An Infrared and Visible Image Fusion Method Based on Progressive Semantic Guided Hierarchical Networks

Abstract

This invention discloses an infrared and visible light image fusion method based on a progressively semantically guided hierarchical network, belonging to the field of computer vision and image processing technology. This method addresses the problems of insufficient global modeling capability, easy loss of detail and texture, and lack of cross-modal semantic consistency in existing fusion techniques by proposing a progressively semantically guided hierarchical fusion network. The steps are as follows: First, input infrared and visible light images, and extract preliminary features through an embedding module; then, use a dual-branch context encoder to extract basic semantic features and detail texture features respectively; in the basic semantic encoding module, perform multi-level information distillation to enhance global semantic expression, while simultaneously using a progressive detail encoding module to gradually enhance local details under semantic guidance; then, fuse the semantic features and detail features across modalities; finally, reconstruct the fused image using a decoder. This method employs a two-stage training strategy to improve the effectiveness of feature extraction and fusion. This invention can effectively improve the information expression capability of the fused image, enhance structural details and semantic consistency, and improve the fusion quality of infrared and visible light images in complex scenes, showing good engineering application prospects.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, specifically to a method for fusing infrared and visible light images based on a progressive semantic-guided hierarchical network. Background Technology

[0002] Infrared and visible light image fusion, as an important technique in multimodal visual processing, combines the brightness of target areas in infrared images with the rich texture details in visible light images, giving the fused image stronger scene resolution capabilities. It has already been applied in various visual perception tasks.

[0003] Existing fusion methods have primarily evolved from traditional multi-scale decomposition methods to deep learning methods. Traditional methods rely on manually designed decomposition and reconstruction strategies, which lack adaptability in complex scenes, fail to retain sufficient detailed information, and have limitations in representing the target region. Deep network methods possess automatic feature extraction capabilities, but convolutional operations are limited by local receptive fields, and their ability to model cross-regional semantic associations remains limited, making it difficult to achieve better fusion results in terms of global structural representation. Furthermore, some adversarial network methods lack stability during training, potentially exhibiting texture loss or slight artifacts.

[0004] From the perspective of comprehensively representing semantic information and high-frequency details, existing methods often struggle to simultaneously leverage the advantages of both: when enhancing salient target regions, the structural details of background regions may be insufficiently represented; conversely, emphasizing local texture enhancement may weaken the overall semantic structure. If cross-modal feature interactions are not adequately designed, the fused image may still have room for improvement in terms of semantic consistency, detail fidelity, and structural coherence. Therefore, existing fusion techniques still have room for further improvement in terms of the overall quality of the fusion results, semantic information representation, and detail recovery capabilities. Developing more efficient collaborative modeling mechanisms for semantics and details remains an important research topic. Summary of the Invention

[0005] This invention proposes a method for fusing infrared and visible light images based on a progressively semantically guided hierarchical network, aiming to address the problems of existing methods in terms of fusion effect, semantic consistency, and detail restoration. This method employs a staged semantically guided network to fuse high-level semantic features and detailed texture information of the image layer by layer, forming a globally and locally unified fusion result. Through the collaborative work of multiple modules, high-quality cross-modal fusion of infrared and visible light images is achieved.

[0006] The method includes the following steps:

[0007] Step 1: Input infrared and visible light images, and perform preliminary feature extraction using the overlapping patch embedding module to generate embedded feature maps;

[0008] Step 2: Input the embedded feature map into the dual-branch context encoder to extract basic semantic features and detailed texture features respectively;

[0009] Step 3: Perform multi-level information distillation and dynamic sparse activation processing on the basic semantic features through the basic semantic encoding module to generate high-level semantic features;

[0010] Step 4: Guided by semantic features, the progressive detail coding module gradually optimizes the detail texture features to generate enhanced detail features;

[0011] Step 5: Perform cross-modal fusion of high-level semantic features and enhanced detail features to generate fused features;

[0012] Step 6: Input the fused features into the dual-branch context decoder to reconstruct and generate the final fused image;

[0013] Step 7: Optimize network parameters through a two-stage training strategy. In the first stage, train the encoder and decoder independently, and in the second stage, jointly optimize the feature fusion module.

