Universal image fusion method and system based on diffusion model

Abstract

The invention relates to the technical field of image processing, and discloses a universal image fusion method and system based on a diffusion model, and the method represents an image fusion task as a posterior sampling model through building an image fusion formula based on the diffusion model. And then establishing a potential diffusion model based on a UNet structure, wherein the model retains high-resolution information from original input in a decoding process of fused image reconstruction. And a diffusion trunk based on a Transform model is constructed, and potential features from different source images are effectively integrated to realize cross-domain interaction, so that a fused image with more information content and visual effect is obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a general image fusion method and system based on a diffusion model. Background Art

[0002] Image fusion is a method for producing high-quality images. It is used to fuse multiple multimodal or digital photographic images with complementary information into one image to obtain a fused image with more information and visual effects. However, existing fusion algorithms often find it difficult to maintain the modal characteristics of images from different sources, resulting in problems such as color distortion and halo artifacts in the fusion results. In addition, for images from different imaging devices, such as visible light, infrared, and polarization cameras, there are semantic and modal differences. Simply fusing multi-source images cannot effectively solve the inconsistency between modalities. These problems will cause existing image fusion algorithms to be unable to adapt to different types of image fusion tasks, and the fusion effect is poor.

[0003] At present, general image fusion methods can be divided into two categories: traditional image fusion methods and data-driven image fusion methods. Traditional fusion methods emphasize defining artificially designed prior conditions, but these prior conditions are often difficult to express accurately. In addition, data-driven methods inevitably require the design of complex fusion rules, especially when images come from different sensors or belong to different fields. And they often cannot make full use of the global information present in the image. For example, existing fusion methods often focus on the fusion of local features, while ignoring the global information interaction between domains during the fusion process.

[0004] It can be seen that the general image fusion method in the prior art has the problems of causing color distortion and halo artifacts, and lacking cross-domain interaction of global information. Summary of the invention

[0005] The present invention provides a general image fusion method and system based on a diffusion model to solve the problems of color distortion and halo artifacts caused by the general image fusion method in the prior art and lack of cross-domain interaction of global information.

[0006] In order to achieve the above object, the present invention is implemented by the following technical solutions:

[0007] In a first aspect, the present invention provides a general image fusion method based on a diffusion model, comprising:

[0008] S1: construct a diffusion model based on Markov chain in latent space, wherein the diffusion model includes a forward noise addition process and a reverse denoising process;

[0009] S2: According to the diffusion model, the image fusion task is expressed as a posterior sampling model based on the diffusion model to obtain an image fusion formula;

[0010] S3: Establishing a network structure, and solving the image fusion formula based on the network structure to obtain a final image fusion result.

[0011] In a second aspect, the present application provides a general image fusion system based on a diffusion model, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the first aspect when executing the computer program.

[0012] Beneficial effects:

[0013] The general image fusion method based on the diffusion model provided by the present invention represents the image fusion task as a posterior sampling model by establishing an image fusion formula based on the diffusion model. Then, a potential diffusion model based on the UNet structure is established, which retains the high-resolution information from the original input during the decoding process of fused image reconstruction. And a diffusion backbone based on the Transformer model is constructed to effectively integrate the potential features from different source images to achieve cross-domain interaction, thereby obtaining a fused image with more information and visual effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 One of the flow charts of a general image fusion method based on a diffusion model according to a preferred embodiment of the present invention;

[0015] Figure 2 This is a second flowchart of a general image fusion method based on a diffusion model according to a preferred embodiment of the present invention;

[0016] Figure 3 Schematic diagram of a Transformer-based denoising network according to a preferred embodiment of the present invention;

[0017] Figure 4 This is one of the comparison results of the method of the present application and the existing method on three image fusion tasks in the preferred embodiment of the present invention;

[0018] Figure 5 FIG2 is a comparison result diagram of the method of the present application and the existing method on three image fusion tasks according to the preferred embodiment of the present invention;

[0019] Figure 6 FIG3 is a comparison result diagram of the method of the present application according to the preferred embodiment of the present invention and the existing methods on three image fusion tasks. DETAILED DESCRIPTION

[0020] The technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the usual meanings understood by persons with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one" or "one" do not indicate quantity restrictions, but indicate the existence of at least one. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship also changes accordingly.

