Infrared image synthesis method and system based on cross attention and reinforcement learning
By extracting foreground and background features based on cross-attention and reinforcement learning, and combining cross-attention mechanism and reinforcement learning, the problem of low computational efficiency in infrared image synthesis is solved, and efficient and natural image synthesis is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INSTITUTE OF TECHNOLOGY (SHENZHEN) (INSTITUTE OF SCIENCE AND TECHNOLOGY INNOVATION HARBIN INSTITUTE OF TECHNOLOGY SHENZHEN)
- Filing Date
- 2025-12-17
- Publication Date
- 2026-05-29
AI Technical Summary
Existing infrared image synthesis methods lack interactivity and computational efficiency, making it difficult to achieve efficient and lightweight image synthesis.
A method based on cross-attention and reinforcement learning is adopted. By extracting foreground and background features, and combining cross-attention mechanism and reinforcement learning, a reasonable infrared image synthesis strategy is generated to avoid repeated calculations in multiple time steps.
It significantly improves the computational efficiency and naturalness of infrared image synthesis, realizing an efficient and lightweight image synthesis process.
Smart Images
Figure CN121353458B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of image processing and pattern recognition technology, and in particular to an infrared image synthesis method and system based on cross-attention and reinforcement learning. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] Image synthesis aims to insert foreground objects into a background image at appropriate positions and sizes to generate natural and realistic composite images, and is widely used in image editing, augmented reality, and post-production. Especially in infrared image synthesis, due to the high cost and difficulty in acquiring and annotating real images, how to efficiently synthesize infrared images with physical consistency and target controllability has become a challenging research problem.
[0004] For infrared image synthesis technology, existing synthesis methods can be mainly divided into three categories: The first category uses direct generative models. This type of method usually does not accept user input, and the model automatically completes the prediction of the position and size of foreground objects. It lacks user participation and is difficult to achieve interactive control. The second category introduces random variables and constructs a mapping from random vectors to the synthesis result to generate diverse synthesis strategies. However, these random vectors themselves lack interpretability, and users cannot understand or guide the generation process, so it is still difficult to achieve effective interaction. The third category uses a reinforcement learning framework, which allows specifying the initial position and size and generating reasonable synthesis results through finite time steps. However, this type of method requires the current synthesized image to be input into the complete network for processing at each time step, which has a large computational cost and slow inference speed, making it difficult to meet the efficiency requirements of practical applications.
[0005] In summary, existing infrared image synthesis methods suffer from insufficient guidance and high computational overhead. Therefore, how to achieve efficient and lightweight infrared image synthesis has become an urgent problem to be solved in the current technology. Summary of the Invention
[0006] To address the shortcomings of existing technologies, the present invention aims to provide an infrared image synthesis method and system based on cross-attention and reinforcement learning, which can significantly improve computational and inference efficiency while ensuring the naturalness of the synthesis, thereby achieving more realistic infrared image synthesis.
[0007] To achieve the above objectives, the present invention is implemented through the following technical solution:
[0008] The first aspect of this invention provides an infrared image synthesis method based on cross-attention and reinforcement learning, comprising the following steps:
[0009] Obtain the foreground infrared image and background infrared image to be synthesized;
[0010] An image synthesis model is constructed and trained using a known dataset. The image synthesis model extracts features from the foreground infrared image and the background infrared image, respectively, and performs feature fusion based on a cross-attention mechanism. The fused features are then subjected to reinforcement learning to obtain the image synthesis strategy.
[0011] The image synthesis model is used to perform multiple time steps on the background infrared image, foreground infrared image and initial transformation parameters to be synthesized based on the learned image synthesis strategy, so as to obtain the synthesized infrared image.
[0012] A second aspect of the present invention provides an infrared image synthesis system based on cross-attention and reinforcement learning, comprising:
[0013] The data acquisition module is configured to acquire the foreground infrared image and the background infrared image to be synthesized;
[0014] The model building module is configured to build an image synthesis model and train the image synthesis model using a known dataset. The image synthesis model extracts features from the foreground infrared image and the background infrared image, respectively, and performs feature fusion based on a cross-attention mechanism. The fused features are then subjected to reinforcement learning to obtain the image synthesis strategy.
[0015] The image synthesis module is configured to use an image synthesis model to perform multiple time steps on the background infrared image, foreground infrared image, and initial transformation parameters to be synthesized based on the learned image synthesis strategy, in order to obtain a synthesized infrared image.