[0014] Preferably, the implementation of the dual-branch context encoder includes the following steps:

[0015] S21: Input image I is embedded using the overlap patching module. in Perform preliminary feature extraction to generate an embedded feature map X. embed Its formula is: X embed =Conv embed (I in ).

[0016] S22: The embedded feature map is input into the transformer block, and feature learning is performed through an improved multi-head self-attention mechanism. The query vector Q and key vector K are L2 normalized, and the attention distribution is adjusted by learnable parameters. Then, the value vector V is updated with weights, and finally, the attention feature X is output through residual connections. attn .

[0017] S23: Input the output features of the transformer block into the global context module, enhance cross-modal semantic consistency through global average pooling and weighting operations, and output feature X. global The calculation formula is: X global =X⊙σ(Conv(AvgPool(X))), where σ is the Sigmoid activation function, and ⊙ represents element-wise multiplication. The specific steps are as follows:

[0018] S231: Obtain global context information C through global average pooling. global Its formula is: C global=σ(Conv(AvgPool(X)));

[0019] S232: Channel attention weights are generated using convolutional layers and the sigmoid activation function;

[0020] S233: Multiply the channel attention weights element-wise with the input features to generate the weighted feature X. global Its formula is: X global =X⊙C global ;

[0021] S24: Input the output features of the global context module into the basic semantic coding module and the progressive detail coding module respectively to generate basic semantic features and detail texture features.

[0022] Preferably, the implementation steps of the basic semantic encoding module are as follows:

[0023] S31: Dynamic weights are generated through multi-level information distillation branches, focusing on information-dense regions, and its output feature X distill The calculation formula is: X distill =σ(Conv 1×1 (GELU(Conv 1×1 (X))))⊙X;

[0024] S32: High-frequency residuals are extracted and edge texture information is enhanced through a high-frequency compensation branch, and its output features The calculation formula is: The specific steps are as follows:

[0025] S321: Extract low-frequency components using 3×3 average pooling;

[0026] S322: Calculate the residual between the input features and the low-frequency components as the high-frequency signal;

[0027] S323: High-frequency signals are enhanced sequentially by 3×3 convolution, ReLU activation, 3×3 convolution, and Tanh activation;

[0028] S33: Redundant features are suppressed through dynamic sparse activation branches, and its output feature X sparse The calculation formula is: X sparse =σ(Conv 1×1 (ReLU(Conv 1×1 (GAP(X)))))⊙X;

[0029] S34: The output features of multi-level information distillation, high-frequency compensation, and dynamic sparse activation are fused. The fusion formula is as follows:

[0030] S35: Perform self-attention and feedforward network computations on the fused features to generate high-level semantic features.

[0031] Preferably, the implementation of the progressive detail coding module includes the following steps:

[0032] S41: Input features X∈R 64×H×W The feature is divided into two sub-features z1 and z2, and the formula is: z1, z2 = split(X);

[0033] S42: Via main path We perform inverse residual block and dilated convolution on z1, and standard convolution on z2 through an auxiliary path (ψ), with the formula: φ(z1)=Conv d＝2 (IRB(z1)), ψ(z2)=Conv d＝1 (IRB(z2));

[0034] S43: Dynamic weighted interaction using the Sigmoid gating mechanism, the formula is: z1←z1+σ(Conv1×1(z2))⊙φ(z1), z2←z2+σ(Conv1×1(z1))⊙ψ(z2);

[0035] S44: After concatenating the outputs of all nodes across layers, generate detailed features X through a 1×1 convolution. detail Its formula is: X detail =Conv 1×1 ([z1 (1) z2 (1) ;...;z1 (L) z2 (L) The specific steps are as follows:

[0036] S441: The input features are processed by inverse residual blocks and dilated convolution through the main path to expand the receptive field;

[0037] S442: Standard convolution processing is performed on the input features through an auxiliary path to preserve local texture details;

[0038] S443: By dynamically adjusting the feature weights of the main road and auxiliary road through a gating mechanism, the gradual integration of details is achieved.