[0022] See also Figure 1 , the present application provides a general image fusion method based on a diffusion model, comprising:

[0023] S1: construct a diffusion model based on Markov chain in latent space, wherein the diffusion model includes a forward noise addition process and a reverse denoising process;

[0024] S2: According to the diffusion model, the image fusion task is expressed as a posterior sampling model based on the diffusion model to obtain an image fusion formula;

[0025] S3: Establishing a network structure, and solving the image fusion formula based on the network structure to obtain a final image fusion result.

[0026] The above-mentioned general image fusion method based on diffusion model, by establishing an image fusion formula based on diffusion model, represents the image fusion task as a posterior sampling model. Then, a potential diffusion model based on UNet structure is established, which retains the high-resolution information from the original input during the decoding process of fused image reconstruction. And a diffusion backbone based on Transformer model is constructed to effectively integrate the latent features from different source images to achieve cross-domain interaction, thereby obtaining a fused image with more information and visual effects.

[0027] The steps of the above method are described in detail below with a complete embodiment.

[0028] 1. Construct a diffusion model based on Markov chain in latent space:

[0029] In this embodiment, the diffusion model is used to learn the approximate distribution of data. The model consists of two processes: a forward diffusion process and a reverse denoising process. In the forward process, according to the predefined variance scheduling strategy {β1,...,β T}, iteratively add Gaussian noise ∈ to a clean sample x0~q(x0). As the time step T increases, x T Close to pure Gaussian noise. The forward diffusion process allows sampling at any time step t, and the formula is as follows:

[0030]

[0031] Where t~{0,…,T},α t =1-β t , is the variance scheduling strategy, x t is the noise image at time t. By using the reparameterization technique: At any time step t, the image x can be reconstructed t , where ∈ is Gaussian noise.

[0032] In the reverse denoising process, the diffusion model needs to be trained to learn the denoising process and implement iterative denoising within a predefined time step range. Therefore, the reverse denoising process is defined as a joint distribution p θ (x 0:T ), whose Markov chain starts from p(x T )=N(x T ; 0,1) start:

[0033]

[0034] Where ∈~N(0,I). For x t The variance of is set to zero. In addition, ∈ θ (x t ,t) is reparameterized into a noise prediction network μ with trainable model parameters θ θ (x t ,t). To this end, the diffusion network μ θ (x t ,t) Given an initial sample Predict the denoised sample x under the condition t . The initial sample can be estimated by reparameterizing the sample. The time step of the reverse denoising process is expressed as:

[0035]

[0036] Then, multiple denoising iterations are performed until the final image x0 is generated. The iteration process starts from xt Starts and ends at x0: x T →…→x t-1 →…→x0.

[0037] 2. Establish image fusion formula based on diffusion model:

[0038] The image fusion task is redefined as a denoising process guided by latent features, and the final fused image is generated through multiple iterations. Specifically, let and represents two aligned source images, represents the fused image. Here, H, W, C in Respectively represent the height, width and number of channels of the input image. out is the number of channels of the fused image, and the latent features extracted from s1 and s2 are expressed as d1 = ε(s1) and d2 = ε(s2), respectively, where ε is the pixel space encoder. Given the latent features d1 and d2 of the two source images, the posterior distribution probability of the fused image f is expressed as p θ (f|d1,d2). According to Bayes’ theorem, the posterior distribution is decomposed into p θ (f∣d1,d2)∝p θ (d1,d2|f)·p θ (f). ∝ indicates proportionality, that is, there is no calculation of normalization constant. In order to find the fusion result f with the largest posterior probability * , we use the maximum a posteriori decomposition method:

[0039]

[0040] Among them, the first term is called the data term, which represents the diffusion process guided by the latent features d1, d2, and the second term is the feature term, which encodes the latent guiding features for the previous term; is a neural network that estimates the fused image.

