An unsupervised ship infrared image generation method and system based on polarization physical guidance, a readable storage medium and a program product

Abstract

This invention proposes an unsupervised method, system, readable storage medium, and program product for generating ship infrared images based on polarization physics guidance. The method includes the following steps: Step 1: acquiring a visible light image and polarization images at multiple polarization angles of the ship; obtaining the degree of linear polarization based on the polarization images at multiple polarization angles; generating a polarization degree image of the ship using the degree of linear polarization; Step 2: extracting features from the visible light image and the polarization degree image respectively to obtain visible light features and polarization features, adaptively fusing the visible light features and polarization features to obtain a thermal infrared image of the ship; Step 3: using a polarization-sensing dual-path discriminator to discriminate the thermal infrared image obtained in Step 2, and only using the thermal infrared image that is discriminated as true as the final thermal infrared image.

Description

Technical Field

[0001] This invention relates to the field of ship infrared image generation technology, and in particular to an unsupervised ship infrared image generation method, system, readable storage medium, and program product based on polarization physics guidance. Background Technology

[0002] Visible light and infrared imagery play a synergistic role in maritime perception and surveillance systems. Visible light images offer rich texture details, facilitating target identification; while infrared images, through their thermal radiation characteristics, can effectively detect ship targets in low-visibility conditions such as nighttime and fog. However, due to limitations in imaging conditions and equipment costs, infrared image data is often difficult to acquire reliably, leading to a data scarcity bottleneck for infrared-based intelligent algorithms. Therefore, utilizing artificial intelligence generation technology to convert widely available visible light images into high-quality infrared images represents a significant breakthrough in enhancing all-weather maritime perception capabilities.

[0003] With the continuous development of Generative Adversarial Networks (GANs) image generation methods in the field of cross-modal image transfer, CycleGAN and its improved models, which do not require paired data, have become a research hotspot. In recent years, researchers have attempted to use unsupervised learning methods for style transfer of unpaired images, but existing research still has limitations: First, the methods mainly rely on visual features of a single modality such as visible light, making it difficult to accurately learn the correspondence between these features and infrared images, resulting in thermodynamic distortion in the generated results; second, the models lack explicit constraints on physical laws, causing the temperature gradient changes at key material boundaries in the generated images to not conform to the laws of heat conduction, resulting in insufficient realism; in addition, existing methods are not sensitive to the subtle thermal features of key structures such as metals, which can easily lead to the blurring or loss of discriminative details under unsupervised conditions. Existing methods generally ignore the significant correlation between the polarization optical features of metal surfaces and their thermal radiation in ship imaging.

[0004] In summary, accurate modeling of physical features and effective fusion of cross-modal information are crucial for ship infrared image generation technology. However, existing technologies mainly rely on single-modal information and lack utilization of polarization physical features, resulting in thermodynamic distortion and insufficient physical plausibility of the generated images, making it difficult to meet the practical application needs such as maritime monitoring. Summary of the Invention

[0005] Purpose of the invention: To address the problems in current ship infrared image generation, such as thermodynamic distortion caused by missing paired data, unreasonable temperature gradients due to lack of physical law constraints, insensitivity to thermal characteristics of metal structures, and neglect of the correlation between polarization optical characteristics and thermal radiation, this invention proposes an unsupervised ship infrared image generation method, system, readable storage medium, and program product based on polarization physics guidance, which achieves the fusion of visible light and polarization information to complete the generation of high-quality unsupervised infrared images.

[0006] Technical Solution: In a first aspect, this invention proposes an unsupervised method for generating ship infrared images based on polarization physics guidance, comprising the following steps:

[0007] Step 1: Acquire a visible light image of the ship and polarization images at multiple polarization angles; obtain the degree of linear polarization based on the polarization images at multiple polarization angles; generate a polarization degree image of the ship using the degree of linear polarization.

[0008] Step 2: Extract features from the visible light image and the polarization image respectively to obtain visible light features and polarization features. Adaptively fuse the visible light features and polarization features to obtain the thermal infrared image of the ship.