[0039] Preferably, the cross-modal feature fusion process in step five includes:

[0040] S51: Weighted fusion of the basic semantic features of infrared and visible light images to generate fused basic semantic features. The formula is:

[0041] S52: Weighted fusion of detail texture features from infrared and visible light images to generate fused detail features. The formula is:

[0042] S53: Input the fused basic semantic features and detailed texture features into the decoder to generate the fused image data_{Fuse}, whose formula is:

[0043] Preferably, the two-stage training strategy in step seven is as follows:

[0044] S71: The first stage involves independently training the encoder and decoder, ensuring the accuracy of feature extraction through reconstruction loss L1, where L1 = L ssim +L mse +L contrastive +L mmd +L perceptual L perceptual For the perceptual loss measured by a pre-trained VGG16 network, L ssim For structural similarity loss, L mse For mean square error loss, L contrastive To compare the losses, L mmd This represents the loss due to the maximum mean difference.

[0045] S72: The second-stage joint optimization feature fusion module optimizes the feature interaction process through the fusion loss L2, where L2 = L fusion +L cc L fusion As a fusion loss, L1 loss based on image brightness and gradient is used to promote cross-modal feature fusion. cc For correlation coefficient loss;

[0046] S73: Employs a cosine annealing learning rate scheduler to dynamically adjust the learning rate, improving training efficiency and model performance. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings will be further described below. Obviously, the drawings are only used to explain the preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0048] Figure 1 This is a schematic diagram of the overall network architecture of the infrared and visible light image fusion method described in this invention.

[0049] Figure 2 This is a schematic diagram of the internal structure of the encoder and decoder described in this invention.

[0050] Figure 3 This is a schematic diagram of the basic semantic encoding module described in this invention, used to extract cross-modal consistent basic semantic information.

[0051] Figure 4 This is a schematic diagram of the progressive detail encoding module described in this invention, used to enhance local texture and detail features.

[0052] Figure 5 This is a qualitative comparison chart of the fusion effect based on the MSRS dataset.

[0053] Figure 6 This is a qualitative comparison chart of the fusion effect based on the TNO dataset.

[0054] Figure 7 This is a qualitative comparison chart of the fusion effect based on the RoadScene dataset. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0056] This invention proposes a method for fusing infrared and visible light images based on a progressive semantic-guided hierarchical network, such as... Figure 1 As shown, the specific implementation method is as follows:

[0057] Step 1: Input infrared and visible light images, and perform preliminary feature extraction using the overlap patch embedding module to generate embedded feature maps. The specific steps are as follows:

[0058] S11: Input infrared image I IR and visible light image I VI The infrared image is a single channel with a size of H×W×1; the visible light image is a three-channel image with a size of H×W×3.

[0059] S12: Perform preliminary feature extraction on the input image using the overlapping patch embedding module. The specific steps are as follows:

[0060] S121: Perform a convolution operation on the input image using a convolutional layer with a kernel size of 3×3, a stride of 1, and padding of 1 to generate an initial feature map:

[0061] S122: Denote the initial feature map as the embedded feature map as X. embed Its calculation formula is: X embed =Conv embed (I in), where Conv embed Indicates the convolution operation, I in Input image (infrared or visible light image);

[0062] S123: The size of the embedded feature map remains consistent with the input, which is H×W×C, where C is the number of embedded channels.