[0016] A third aspect of the present invention provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and to execute steps in the infrared image synthesis method based on cross-attention and reinforcement learning as described in the first aspect of the present invention.
[0017] A fourth aspect of the present invention provides a computer device comprising:
[0018] A processor, adapted to execute computer programs;
[0019] A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the infrared image synthesis method based on cross-attention and reinforcement learning as described in the first aspect of the present invention.
[0020] The above one or more technical solutions have the following beneficial effects:
[0021] This invention discloses an infrared image synthesis method and system based on cross-attention and reinforcement learning, overcoming the shortcomings of existing methods such as lack of interactivity, low inference efficiency, and insufficient rationality. This method extracts foreground and background features only once during the inference process, avoiding the overhead of repeated computations across multiple time steps and significantly improving overall operational efficiency. Furthermore, by fusing global features of the foreground and background and introducing a position-aware mechanism, it can achieve natural, rational, and efficient image synthesis.
[0022] This invention discloses an infrared image synthesis method and system based on cross-attention and reinforcement learning. It only requires extracting the background feature map and foreground feature vector once. Based on this, the method encodes the current transformation parameters into a geometric embedding vector at each time step, adds it to the foreground feature vector, and then dynamically fuses it with the background feature map through a cross-attention mechanism to generate the current state representation. The environmental feedback reward is used as the optimization objective to guide the policy network to gradually generate a more reasonable synthesis scheme, thereby significantly improving the computational efficiency while ensuring the naturalness of the synthesis.
[0023] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0025] Figure 1 This is a flowchart of the infrared image synthesis method based on cross-attention and reinforcement learning in Embodiment 1 of the present invention. Detailed Implementation
[0026] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0027] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0028] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0029] Example 1:
[0030] Embodiment 1 of the present invention provides an infrared image synthesis method based on cross-attention and reinforcement learning, such as... Figure 1 As shown, it includes the following steps:
[0031] S1: Obtain the foreground infrared image and background infrared image to be synthesized.
[0032] In one specific implementation, the background infrared image is obtained directly using an infrared device. The foreground outline is manually marked on the infrared image to obtain a foreground mask, and then the foreground is extracted from the infrared image to obtain the foreground infrared image.
[0033] S2: Construct an image synthesis model and train it using a known dataset.
[0034] In one specific implementation, the process of obtaining the known training dataset includes three steps: First, a foreground and a background are randomly selected from the foreground and background infrared images in S1 to construct a foreground-background pair. This step is performed multiple times to obtain a large number of foreground-background pairs. Second, for a foreground-background pair, a foreground mask is used to place the foreground onto the background at a random size and position. This step is performed multiple times to obtain diverse synthetic images. Third, the synthetic images obtained in the second step are manually labeled with a reasonableness rating, where 0 indicates unreasonableness and 1 indicates reasonableness. The known training dataset includes a large number of foreground infrared images and foreground masks, background infrared images, synthetic images, and reasonableness scores.
[0035] The image synthesis model is then trained using the known dataset, specifically including the following steps:
[0036] S21: The image synthesis model extracts foreground infrared image features and background infrared image features respectively.
[0037] It is important to note that the image synthesis model extracts foreground and background infrared image features only once. Specifically, for a foreground-background pair, foreground and background features are extracted only once before multiple time steps (i.e., before repeatedly executing S221, S222, S231, and S232), not only once during the entire training process. This embodiment uses a cross-attention mechanism to dynamically fuse foreground, background, and spatial relationships at each step, thereby ensuring feature utilization comparable to multiple extractions while avoiding redundant computation.
[0038] S211: Use a convolutional neural network to extract background infrared image features, add the background infrared image features to the background position code, and then input them into the Transformer encoder to obtain global background features.
[0039] In one specific implementation, the background infrared image and a mask with all values set to 0 are stitched together and then input into a pre-trained convolutional neural network with fixed parameters to extract background image features. A sine function is then used to obtain the background location code, which is added to the background image features. The shape of the added result is then converted into a background feature sequence, which is finally input into a Transformer encoder to obtain global background features. This process is performed only once.
[0040] S212: Use a convolutional neural network to extract foreground infrared image features and input them into a global average pooling layer to obtain a foreground feature vector.
[0041] In one specific implementation, the foreground infrared image and the foreground mask are input into a pre-trained, parameter-fixed convolutional neural network to extract foreground image features. These features are then input into a global average pooling layer and their shape is converted into a vector to obtain the foreground feature vector. This process is performed only once.