[0009] Step 3: Use a polarization-sensing dual-path discriminator to discriminate the thermal infrared images obtained in Step 2, and only use the thermal infrared images that are discriminated as true as the final thermal infrared images;

[0010] The polarization-aware dual-path discriminator comprises a global discriminant path and a DoLP-weighted local discriminant path. The global discriminant path is based on the standard PatchGAN architecture and includes five convolutional layers. These five convolutional layers progressively extract multi-scale features, ultimately outputting a two-dimensional discriminant matrix. Each element in this matrix corresponds to the probability distribution of the ground truth of a local region in the input image. The DoLP-weighted local discriminant path is used to stitch features together from the generated thermal infrared image and polarization image. An attention mechanism dynamically adjusts the weights of the global discriminant path and the DoLP-weighted local discriminant path to obtain dual-path features, which are then used to discriminate the thermal infrared image.

[0011] Furthermore, the linear polarization degree is obtained from the polarization image based on multiple polarization angles, and is expressed as:

[0012]

[0013] In the formula, The first three parameters of the Stokes vector are represented as follows:

[0014]

[0015]

[0016]

[0017] In the formula, These are polarization images at different polarization angles.

[0018] Furthermore, the step of extracting features from the visible light image and the polarization image respectively to obtain visible light features and polarization features, and adaptively fusing the visible light features and polarization features to obtain the thermal infrared image of the ship, includes:

[0019] By employing different convolutional kernel initialization strategies, feature extraction is performed on visible light images and polarization images to obtain visible light features and polarization features.

[0020] The contribution weights of visible light features are dynamically adjusted, and visible light features and polarization features are adaptively fused to obtain fused features.

[0021] The fused features are transformed through a network consisting of multiple stacked residual blocks. The transformed features are then spatially upsampled through transposed convolution operations. Each upsampling process fuses the previous layer feature maps at the corresponding scale. Finally, a thermal infrared image is output through a 7×7 convolution kernel and a hyperbolic tangent function.

[0022] Furthermore, the polarization-sensing dual-path discriminator is trained using a total loss function consisting of an adversarial loss function, a cycle consistency loss function, and a gradient consistency loss function.

[0023] Furthermore, the gradient consistency loss function includes a gradient matching term and a physical regularization term.

[0024] Secondly, this invention proposes an unsupervised ship infrared image generation system based on polarization physics guidance, comprising:

[0025] The image acquisition module is configured to acquire visible light images and polarization images of the ship at multiple polarization angles;

[0026] The polarization degree image generation module is configured to obtain the linear polarization degree based on a polarization image with multiple polarization angles; and to generate a polarization degree image of the ship using the linear polarization degree.

[0027] The thermal infrared image generation module is configured to extract features from the visible light image and the polarization image respectively to obtain visible light features and polarization features, and to adaptively fuse the visible light features and polarization features to obtain the thermal infrared image of the ship.

[0028] The discrimination module is configured to use a polarization-sensing dual-path discriminator to discriminate the thermal infrared images obtained by the thermal infrared image generation module, and only use the thermal infrared images that are discriminated as true as the final thermal infrared images.

[0029] The polarization-aware dual-path discriminator comprises a global discriminant path and a DoLP-weighted local discriminant path. The global discriminant path is based on the standard PatchGAN architecture and includes five convolutional layers. These five convolutional layers progressively extract multi-scale features, ultimately outputting a two-dimensional discriminant matrix. Each element in this matrix corresponds to the probability distribution of the ground truth of a local region in the input image. The DoLP-weighted local discriminant path is used to stitch features together from the generated thermal infrared image and polarization image. An attention mechanism dynamically adjusts the weights of the global discriminant path and the DoLP-weighted local discriminant path to obtain dual-path features, which are then used to discriminate the thermal infrared image.