[0063] Step 2: Input the embedded feature map into the dual-branch context encoder, whose internal structure is as follows: Figure 2 As shown, basic semantic features and detailed texture features are extracted respectively. The specific steps are as follows:

[0064] S21: Embed the feature map X embed The input transformer block undergoes feature learning through a multi-head self-attention mechanism and a feedforward network. The specific steps are as follows:

[0065] S211: Generate query (Q), key (K), and value (V) through convolution, calculate similarity attention weights after normalizing Q and K, and update V with weights to obtain multi-head attention features X. attn ;

[0066] S212: The attention output is processed through a feedforward network (FFN) to generate the final output feature X of the transformer block. ffn ;

[0067] S22: Input the output features of the transformer block into the global context module to enhance cross-modal semantic consistency. The specific steps are as follows:

[0068] S221: Extract global context information C using global average pooling. global The formula is: C global =σ(Conv(AvgPool(X))) where σ is the Sigmoid activation function;

[0069] S222: Multiply the global context information element-wise with the input features to generate the weighted feature X. global The formula is: X global =X⊙C global ;

[0070] S23: Input the output features of the global context module into the basic semantic coding module and the progressive detail coding module respectively to generate basic semantic features. B and detailed texture features D .

[0071] Step 3: Using the basic semantic encoding module, the structure of which is as follows: Figure 3As shown, multi-level information distillation and dynamic sparse activation processing are performed on the basic semantic features to generate high-level semantic features. The specific steps are as follows:

[0072] S31: Dynamic weights are generated through multi-level information distillation branches, focusing on information-dense regions. The input feature X generates dynamic weights through a 1×1 convolutional layer and the GELU activation function, and the output feature X is... distill The calculation formula is: X distill =σ(Conv 1×1 (GELU(Conv 1×1 (X))))⊙X;

[0073] S32: Extract high-frequency residuals and enhance edge texture information through a high-frequency compensation branch. The specific steps are as follows:

[0074] S321: Extracting low-frequency components using 3×3 average pooling (AvgPool) 3×3 (X);

[0075] S322: Calculate the high-frequency residual X hf =X-AvgPool 3×3 (X);

[0076] S323: Perform 3×3 convolution, ReLU activation, 3×3 convolution, and Tanh activation sequentially on the high-frequency residuals to output features.

[0077] S33: Suppress redundant features through dynamic sparse activation branches. The specific steps are as follows:

[0078] S331: Perform global average pooling on the input features to generate channel attention weights;

[0079] S332: Multiply the channel attention weights element-wise with the input features to output feature X. sparse ;

[0080] S34: The output features of multi-level information distillation, high-frequency compensation, and dynamic sparse activation are fused together, as shown in the formula:

[0081] S35: Perform self-attention and feedforward network computation on the fused features to generate high-level semantic features. B .

[0082] Step 4, as follows Figure 4 As shown, guided by semantic features, the progressive detail encoding module gradually optimizes the detail texture features to generate enhanced detail features. The specific steps are as follows:

[0083] S41: Input features X∈R 64×H×WThe feature is divided into two sub-features z1 and z2, and the formula is: z1, z2 = split(X);

[0084] S42: Perform inverse residual block (IRB) and dilated convolution on z1 through the main path (φ) to expand the receptive field;

[0085] S43: Perform standard convolution on z2 through the auxiliary path (ψ) to preserve local texture details;

[0086] S44: Dynamically weighted interaction using the Sigmoid gating mechanism, the formula is: z1←z1σ+(Conv 1×1 (z2))⊙φ(z1)),z2←z2σ+(Conv 1×1 (z1))⊙ψ(z2);

[0087] S45: After concatenating the outputs of all nodes across layers, generate detailed features X through a 1×1 convolution. detail .

[0088] Step 5: Perform cross-modal fusion of high-level semantic features and enhanced detail features to generate fused features. The specific steps are as follows:

[0089] S51: Weighted fusion of the basic semantic features of infrared and visible light images to generate fused basic semantic features.

[0090] S52: Weighted fusion of detail texture features from infrared and visible light images to generate fused detail features.

[0091] S53: Input the fused features into the decoder to generate the fused image data. Fuse .

[0092] Step 6: Input the fused features into the dual-branch context decoder to reconstruct the final fused image. The specific steps are as follows:

[0093] S61: Reconstruct the fused image using a dual-branch context decoder. The specific steps are as follows:

[0094] S611: The fused features are processed by a lightweight converter block and a channel compression module;

[0095] S612: Output the final fused image I through the convolutional layer. fuse The dimensions are H×W×3.