[0042] S22: Feature fusion based on cross-attention mechanism.
[0043] S221: Encode the current transformation parameters into a geometric embedding vector and add it to the foreground feature vector to obtain the position-aware foreground features.
[0044] In one specific implementation, the current transformation parameters are converted into bounding box coordinates, and then the bounding box coordinates are processed using a sine function to obtain a geometric embedding vector. The geometric embedding vector is then added to the foreground feature vector to obtain the position-aware foreground features.
[0045] It is important to note that during training, the transformation parameters for the first time step are initialized through random sampling to simulate different positions and sizes of the foreground within the background. In this way, the model can learn the interaction patterns between the foreground and background under various possible transformations during training, thereby improving the policy network's ability to generate reasonable synthesized images. The transformation parameters for subsequent time steps are obtained by acquiring new transformation parameters in step S231.
[0046] This embodiment introduces geometric embedding to explicitly encode the positional information of the foreground, enabling subsequent feature interactions to not only depend on the content but also to perceive the spatial position of the foreground in the background, thereby improving the accuracy of the synthesis.
[0047] Specifically, using transformation parameters This indicates the position and size of the foreground within the background, where This indicates the scaling ratio of the entire foreground. Therefore, scaling the width of the foreground... for And height for ,in and These represent the width and height of the foreground and background images, respectively. Assuming... This indicates the coordinates of the top-left pixel of the foreground in the background image, using... and This indicates the relative horizontal and vertical positions where the foreground should be placed on the background infrared image. At time step... Transform parameters Convert to bounding box coordinates ,in , as well as Then, a sine function is used to process the bounding box coordinates to obtain a geometric embedding vector, which is then added to the foreground feature vector to obtain the position-aware foreground features.
[0048] S222: The cross-attention mechanism is used to dynamically fuse global background features and position-aware foreground features to obtain the current observation.
[0049] In one specific implementation, since the foreground features have incorporated spatial location information, the cross-attention mechanism can adaptively find the region corresponding to the foreground in the global background, achieving joint alignment of semantics and geometry, thereby obtaining a more reasonable synthesis effect.
[0050] Position-aware foreground features are used as the Query (Q) in the cross-attention mechanism, and global background features are used as the Key (K) and Value (V) in the cross-attention mechanism. The current observation is obtained by dynamic fusion through a fusion formula. :
[0051] .
[0052] in, These are the feature dimensions for Query, Key, and Value.
[0053] S23: Reinforcement learning is performed on the fused features to obtain an image synthesis strategy.
[0054] S231: Input the current observation into the Actor-Critic network to generate action probability distribution and state value estimate, and select an action to execute to obtain new transformation parameters or terminate.
[0055] In one specific implementation, the current observation is input into the LSTM in the Actor-Critic network to incorporate historical observations. and current observations Get the current state Then, the input is fed into the Actor branch to obtain a 7-dimensional action probability distribution. These 7 actions are: zoom in, zoom out, move right, move left, move down, move up, and terminate. Input to the Critic branch to obtain the state value estimate Then sample an action from a multivariate distribution. And execute it, thereby obtaining new transformation parameters or terminating.
[0056] In the Actor-Critic network, the first layer is an LSTM layer used to extract state features by combining current observations and historical information. The output of the LSTM is then fed into two parallel branches—the Actor branch and the Critic branch—each containing a fully connected layer. The Actor branch generates the action probability distribution, while the Critic branch outputs the state value estimate.
[0057] S232: Use the new transformation parameters to obtain the synthesized image and the synthesized mask, and input them into the pre-trained image synthesis evaluation network to obtain a reasonableness score and calculate the reward.
[0058] In one specific implementation, a spatial transformation network is used to process the new transformation parameters. This spatial transformation network includes a localization network, a mesh generator, and a sampler. The localization network outputs the spatial transformation parameters (in this embodiment, the Actor branch), and the mesh generator generates a sampling coordinate mesh of the target image based on the spatial transformation parameters. The sampler uses bilinear interpolation to sample corresponding pixels from the foreground image and the mask to obtain the transformed foreground image and the composite mask.
[0059] The specific steps include:
[0060] First, the new transformation parameters are used with a localization network. Transformation parameters at time t+1 The formula is:
[0061] .
[0062] in, These are the spatial transformation parameters.
[0063] Then, using the mesh generator according to... Generate a sampling coordinate grid.