[0030] Furthermore, the step of extracting features from the visible light image and the polarization image respectively to obtain visible light features and polarization features, and adaptively fusing the visible light features and polarization features to obtain the thermal infrared image of the ship, includes:

[0031] By employing different convolutional kernel initialization strategies, feature extraction is performed on visible light images and polarization images to obtain visible light features and polarization features.

[0032] The contribution weights of visible light features are dynamically adjusted, and visible light features and polarization features are adaptively fused to obtain fused features.

[0033] The fused features are transformed through a network consisting of multiple stacked residual blocks. The transformed features are then spatially upsampled through transposed convolution operations. Each upsampling process fuses the previous layer feature maps at the corresponding scale. Finally, a thermal infrared image is output through a 7×7 convolution kernel and a hyperbolic tangent function.

[0034] Furthermore, the polarization-aware dual-path discriminator is trained using a total loss function consisting of an adversarial loss function, a cycle consistency loss function, and a gradient consistency loss function, wherein the gradient consistency loss function includes a gradient matching term and a physical regularization term.

[0035] Thirdly, the present invention proposes a computer-readable storage medium storing computer instructions for causing a processor to execute an unsupervised method for generating ship infrared images based on polarization physics guidance.

[0036] Fourthly, the present invention proposes a computer program product, which includes a computer program that, when executed by a processor, implements an unsupervised method for generating ship infrared images based on polarization physics guidance.

[0037] Beneficial Effects: Compared with existing technologies, this invention establishes a physical mapping relationship between visible light-polarization synergistic features and infrared thermal radiation through a polarization feature fusion module, designs a polarization-sensing dual-path discriminator to improve the ability to distinguish the consistency of thermophysical laws, and introduces gradient consistency loss to enhance the adherence of the generated results to physical laws such as heat conduction. It effectively solves the problem of scarce ship infrared image samples, significantly improves the thermophysical consistency and structural reliability of the generated images, and provides high-quality data support for applications such as maritime monitoring and target recognition. Specifically, it has the following advantages:

[0038] (1) The method of the present invention can realize the unsupervised generation of ship infrared images and automatically convert widely available visible light images, which can greatly alleviate the problem of scarcity of infrared image data, provide high-quality data support for applications such as maritime monitoring and target recognition, and improve all-weather perception capabilities.

[0039] (2) The method of the present invention establishes a physical mapping relationship between visible light, polarization and infrared by fusing visible light and polarization multi-source information, which significantly improves the thermophysical consistency and structural reliability of the generated image and maintains a temperature gradient change that conforms to the law of thermal conduction at the metal boundary;

[0040] (3) The method of the present invention does not require paired training data, saving data acquisition and labeling costs; and can overcome the physical distortion problem caused by single modal features, and preferentially maintain the thermal radiation characteristics of high reflectivity areas through polarization guidance mechanism; the method of the present invention does not require a large number of infrared imaging devices, and only a visible light camera and a polarization camera are needed to generate high-quality infrared images, reducing system deployment costs.

[0041] The system, storage medium, and product provided in the embodiments of this application also have the above-mentioned technical effects.

[0042] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a network structure diagram of an unsupervised ship infrared image method based on polarization physics guidance proposed in this invention;

[0045] Figure 2 Enhance the design of the dual-branch generator;

[0046] Figure 3 Design diagram of a polarization-sensing dual-path discriminator. Detailed Implementation

[0047] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0048] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0049] like Figure 1 As shown, this embodiment discloses an unsupervised ship infrared image method based on polarization physics guidance, the main steps of which include:

[0050] Step 1: Acquire visible light images and four-angle polarization images of the ship using a Hikvision visible light camera and a LUCID polarization camera. Extract the degree of linear polarization using the four-angle polarization images. Specific operations include:

[0051] First, a set of polarization images at different polarization angles are acquired using a polarization camera. .