[0041] In this embodiment, the diffusion process is used for implicit learning. Specifically, the forward diffusion process of the image fusion task is defined as a Gaussian transition, denoted as where β t is a predefined noise variance scheduling strategy. Therefore, combining formula (2) and formula (3), the forward diffusion process and backward denoising process of the image fusion task can be obtained as follows:

[0042]

[0043] Formula (26) is expanded into:

[0044]

[0045] in represents the potential features d1 and d2 after adding noise at time t. D0 represents the clean image. Subsequently, in the back diffusion stage, the initial random Gaussian noise D is generated T , by constructing a denoising network μ θ (·) Come and learn D T The distribution characteristics of the noise in the image are used to perform iterative denoising, and finally D T →…→D t-1 →…→D0, the reverse denoising process of the image fusion task is expressed as:

[0046]

[0047] Formula (28) is expanded into:

[0048]

[0049] After establishing the forward and reverse processes of the diffusion model, the fusion result is estimated by the latent features in the reverse process. Specifically, based on the latent feature D obtained from Equation (9) t-1 , design fusion network To estimate the fusion result Through multiple iterations, the fusion network gradually estimates the approximate fusion result in is the estimated potential feature. Finally, the decoder The estimated potential fusion results Mapping from low-dimensional space to high-dimensional space f * .

[0050] 3. Establish network structure:

[0051] Pixel Space Encoder: It is generally believed that it is time-consuming to iteratively run the diffusion model in the pixel space. Performing image fusion tasks in a low-resolution latent space is crucial to reduce computational complexity and improve inference speed. Considering that VAE-GAN is used as a compression model, it needs to be pre-trained to adapt to different image fusion tasks. Therefore, the present invention adopts the UNet model as a pixel-space autoencoder to replace the VAE-GAN model. Compared with the compression model using VAE-GAN as latent diffusion, there are two significant differences: first, the proposed UNet maintains multi-scale details from the encoder to the decoder through skip connections. This can better capture the input image information and provide additional details to the decoder to reconstruct a more accurate fused image. In addition, no pre-training is required; it is trained directly as part of the entire framework in an end-to-end manner. This training strategy does not involve adversarial optimization, making model training more stable than latent diffusion training. The definitions are as follows:

[0052] d1,d2,h=ε(s1,s2) (30);

[0053]

[0054] Where h represents the intermediate features of encoder ε, f * Representation decoder Reconstructed fused image.

[0055] Denoising network: The denoising network mainly consists of two parts: a denoising module and a fusion head module. The denoising module is a four-stage encoder-decoder structure. Each stage includes multiple diffuse transformer (DFT) modules, including diffuse feedforward network (DFN), diffuse self-attention mechanism (DSA) and multi-layer perceptron (MLP). DFN helps the model capture potential features at different positions and enhance its representation ability. DSA assigns attention weights to potential features at different positions to capture pixel dependencies. MLP integrates time into the DFT block by embedding the time step t as the channel scale γ and shift parameter β into the DFT block. Given the time step t of the diffusion model, the MLP is encoded as the channel scale and shift parameters Then, γ1 and β1 are linearly embedded into the given input feature D t Self-attention in:

[0056]

[0057] here is the self-attention mechanism. α is a learnable scaling parameter. C1 and C d They represent 1×1 convolution and 3×3 depthwise convolution respectively. Split is the splitting operation; LN(·) represents layer normalization.

[0058] Next, γ2 and β2 are embedded into the latent feature D t , and use DFN to generate the output:

[0059] DFN(D t ,t)=C1φC1(γ2(LN(D t ))+β2)+D t (33)

[0060] Here φ is the nonlinear activation function GELU.

