G) between the real optical image and the SAR image. The L1 loss L1(G) is specifically set as follows:

[0082] L1(G)=E x,y [||yG(x)||1];

[0083] Step 4: Obtain the total loss function L(G), which is specifically set as follows:

[0084] L(G)=L cGAN (G)+λ L1 L1(G)+λ MMD L MMD (G)+λ features L features (G)

[0085] Among them, L MMD (G) represents the MMD constraint of the domain adaptation module, L features (G) represents the hierarchical feature loss of the decoder for both the generator and the discriminator, λ L1 , λ MMD and λ features Both represent hyperparameters.

[0086] By introducing MMD constraints, the difference between features extracted from real optical images and features extracted from SAR images is measured and reduced. The multi-layer domain adaptation structure enables the model to better adapt to the feature differences of images in different scenarios, thereby improving the accuracy of mapping.

[0087] After the above processing, the preliminary mapping network can convert SAR images into pseudo-optical images with optical features, effectively suppressing speckle noise in the original SAR image and its negative impact during the mapping process. However, the image has problems such as blurred details and low spatial resolution.

[0088] S2: Construct a feature extraction model to obtain a post-processing fine-tuning network;

[0089] Step S2 specifically includes the following steps:

[0090] S21: Construct a feature extraction model based on self-attention mechanism and long jump connection mechanism to improve the model's ability to perceive large-scale features and information sensitivity. Combine the Canny operator extracted from SAR image to enhance the contour of ground objects, thereby improving the quality of the final mapped image.

[0091] The input end of the feature extraction model is designed for SAR image feature adaptation and speckle suppression. Through specific structural design, the model can better capture the features of SAR images while suppressing speckle noise, providing high-quality feature input for the subsequent mapping process.

[0092] The self-attention mechanism enhances the model's understanding of global image information through the Transformer module, while long skip connections preserve the detailed information of shallow networks, improving the accuracy of feature extraction. The long skip connection mechanism establishes connections between shallow and deep networks, restoring the information loss caused by downsampling, strengthening shallow information, preserving image details, and simplifying training.

[0093] The similarity calculation function Attention(Q,K,V) in the self-attention mechanism is specifically set as follows:

[0094]

[0095] Where Q represents the query matrix, K represents the key matrix, V represents the value matrix, and K... T d represents the transpose of the key matrix. k The dimension representing the key;

[0096] S22: Design of noise prediction loss function L based on conditional diffusion model simple (θ), thus obtaining the post-processing fine-tuning network, and the noise prediction loss function L. simple (θ) is specifically set as follows:

[0097]

[0098] in, This represents the expected loss function value obtained through calculation, where ∈ represents the actual noise added at step t. θ Represents a noise prediction network. Indicates noise level. denoted as pseudo-optical image, t represents the number of noise addition steps, and x represents SAR image.

[0099] S3: The initial mapping network and the post-processing fine-tuning network are trained using real optical images and SAR images. During the training process, data augmentation is performed on the real optical image-SAR image pairs with a certain probability, such as flipping and translating. This increases the amount of training data while avoiding the leakage of augmentation styles into the mapping results. This data augmentation method enriches the training samples of the model and improves the generalization ability of the model.

[0100] Step S3 specifically includes the following steps:

[0101] S31: Input the real optical image and SAR image into the initial mapping network;

[0102] S32: The preliminary mapping network enhances the image information of the SAR image through the first-layer fusion module. The preliminary mapping network uses optical prior knowledge and combines the dilated convolution module to suppress speckle noise in the SAR image and generate a pseudo-optical image.

[0103] S33: The pseudo-optical image, real optical image, and SAR image are transmitted to the post-processing fine-tuning network. The post-processing fine-tuning network improves the image quality of the pseudo-optical image through self-attention mechanism, long jump connection mechanism, and data augmentation strategy to generate the final mapping result.

[0104] In step S33, the data augmentation strategy is set to a probabilistic scheduling strategy (w,h)~p for image patch cropping based on PatchDiffusion. (w,h) When training data is limited, random image segmentation and data augmentation increase the amount of paired data, prevent style leakage, improve training efficiency, and help the model learn more comprehensive features. This enhances mapping performance under data-constrained conditions. The probabilistic scheduling strategy (w,h) ~ p (w,h) Specifically set as follows:

[0105]

[0106] Where p represents the ratio of the training images that retain their original size, W represents the width of the original pseudo-optical image-real optical image pair, H represents the height of the original pseudo-optical image-real optical image pair, w represents the width of the original pseudo-optical image-real optical image pair after being cropped into image blocks, and h represents the height of the original pseudo-optical image-real optical image pair after being cropped into image blocks.

[0107] After processing by the initial mapping network and the post-processing fine-tuning network, the final mapping result combines the advantages of both networks, enabling effective suppression of speckle noise, accurate capture of optical image features, effective supplementation of detail information, and improvement of image quality.

[0108] S4: Test and infer the initial mapping network and the post-processing fine-tuning network using SAR images to complete the conversion from SAR images to optical images.

[0109] Step S4 specifically includes the following steps:

[0110] S41: Input the SAR image into the preliminary mapping network;

[0111] S42: The initial mapping network converts the SAR image into a pseudo-optical image with optical features;

[0112] S43: The pseudo-optical image and SAR image are transmitted to the post-processing fine-tuning network. The post-processing fine-tuning network optimizes the pseudo-optical image and generates an optical image corresponding to the SAR image, thus completing the conversion from SAR image to optical image.