[0052] Then, the first three parameters of the Stokes vector are calculated. :

[0053]

[0054]

[0055]

[0056] Finally, the degree of linear polarization is obtained by the following formula:

[0057]

[0058] Step 2: Generate a ship polarization image using linear polarization. Extract features from this image and the visible light image separately using independent encoding paths via a dual-branch path, then adaptively fuse them to obtain a thermal infrared image that conforms to physical laws. Specific operations include:

[0059] like Figure 2 As shown, the two branches extract features from the ship polarization image and the visible light image through independent encoding paths, and use different convolution kernel initialization strategies to adapt to the differences in the characteristics of their respective modes:

[0060]

[0061]

[0062] in, and These represent the primary features extracted by the visible light branch and the polarization branch, respectively. ReflectPad represents the reflection padding layer. Both the visible light branch and the polarization branch feature extraction involve operations on the reflection padding layer, which ensures the complete preservation of edge information.

[0063] Next, the features from the two branches are deeply fused using a specially designed PFFM module. The PFFM module utilizes prior knowledge of the metallic region provided by polarization information to dynamically adjust the contribution weights of the visible light features. Its calculation process can be expressed as follows:

[0064]

[0065]

[0066] in, Attention map of the metallic region. is the Sigmoid activation function, and ⊗ represents element-wise multiplication.

[0067] Features after fusion Subsequently, a deep feature transformation is performed using a ResNet with nine residual blocks. The continuous stacking of these nine residual blocks gives the network powerful nonlinear mapping capabilities, enabling it to learn complex transformation relationships from the input domain to the output domain. Next, the network progressively upsamples the feature maps spatially through transposed convolution operations. Each upsampling stage fuses the feature maps from the previous layer at the corresponding scale, enhancing detail reproduction through cross-layer information interaction. Finally, a physically accurate thermal infrared image is output using a 7×7 convolutional kernel and a hyperbolic tangent function (tanh).

[0068]

[0069] Step 3: The thermal infrared image generated in Step 2 is processed by a polarization-sensing dual-path discriminator for image discrimination. The polarization-sensing dual-path discriminator includes a global discrimination path and a DoLP-weighted local discrimination path. The contribution weights of the two paths are dynamically adjusted through an attention mechanism. Specific operations include:

[0070] like Figure 3 As shown, the polarization-aware dual-path discriminator mentioned in this step introduces a DoLP-weighted local discriminant path based on the traditional PatchGAN, which works in conjunction with the original global discriminant path. Specifically, the global discriminant path maintains the ability to perceive the overall consistency of the image, while the DoLP-weighted local discriminant path enhances the discrimination sensitivity of key regions such as metal-nonmetal interfaces and structural edges through the DoLP weight map.

[0071] The global discrimination path, based on the standard PatchGAN architecture, is responsible for global realism discrimination of the entire image. This path contains five convolutional layers, which progressively extract multi-scale features through five convolutional operations, ultimately outputting a two-dimensional discrimination matrix. Each element of this matrix corresponds to the realism probability distribution of a local region in the input image, thereby achieving joint perception of the image's global structure and local consistency.

[0072] The DoLP-weighted local discrimination path specifically enhances the discrimination of thermal features in metallic regions. The input to this path is the stitched features of the generated thermal infrared image and the corresponding DoLP image, which are then weighted as follows:

[0073]

[0074]

[0075] in, It has thermal infrared characteristics. Weighted mask for metallic regions, and These are learnable parameters. This design allows the discriminator to focus on the authenticity of the thermal features of the metallic region, consistent with the physical characteristics of thermal infrared imaging. The features from both paths are dynamically adjusted in the final discrimination stage through an attention mechanism to adjust the contribution weights of the two paths.

[0076]

[0077]

[0078]

[0079] in For the Sigmoid function, , , , These are learnable parameters. This adaptive fusion mechanism ensures that the discriminator maintains globally consistent judgments while enhancing its ability to discriminate key regions.

[0080] Step 4: Add gradient consistency loss, which includes gradient matching and physical regularization terms, to ensure that the temperature distribution of the generated image conforms to the physical laws of heat conduction. Specific operations include:

[0081] Gradient consistency loss is introduced into the original loss function of CycleGAN to enhance the adherence of the generated results to physical laws such as heat conduction. Adversarial loss ensures high visual quality of the generated images, while cyclic consistency loss constrains the bidirectional reconstruction consistency between visible and infrared modes. This comprehensive loss design effectively improves the overall performance of the generated images in terms of visual quality, structural consistency, and physical reliability.