[0061] Fusion Head: The fusion head module is a decoder composed of four-stage Cross-Domain Fusion Transformer (CFT) blocks, which aims to integrate hierarchical latent features from the denoising network decoder and promote cross-modal interaction for feature fusion. Its structure is very similar to the DFT block, except that the CFT block uses a multi-head cross-attention mechanism to guide the interaction of latent features and uses SKFF for feature aggregation. Specifically, due to the semantic differences in the extracted multimodal latent diffusion features, simply merging multiple information sources cannot effectively model cross-modal inconsistencies. Therefore, cross-modal MCA is introduced to capture the complementarity between multiple latent feature domains and achieve semantic interaction. Among them, the query, key and value of the latent feature d1 are denoted as Q1, K1 and V1, respectively, and the query, key and value of the latent feature d2 are denoted as Q2, K2 and V2, respectively. Therefore, the obtained represents the set of latent features d1, participated by other latent features d2:

[0062]

[0063] Similarly, is calculated as:

[0064]

[0065] Subsequently, the SKFF module is used to integrate the cross-domain features generated by different branches.

[0066]

[0067] here is the potential fusion feature obtained by the l-th layer CFT block.

[0068] Final output Obtained by:

[0069]

[0070] Here Tanh represents the tanh activation function, Conv is a 3×3 convolution, and Concat is an element-wise concatenation operation. When t=0, the potential fusion result is It can be obtained by equation (37). Next, the fused image f * With decoder Reconstructed by equation (31).

[0071] Loss function: supervised data and noise components are necessary for training diffusion models. Therefore, the loss function of the training objective mainly includes diffusion loss and fusion loss to constrain the network:

[0072] L total =L diff +L fusion (38);

[0073] During forward diffusion, the trainable parameters mainly consist of the latent denoising model defined above. Therefore, the model is trained by minimizing the loss between the diffusion output and the denoised prediction:

[0074] L diff =||∈-μ θ (D t ,t)|| 2 (39);

[0075] Where D t is the forward process defined by equation (26) and is calculated over multiple sample steps.

[0076] The purpose of the fusion loss term is to make the generated fusion result closer to the source image in terms of gradient information, structure, pixel intensity distribution, etc. Therefore, the formula is expressed as:

[0077] L fusion =λ1L ssim +λ2L grad +λ3L int (40);

[0078] Where λ1, λ2, λ3 are hyperparameters that control the tradeoff of each sub-loss term. ssim ,L grad ,L int denote the structural similarity loss, gradient loss, and pixel intensity loss respectively. They are defined as:

[0079]

[0080] Where ssim(·) represents the structural similarity operation, which is used to measure the similarity between two images. ▽ is the Sobel gradient operator. |·| is the absolute operation, ||·||1 is the l1 norm, and max(·) is the maximum operation. M(·) is an element fusion operation, where VI-NIR and MEF tasks use element averaging, while other tasks use element maximum operations. In a feasible implementation, w1 and w2 are set to 0.5.

[0081] In one example, the comparison results of the present invention method and the existing methods on three image fusion tasks are as follows: Figure 4-6 As shown, Figure 4 This is a visualization comparison result diagram of the method of the present application and the existing method on the near-infrared and visible light image fusion task; Figure 5 This is a visualization comparison result diagram of the method of this application and the existing methods on the medical image fusion task; Figure 6 This is a visualization comparison result diagram of the method of this application and the existing method on the infrared and visible light image fusion task.

[0082] In summary, the above-mentioned general image fusion method based on the diffusion model, in order to adapt to different image fusion tasks, represents the image fusion task as a posterior sampling model based on the diffusion model, which can be further decomposed into an unconditional denoising problem and a diffusion generation problem. In order to solve the large amount of computing time required for iterative operations of the diffusion model at the pixel level, and the challenge of using VAE-GAN as a compression model and pre-training to adapt to different image fusion tasks, the present invention proposes a pixel space encoder based on the UNet framework. The encoder can effectively compress the input image to a low-resolution latent space, while maintaining and restoring the high-resolution information in the original input during the decoding stage. This strategy not only significantly reduces the computational burden, but also improves the efficiency and accuracy of the image fusion task.

[0083] Furthermore, in order to solve the problem that there are semantic differences in the features of images of different modalities, simply fusing multiple information sources cannot effectively model the inconsistency between cross-modalities. The present invention proposes a Transformer-based denoising network to capture long-range diffusion dependencies and learn hierarchical diffusion representations to understand the cross-domain information between different images, thereby achieving complementarity and semantic interaction of cross-modal image features.