[0096] Step 7: Optimize network parameters using a two-stage training strategy. The specific steps are as follows:

[0097] S71: The encoder and decoder are trained independently in the first stage, and the accuracy of feature extraction is ensured by reconstruction loss L1;

[0098] S72: The second-stage joint optimization feature fusion module optimizes the feature interaction process through the fusion loss L2.

[0099] S73: Employs a cosine annealing learning rate scheduler to dynamically adjust the learning rate, thereby improving training efficiency.

[0100] In one embodiment of the present invention, typical scenarios were selected from three public datasets—MSRS, TNO, and RoadScene—for experiments to verify the effectiveness of the fusion method of the present invention. The comparison methods included seven mainstream algorithms: ReCoNet, DID, U2F, TarD, DeF, CDD, and DAF. The comparison results are as follows: Figure 5 , Figure 6 , Figure 7 As shown.

[0101] MSRS dataset corresponding to Figure 5 In infrared images, target thermal radiation features are clear but background details are lacking, while visible light images have complete structural information but insufficient target recognition in low-light areas. The ReCoNet method results in target blurring in low-light areas, the DID method suffers from blurred target edges and loss of background details, and the U2F method over-enhances infrared information, causing building textures to be obscured. The method of this invention, however, achieves clear target outlines, fully preserves road and building structures, maintains balanced brightness without overexposure or blurring, and achieves efficient coordination between target saliency and background detail. (The TNO dataset corresponds to nighttime low-light scenes.) Figure 6 In infrared images, high-temperature targets are clearly identifiable but lack environmental details, while visible light images are rich in structural information but targets in low-light areas are easily concealed. The CDD method causes visual clutter due to excessive background enhancement, and the DAF method is deficient in preserving details of the camouflaged environment. The method of this invention clearly preserves the thermal radiation characteristics of infrared targets, accurately identifies distant building outlines and tree structures, exhibits outstanding edge sharpening effects, and provides excellent naturalness and recognizability in the fused image. The RoadScene dataset corresponds to road traffic scenes. Figure 7Infrared images highlight the thermal radiation of traffic targets but lack texture information, while visible light images have rich texture but poor target visibility in low-light areas. The ReCoNet method suffers from overexposure in strong light areas, and the U2F method leads to excessive smoothing of targets and loss of edge details. The method of this invention effectively improves the visibility of traffic targets, preserving road markings and building outlines intact, with balanced brightness and no overexposure, while also maintaining the infrared salience of distant targets and visible light details in the background. Comprehensive comparison results show that the method of this invention can achieve efficient complementarity of infrared and visible light information in different scenarios, significantly outperforming existing mainstream methods in terms of target recognition, detail preservation, contrast enhancement, and edge sharpness. The fused image quality is superior, and it possesses stronger scene resolution capabilities.

[0102] To objectively evaluate the fusion effect, this invention uses the following seven indicators:

[0103] Information entropy (EN) reflects the richness of image information, and is calculated as follows: Where p(i) is the probability of gray level i occurring. Standard deviation (SD) measures contrast and detail intensity. This represents the pixel mean. Spatial frequency (SF) is used to evaluate texture detail and consists of row frequency (RF) and column frequency (CF). Structural Content Difference (SCD) assesses the retention of structural information, and is calculated as follows: SCD = ∑(|S F -S A |+|S F -S B |), where S F To fuse image structural features, S A S B Features of the source image. Visual fidelity (VIF) assesses cross-scale information retention: Gradient similarity (Qabf): Structural similarity (SSIM) assessment combines brightness, contrast, and structure:

[0104] Higher values ​​for each indicator indicate higher fusion quality. Among them, EN, SD, SF, and SCD reflect the amount of information, while VIF, Qabf, and SSIM assess fidelity. The SCD indicator effectively supplements the quantitative assessment of the degree of structural information preservation by comparing the structural feature differences between the fused image and the source image.