[0082] The adversarial loss is built upon a generative adversarial network (GAN) framework, achieving cross-modal conversion from visible light ship images to infrared images through game-theoretic optimization between the generator and discriminator. Its mathematical expression is as follows:

[0083]

[0084]

[0085] in, It represents The magnitude of the loss in the generator when converting visible light into infrared light; It represents The generator loses light when converting infrared light into visible light. It is the expected value of the real sample of the visible light image. It is the expected value of the real sample of the infrared image.

[0086] In the task of generating infrared images of ships, the original GAN ​​learns the mapping relationship from the visible light domain to the infrared domain, with the goal of making the distribution of generated images consistent with the distribution of real infrared ship images. However, this mapping has high uncertainty; the visible light image of the same ship may be mapped to multiple infrared outputs that conform to the distribution but have inconsistent structures. Cyclic consistency loss effectively solves this problem by introducing physical constraints, becoming a key regularization method for training cross-modal conversion networks for ship images.

[0087] This loss is calculated by constructing a closed-loop reconstruction process of visible light-infrared-visible light, to determine the difference between the original input ship image and the cyclically reconstructed image:

[0088]

[0089] In the formula, It is the distribution of visible light image data of ships. It is the distribution of ship infrared image data.

[0090] To ensure that the temperature distribution of the generated image conforms to the physical laws of heat conduction, a gradient consistency loss is introduced:

[0091]

[0092] The loss function comprises two important parts: a gradient matching term ensures that the generated image maintains consistency with the real image in terms of edge and texture details; and a physical regularization term applies smoothness constraints through a second-order differential operator to avoid generating temperature abrupt changes that do not conform to physical laws. This design allows the model to not only learn the statistical regularities in the data but also to comply with the fundamental physical laws of heat conduction.

[0093] The total loss function is as follows:

[0094]

[0095] in, These are the weights for the adversarial loss function, the cycle consistency loss function, and the gradient consistency loss function, respectively.

[0096] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0097] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0098] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for generating unsupervised ship infrared images based on polarization physics guidance, characterized in that: Includes the following steps: Step 1: Acquire visible light images and polarization images at multiple polarization angles of the ship; Based on polarization images with multiple polarization angles, the degree of linear polarization is obtained; the degree of linear polarization is then used to generate a polarization image of the ship. Step 2: Extract features from the visible light image and the polarization image respectively to obtain visible light features and polarization features. Adaptively fuse the visible light features and polarization features to obtain the thermal infrared image of the ship. Step 3: Use a polarization-sensing dual-path discriminator to discriminate the thermal infrared images obtained in Step 2, and only use the thermal infrared images that are discriminated as true as the final thermal infrared images; The polarization-aware dual-path discriminator comprises a global discriminant path and a DoLP-weighted local discriminant path. The global discriminant path is based on the standard PatchGAN architecture and includes five convolutional layers. These five convolutional layers progressively extract multi-scale features, ultimately outputting a two-dimensional discriminant matrix. Each element in this matrix corresponds to the probability distribution of the ground truth of a local region in the input image. The DoLP-weighted local discriminant path is used to stitch features together from the generated thermal infrared image and polarization image. An attention mechanism dynamically adjusts the weights of the global discriminant path and the DoLP-weighted local discriminant path to obtain dual-path features, which are then used to discriminate the thermal infrared image.

2. The method for generating unsupervised ship infrared images based on polarization physics guidance according to claim 1, characterized in that: The linear polarization degree, obtained from the polarization image based on multiple polarization angles, is expressed as: In the formula, The first three parameters of the Stokes vector are represented as follows: In the formula, These are polarization images at different polarization angles.