[0084] The embodiment of the present application also provides a general image fusion system based on a diffusion model, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the above method when executing the computer program. The general image fusion system based on a diffusion model can implement various embodiments of the above method and achieve the same beneficial effects, which will not be described in detail here.

[0085] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A general image fusion method based on diffusion model, characterized in that: include: S1: construct a diffusion model based on Markov chain in latent space, wherein the diffusion model includes a forward noise addition process and a reverse denoising process; S2: According to the diffusion model, the image fusion task is expressed as a posterior sampling model based on the diffusion model to obtain an image fusion formula; S3: Establishing a network structure, and solving the image fusion formula based on the network structure to obtain a final image fusion result.

2. The general image fusion method based on diffusion model according to claim 1, characterized in that: The S1 includes: In the forward noise addition process, according to the predefined noise variance scheduling strategy {β1,...,β T }, iteratively add Gaussian noise ∈ to a clean sample x0~q(x0), as the time step T increases, x T Close to pure Gaussian noise, the forward diffusion process allows sampling at any time step t, as follows: In the formula, x t is the noise image at time t, t~{0,…,T},α t =1-β t , By using reparameterization techniques: Reconstruct the noisy image x at any time step t t , N represents Gaussian distribution, I represents the identity matrix, Represents the variance scheduling strategy α t The average value of α i represents the i-th noise variance; In the reverse denoising process, the reverse denoising process is defined as a joint distribution p θ (x 0:T ), whose Markov chain starts from p(x T )=N(x T ; 0,I) start: Among them, ∈~N(0,I); For x t The variance of is set to zero. In addition, ∈ θ (x t ,t) is reparameterized into a noise prediction network μ with trainable model parameters θ θ (x t ,t), diffusion network μ θ (x t ,t) Given an initial sample Predict the denoised sample x under the condition t-1 ; The time step of the reverse denoising process is expressed as: Then, multiple denoising iterations are performed until the final image x0 is generated. The iteration process starts from x t Starts and ends at x0: x T →…→x t-1 →…→x0.

3. The general image fusion method based on diffusion model according to claim 1, characterized in that: The S2 includes: set up and represents two aligned source images, Represents the fused image; where H, W, C in Respectively represent the height, width and number of channels of the input image, represents the field of real numbers, C out is the number of channels of the fused image, and the potential features extracted from s1 and s2 are expressed as d1 = E(s1), d2 = E(s2), where E is the pixel space encoder; Given the latent features d1 and d2 of the two source images, the posterior distribution probability of the fused image f is expressed as p θ (f|d1,d2); According to Bayes' theorem, the posterior distribution is decomposed into p θ (f∣d1,d2)∝p θ (d1,d2|f)·p θ (f); In order to find the fusion result f with the maximum posterior probability * , using the maximum a posteriori decomposition method: Among them, the first term logp(d1,d2|f) is called the data term, which represents the diffusion process guided by the potential features d1,d2, and the second term logp(f) is the feature term, which encodes the potential guiding features of the previous term; is a neural network that estimates the fused image; Implicit learning using diffusion process The details are as follows: The forward diffusion process of the image fusion task is defined as a Gaussian transition, denoted as where β t is a predefined noise variance scheduling strategy. Combining formula (2) with formula (3), the forward diffusion process of the image fusion task is as follows: Formula (6) is expanded to: in, represents the potential features d1 and d2 after adding noise at time t, D0 represents the clean image, and then, in the back diffusion stage, the initial random Gaussian noise D is generated. T , by constructing a denoising network μ θ (·) Come and learn D T The distribution characteristics of the noise in the image are used to perform iterative denoising, and finally D T →…→D t-1 →…→D0, the reverse denoising process of the image fusion task is expressed as: Formula (8) is expanded to: After establishing the forward and reverse processes of the diffusion model, the fusion result is estimated by the latent features in the reverse process. Based on the latent feature D obtained from formula (9), t-1 , designing converged networks To estimate the fusion result Through multiple iterations, the fusion network gradually estimates the approximate fusion result in To estimate the latent features, finally, the decoder The estimated potential fusion results Mapping from low-dimensional space to high-dimensional space f * .