[0064] The sampler then uses bilinear interpolation to sample corresponding pixels from the foreground image and the mask to obtain a new foreground image. and synthetic mask .
[0065] Finally, the new synthesized image is obtained using the following formula:
[0066] .
[0067] in, For the new synthetic image, Background infrared image.
[0068] New synthetic image and synthetic mask Input into a pre-trained image synthesis evaluation network The formula for predicting reasonableness scores and calculating rewards is as follows:
[0069] .
[0070] If the current time step reaches the maximum number of time steps Alternatively, if a termination action is performed, proceed to step S233; otherwise, return to step S221 to reacquire the location-aware foreground features. In this embodiment, the image synthesis evaluation network is only used during the training process to provide environmental rewards to optimize the policy network, and does not participate in computation during the inference phase, thereby improving synthesis efficiency.
[0071] In this embodiment, the image synthesis evaluation network is built based on the Swin-Transformer and consists of multiple consecutive Swin-Transformer modules and a fully connected layer. The input to each Swin-Transformer module is the stitched result of the synthesized image and the synthesized mask. The Swin-Transformer module extracts feature representations of the image and mask through self-attention operations, and finally outputs a scalar as the reasonableness score of the synthesized image through the fully connected layer. During the network training phase, the output reasonableness score is optimized by calculating an L1 loss with the labeled reasonableness score, enabling the network to accurately evaluate the reasonableness of the foreground within the background. During the image synthesis strategy training process, the network has fixed parameters. The synthesized image and corresponding mask generated at each time step are input into the network, and the output reasonableness score is used to calculate the reward, guiding the Actor-Critic network to optimize the object synthesis strategy.
[0072] S233: Using environmental feedback rewards as the optimization objective, update network parameters and learn reasonable image synthesis strategies.
[0073] In one specific implementation, environmental feedback rewards are used. As an optimization objective, the image synthesis strategy is optimized by maximizing the advantage function and policy entropy to update the Actor parameters and minimizing the error between the prediction value and the cumulative reward to update the Critic parameters, thereby achieving reasonable and efficient image synthesis.
[0074] Specifically, at each time step Cumulative rewards The calculation formula is:
[0075] .
[0076] in, It is a discount factor. It's a reward. Indicates the state The value output below, The network parameters representing the Critic branch are... The value range is 0 to .
[0077] Update Actor parameters by using environmental feedback rewards as the optimization objective. The formula is:
[0078] .
[0079] in, To maximize the advantage function, For policy entropy, This represents the gradient operation with respect to the network parameters, used to calculate the derivative of the loss function or optimization objective with respect to the parameters. This indicates a parameter update operation, which adjusts the current parameter values according to the gradient direction to optimize the network.
[0080] The Critic parameter is updated by minimizing the error between the predicted value and the cumulative reward. This optimizes the synthesis strategy. This achieves efficient and reasonable image synthesis. The formula is:
[0081] + .
[0082] S3: Using an image synthesis model, based on the learned image synthesis strategy, perform multiple time steps on the background infrared image, foreground infrared image, and initial transformation parameters to obtain the synthesized infrared image.
[0083] In one specific implementation, after training is complete, the background infrared image, the foreground infrared image, and the user-specified initial transformation parameters are used as input, and the learned image synthesis strategy is executed to maximize... The transformation parameters are updated at each time step to obtain a reasonable synthesis result. During testing, the action with the highest probability is selected based on the learned policy to find a reasonable synthesis result.
[0084] Example 2:
[0085] Embodiment 2 of the present invention provides an infrared image synthesis system based on cross-attention and reinforcement learning, comprising:
[0086] The data acquisition module is configured to acquire the foreground infrared image and the background infrared image to be synthesized;
[0087] The model building module is configured to build an image synthesis model and train the image synthesis model using a known dataset. The image synthesis model extracts features from the foreground infrared image and the background infrared image, respectively, and performs feature fusion based on a cross-attention mechanism. The fused features are then subjected to reinforcement learning to obtain the image synthesis strategy.
[0088] The image synthesis module is configured to use an image synthesis model to perform multiple time steps on the background infrared image, foreground infrared image, and initial transformation parameters to be synthesized based on the learned image synthesis strategy, in order to obtain a synthesized infrared image.
[0089] Example 3:
[0090] Embodiment 3 of the present invention provides a computer-readable storage medium storing a computer program adapted for loading by a processor and executing the steps of the infrared image synthesis method based on cross-attention and reinforcement learning as described in Embodiment 1 of the present invention.