3. The method for generating unsupervised ship infrared images based on polarization physics guidance according to claim 1, characterized in that: The process of extracting features from the visible light image and the polarization image to obtain visible light features and polarization features, and then adaptively fusing the visible light features and polarization features to obtain a thermal infrared image of the ship, includes: By employing different convolutional kernel initialization strategies, feature extraction is performed on visible light images and polarization images to obtain visible light features and polarization features. The contribution weights of visible light features are dynamically adjusted, and visible light features and polarization features are adaptively fused to obtain fused features. The fused features are transformed through a network consisting of multiple stacked residual blocks. The transformed features are then spatially upsampled through transposed convolution operations. Each upsampling process fuses the previous layer feature maps at the corresponding scale. Finally, a thermal infrared image is output through a 7×7 convolution kernel and a hyperbolic tangent function.

4. The method for generating unsupervised ship infrared images based on polarization physics guidance according to claim 1, characterized in that: The polarization-sensing dual-path discriminator is trained using a total loss function consisting of an adversarial loss function, a cycle consistency loss function, and a gradient consistency loss function.

5. The method for generating unsupervised ship infrared images based on polarization physics guidance according to claim 4, characterized in that: The gradient consistency loss function includes a gradient matching term and a physical regularization term.

6. An unsupervised ship infrared image generation system based on polarization physics guidance, characterized in that: include: The image acquisition module is configured to acquire visible light images and polarization images of the ship at multiple polarization angles; The polarization degree image generation module is configured to obtain the linear polarization degree based on a polarization image with multiple polarization angles; and to generate a polarization degree image of the ship using the linear polarization degree. The thermal infrared image generation module is configured to extract features from the visible light image and the polarization image respectively to obtain visible light features and polarization features, and to adaptively fuse the visible light features and polarization features to obtain the thermal infrared image of the ship. The discrimination module is configured to use a polarization-sensing dual-path discriminator to discriminate the thermal infrared images obtained by the thermal infrared image generation module, and only use the thermal infrared images that are discriminated as true as the final thermal infrared images. The polarization-aware dual-path discriminator comprises a global discriminant path and a DoLP-weighted local discriminant path. The global discriminant path is based on the standard PatchGAN architecture and includes five convolutional layers. These five convolutional layers progressively extract multi-scale features, ultimately outputting a two-dimensional discriminant matrix. Each element in this matrix corresponds to the probability distribution of the ground truth of a local region in the input image. The DoLP-weighted local discriminant path is used to stitch features together from the generated thermal infrared image and polarization image. An attention mechanism dynamically adjusts the weights of the global discriminant path and the DoLP-weighted local discriminant path to obtain dual-path features, which are then used to discriminate the thermal infrared image.

7. The unsupervised ship infrared image generation system based on polarization physics guidance according to claim 6, characterized in that: The process of extracting features from the visible light image and the polarization image to obtain visible light features and polarization features, and then adaptively fusing the visible light features and polarization features to obtain a thermal infrared image of the ship, includes: By employing different convolutional kernel initialization strategies, feature extraction is performed on visible light images and polarization images to obtain visible light features and polarization features. The contribution weights of visible light features are dynamically adjusted, and visible light features and polarization features are adaptively fused to obtain fused features. The fused features are transformed through a network consisting of multiple stacked residual blocks. The transformed features are then spatially upsampled through transposed convolution operations. Each upsampling process fuses the previous layer feature maps at the corresponding scale. Finally, a thermal infrared image is output through a 7×7 convolution kernel and a hyperbolic tangent function.

8. The unsupervised ship infrared image generation system based on polarization physics guidance according to claim 6, characterized in that: The polarization-sensing dual-path discriminator is trained using a total loss function consisting of an adversarial loss function, a cycle consistency loss function, and a gradient consistency loss function, wherein the gradient consistency loss function includes a gradient matching term and a physical regularization term.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the unsupervised ship infrared image generation method based on polarization physics guidance as described in any one of claims 1-5.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the unsupervised ship infrared image generation method based on polarization physics guidance as described in any one of claims 1-5.

Images