[0113] Therefore, the present invention employs the above-mentioned SAR to optical image conversion technology based on preliminary mapping and post-processing fine-tuning, which not only reduces the inter-domain difference between SAR images and optical images, making the generated pseudo-optical images more optically distinctive, but also increases the detail quality of the final converted optical image, thereby improving the interpretability of SAR images for non-professionals, assisting data users in interpretation and analysis, and improving interpretation efficiency and the application efficiency of SAR images.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the method of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the method of the present invention, and these modifications or equivalent substitutions should not cause the modified method to deviate from the spirit and scope of the method of the present invention.

Claims

1. A SAR-to-optical image conversion technique based on preliminary mapping and post-processing fine-tuning, characterized in that, Includes the following steps: S1: Construct a mapping path based on a generative adversarial network to obtain a preliminary mapping network; S2: Construct a feature extraction model to obtain a post-processing fine-tuning network; S3: Train the initial mapping network and the post-processing fine-tuning network using real optical images and SAR images; S4: Test and infer the initial mapping network and the post-processing fine-tuning network using SAR images to complete the conversion from SAR images to optical images.

2. The SAR-to-optical image conversion technology based on preliminary mapping and post-processing fine-tuning according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11: Set the first-layer fusion module in the encoder of the generator, and set the dilated convolution module in the middle layer of the generator; S12: A domain adaptation module and a feature loss module are set between the generator and the optical image reconstruction network. The domain adaptation module, the feature loss module, and the optical image reconstruction network together constitute the optical image reconstruction module. S13: Design loss constraints based on generative adversarial networks.

3. The SAR-to-optical image conversion technology based on preliminary mapping and post-processing fine-tuning according to claim 2, characterized in that, Step S13 specifically includes the following steps: Step 1: Design the loss function L of the discriminator cGAN (D), the discriminant's loss function L cGAN (D) Specifically set as follows: L cGAN (D)=-E x,y [logD(x,y)]-E x [log(1-D(x,G(x)))] Among them, E x,y E represents the expected value of the corresponding calculated result after inputting a SAR image and a real optical image. x This represents the expected result of the calculation after inputting the SAR image, where x represents the SAR image, y represents the real optical image, D(·) represents the generator output, and G(·) represents the discriminator output. Step 2: Design the generator's adversarial loss L cGAN (G), the generator's adversarial loss L cGAN (G) is specifically set as follows: L cGAN (G)=E x [log(D(x,G(x)))]; Step 3: Design the L1 loss L1(G) between the real optical image and the SAR image. The L1 loss L1(G) is specifically set as follows: L1(G)=E x,y [||yG(x)||1]; Step 4: Obtain the total loss function L(G), which is specifically set as follows: L(G)=L cGAN (G)+λ L1 L1(G)+λ MMD L MMD (G)+λ features L features (G) Among them, L MMD (G) represents the MMD constraints of the domain adaptation module, L features (G) represents the hierarchical feature loss of the decoder for both the generator and the discriminator, λ L1 , λ MMD and λ features Both represent hyperparameters.

4. The SAR-to-optical image conversion technology based on preliminary mapping and post-processing fine-tuning according to claim 2, characterized in that, Step S2 specifically includes the following steps: S21: Construct a feature extraction model based on self-attention and long jump connections. The similarity calculation function Attention(Q,K,V) in the self-attention mechanism is specifically set as follows: Where Q represents the query matrix, K represents the key matrix, V represents the value matrix, and K... T d represents the transpose of the key matrix. k The dimension representing the key; S22: Design of noise prediction loss function L based on conditional diffusion model simple (θ), thus obtaining the post-processing fine-tuning network, and the noise prediction loss function L. simple (θ) is specifically set as follows: in, This represents the expected loss function value obtained through calculation, where ∈ represents the actual noise added at step t. θ Represents a noise prediction network. Indicates noise level. denoted as pseudo-optical image, t represents the number of noise addition steps, and x represents SAR image.

5. The SAR-to-optical image conversion technology based on preliminary mapping and post-processing fine-tuning according to claim 4, characterized in that, Step S3 specifically includes the following steps: S31: Input the real optical image and SAR image into the initial mapping network; S32: The preliminary mapping network enhances the image information of the SAR image through the first-layer fusion module. The preliminary mapping network uses optical prior knowledge and combines the dilated convolution module to suppress speckle noise in the SAR image and generate a pseudo-optical image. S33: The pseudo-optical image, real optical image, and SAR image are transmitted to the post-processing fine-tuning network. The post-processing fine-tuning network improves the image quality of the pseudo-optical image through self-attention mechanism, long jump connection mechanism, and data augmentation strategy to generate the final mapping result.

6. The SAR-to-optical image conversion technology based on preliminary mapping and post-processing fine-tuning according to claim 5, characterized in that, In step S33, the data augmentation strategy is set to a probabilistic scheduling strategy (w,h)~p for image patch cropping based on PatchDiffusion. (w,h) Probabilistic scheduling strategy (w,h) ~ p (w,h) Specifically set as follows: Where p represents the ratio of the training images that retain their original size, W represents the width of the original pseudo-optical image-real optical image pair, H represents the height of the original pseudo-optical image-real optical image pair, w represents the width of the original pseudo-optical image-real optical image pair after being cropped into image blocks, and h represents the height of the original pseudo-optical image-real optical image pair after being cropped into image blocks.

7. The SAR-to-optical image conversion technology based on preliminary mapping and post-processing fine-tuning according to claim 1, characterized in that, Step S4 specifically includes the following steps: S41: Input the SAR image into the preliminary mapping network; S42: The initial mapping network converts the SAR image into a pseudo-optical image with optical features; S43: The pseudo-optical image and SAR image are transmitted to the post-processing fine-tuning network. The post-processing fine-tuning network optimizes the pseudo-optical image and generates an optical image corresponding to the SAR image, thus completing the conversion from SAR image to optical image.

Images