4. The general image fusion method based on diffusion model according to claim 1, characterized in that: The establishing of the network structure comprises: determining a pixel space encoder; Build a denoising network; Construct loss function; A final network structure is determined based on the pixel space encoder, the denoising network and the loss function.

5. The general image fusion method based on diffusion model according to claim 4, characterized in that: The determining of the pixel space encoder comprises: The UNet model is trained as a pixel-space autoencoder. The training strategy does not involve adversarial optimization and is defined as follows: d1,d2,h=E(s1,s2) (10); Among them, h represents the intermediate feature of encoder E, f * Representation decoder The reconstructed fused image, Represents the final potential fusion result.

6. The general image fusion method based on diffusion model according to claim 4, characterized in that: The denoising network includes a denoising module and a fusion head module. The denoising module is a four-stage encoder-decoder structure. Each stage includes multiple diffusion transformers DFT, including a diffusion feedforward network DFN, a diffusion self-attention mechanism DSA and a multi-layer perceptron MLP. DFN is used to capture potential features at different positions and enhance its representation ability. DSA is used to assign attention weights to potential features at different positions to capture pixel dependencies. MLP is used to integrate time into the DFT block by embedding the time step t as the channel scale γ and the shift parameter β into the DFT block. Given a diffusion model time step t, the MLP is encoded as a channel-scale and shift parameters Then, the channel dimensions γ1 and β1 are linearly embedded into the given input feature D t Self-attention in: In the formula, is the self-attention mechanism, α is a learnable scaling parameter, C1 and C d They represent 1×1 convolution and 3×3 depthwise convolution respectively, Split is the splitting operation, and LN(·) represents layer normalization; Next, γ2 and β2 are embedded into the latent feature D t and use DFN to generate the output: DFN(D t ,t)=C1φC1(γ2(LN(D t ))+β2)+D t (13) Where φ is the nonlinear activation function GELU; The fusion head module is a decoder composed of four-level cross-domain fusion transformer CFT blocks, which is used to integrate the hierarchical latent features from the denoising network decoder and promote cross-modal interaction of feature fusion. Cross-modal MCA is introduced to capture the complementarity between multiple latent feature domains and realize semantic interaction. The query, key and value of the latent feature d1 are denoted as Q1, K1 and V1 respectively, and the query, key and value of the latent feature d2 are denoted as Q2, K2 and V2 respectively. Therefore, the obtained represents the set of latent features d1, participated by other latent features d2: In the formula, Split is the splitting operation, MCA represents the cross-modal attention mechanism, and FFN is the feedforward neural network; represents the potential features d1 and d2 after adding noise at time t, express The result calculated by the cross-domain fusion transformer CFT block; in, is calculated as: Then, the SKFF module is used to integrate the cross-domain features generated by different branches: here The potential fusion features obtained for the l-th layer CFT block; Final output Obtained by: Where Tanh represents the tanh activation function, Conv is a 3×3 convolution, and Concat is an element-wise concatenation operation; when t=0, the potential fusion result is It can be obtained by equation (17). Next, the fused image f * With decoder Reconstructed by equation (11).

7. The general image fusion method based on diffusion model according to claim 4, characterized in that: The constructing of the loss function comprises: The loss function of the training objective is used to constrain the network. The loss function includes diffusion loss and fusion loss as follows: L total =L diff +L fusion (18); Where, L diff represents diffusion loss, L fusion represents the fusion loss; The diffusion loss satisfies the following relationship: L diff =||∈-μ θ (D t ,t)|| 2 (19); Where D t It is the forward process defined by equation (7) and is calculated by multiple sample steps; The fusion loss satisfies the following relationship: L fusion =λ1L ssim +λ2L grad +λ3L int (20); Among them, λ1,λ2,λ3 are hyperparameters that control the trade-off of each sub-loss term, L ssim ,L grad ,L int They represent structural similarity loss, gradient loss and pixel intensity loss respectively.

8. A general image fusion system based on a diffusion model, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method described in any one of claims 1 to 7 are implemented.

Images