[0091] Example 4:
[0092] Embodiment 4 of the present invention provides a computer device, the device comprising:
[0093] A processor, adapted to execute computer programs;
[0094] A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the steps in the infrared image synthesis method based on cross-attention and reinforcement learning as described in Embodiment 1 of the present invention.
[0095] The steps and methods involved in Examples 2, 3 and 4 above correspond to those in Example 1. For specific implementation details, please refer to the relevant description section of Example 1.
[0096] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0097] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data processing device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).
[0098] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. An infrared image synthesis method based on cross-attention and reinforcement learning, characterized in that, Includes the following steps: Obtain the foreground infrared image and background infrared image to be synthesized; An image synthesis model is constructed and trained using a known dataset. The image synthesis model extracts features from the foreground infrared image and the background infrared image, respectively, and performs feature fusion based on a cross-attention mechanism. The fused features are then subjected to reinforcement learning to obtain the image synthesis strategy. The image synthesis model extracts foreground infrared image features and background infrared image features only once. The specific steps for the image synthesis model to extract foreground infrared image features and background infrared image features are as follows: Background infrared image features are extracted using a convolutional neural network. The background infrared image features are then added to the background position code and input into the Transformer encoder to obtain global background features. Convolutional neural networks are used to extract foreground infrared image features, which are then input into a global average pooling layer to obtain a foreground feature vector. The image synthesis model is used to perform multiple time steps on the background infrared image, foreground infrared image and initial transformation parameters to be synthesized based on the learned image synthesis strategy, so as to obtain the synthesized infrared image.
2. The infrared image synthesis method based on cross-attention and reinforcement learning as described in claim 1, characterized in that, The specific steps for feature fusion based on the cross-attention mechanism are as follows: The current transformation parameters are encoded into a geometric embedding vector and added to the foreground feature vector to obtain the position-aware foreground features; The cross-attention mechanism is used to dynamically fuse global background features and position-aware foreground features to obtain the current observation.
3. The infrared image synthesis method based on cross-attention and reinforcement learning as described in claim 2, characterized in that, The specific steps for obtaining the image synthesis strategy by performing reinforcement learning on the fused features are as follows: The current observation is input into the Actor-Critic network to generate action probability distributions and state value estimates, and actions are selected for execution to obtain new transformation parameters; The synthesized image and synthesized mask are obtained using new transformation parameters and then input into a pre-trained image synthesis evaluation network to obtain a reasonableness score and calculate a reward. Using environmental feedback rewards as the optimization objective, the network parameters are updated to learn reasonable image synthesis strategies.
4. The infrared image synthesis method based on cross-attention and reinforcement learning as described in claim 3, characterized in that, The image synthesis evaluation network is only used during the training process.
5. The infrared image synthesis method based on cross-attention and reinforcement learning as described in claim 3, characterized in that, Using environmental feedback rewards as the optimization objective, the Actor parameters are updated by maximizing the advantage function and policy entropy, and the Critic parameters are updated by minimizing the error between the prediction value and the cumulative reward, thereby optimizing the image synthesis strategy.
6. An infrared image synthesis system based on cross-attention and reinforcement learning, characterized in that, include: The data acquisition module is configured to acquire the foreground infrared image and the background infrared image to be synthesized; The model building module is configured to build an image synthesis model and train the image synthesis model using a known dataset. The image synthesis model extracts features from the foreground infrared image and the background infrared image, respectively, and performs feature fusion based on a cross-attention mechanism. The fused features are then subjected to reinforcement learning to obtain the image synthesis strategy. The image synthesis model extracts foreground infrared image features and background infrared image features only once. The specific steps for the image synthesis model to extract foreground infrared image features and background infrared image features are as follows: Background infrared image features are extracted using a convolutional neural network. The background infrared image features are then added to the background position code and input into the Transformer encoder to obtain global background features. Convolutional neural networks are used to extract foreground infrared image features, which are then input into a global average pooling layer to obtain a foreground feature vector. The image synthesis module is configured to use an image synthesis model to perform multiple time steps on the background infrared image, foreground infrared image, and initial transformation parameters to be synthesized based on the learned image synthesis strategy, in order to obtain a synthesized infrared image.
7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and executed as described in any one of claims 1-5: the infrared image synthesis method based on cross-attention and reinforcement learning.
8. A computer device, characterized in that, include: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the infrared image synthesis method based on cross-attention and reinforcement learning as described in any one of claims 1-5.