[0105] In one embodiment of the present invention, the above quantitative indicators are used to evaluate the image fusion effect of the proposed method and existing mainstream methods (including ReCoNet, DID, U2F, etc.) on the MSRS, TNO and RoadScene datasets, respectively. The specific results are shown in Tables 1, 2 and 3. The RANK indicator refers to the overall ranking, which is calculated by averaging the rankings of each model on all evaluation indicators, and reflects the overall performance of the model.

[0106] Table 1 Comparison of evaluation indicators obtained by different methods on the MSRS dataset.

[0107]

[0108] Table 2 Comparison of evaluation metrics obtained by different methods on the TNO dataset

[0109]

[0110] Table 3 Comparison of evaluation metrics obtained by different methods on the RoadScene dataset

[0111]

[0112] As shown in Tables 1-3, in the three datasets, the information entropy values ​​of the proposed method (MSRS: 6.7810, TNO: 7.1843, RoadScene: 7.4576) are significantly better than the corresponding values ​​of the comparative methods, indicating that the fusion result retains more complete cross-modal information. The standard deviation (SD) index reaches the highest values ​​on MSRS (45.5845), TNO (46.5713), and RoadScene (55.6058), verifying the advantages of the present invention in enhancing image contrast. The spatial frequency (SF) index is comparable to the optimal method (CDD), showing the effective preservation of texture details. The structural content difference (SCD) and structural similarity (SSIM) indices are both among the top, proving that the semantic guidance mechanism effectively maintains the structural integrity of the image.

[0113] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for fusing infrared and visible light images based on a progressive semantically guided hierarchical network, characterized in that, A progressively semantically guided hierarchical fusion network is adopted. A basic semantic encoding module extracts high-level semantic information and guides a progressive detail encoding module to gradually recover local texture details, forming a hierarchical feature representation system from global to local. A global context module enhances cross-modal semantic consistency. The training process is constrained by joint perceptual loss and structural similarity loss, achieving high-quality fusion of infrared and visible light images. The specific steps are as follows: Step 1: Input infrared and visible light images, and perform preliminary feature extraction using the overlapping patch embedding module to generate embedded feature maps; Step 2: Input the embedded feature map into the dual-branch context encoder to extract basic semantic features and detailed texture features respectively; Step 3: Perform multi-level information distillation and dynamic sparse activation processing on the basic semantic features through the basic semantic encoding module to generate high-level semantic features; Step 4: Guided by semantic features, the progressive detail coding module gradually optimizes the detail texture features to generate enhanced detail features; Step 5: Perform cross-modal fusion of high-level semantic features and enhanced detail features to generate fused features; Step 6: Input the fused features into the dual-branch context decoder to reconstruct and generate the final fused image; Step 7: Optimize network parameters through a two-stage training strategy. In the first stage, train the encoder and decoder independently, and in the second stage, jointly optimize the feature fusion module.

2. The infrared and visible light image fusion method based on a progressive semantic-guided hierarchical network according to claim 1, characterized in that, The implementation steps of the dual-branch context encoder in step two are as follows: S21: Input image I is embedded using the overlap patching module. in Perform preliminary feature extraction to generate an embedded feature map X. embed Its formula is: X embed =Conv embed (I in ); S22: The embedded feature map is input into the transformer block, and feature learning is performed through an improved multi-head self-attention mechanism. The query vector Q and key vector K are L2 normalized, and the attention distribution is adjusted by learnable parameters. Then, the value vector V is updated with weights, and finally, the attention feature X is output through residual connections. attn ; S23: Input the output features of the transformer block into the global context module, enhance cross-modal semantic consistency through global average pooling and weighting operations, and output feature X. global The calculation formula is: X global =X⊙σ(Conv(AvgPool(X))), where σ is the Sigmoid activation function, and ⊙ represents element-wise multiplication. The specific steps are as follows: S231: Obtain global context information C through global average pooling. global Its formula is: C global =σ(Conv(AvgPool(X))); S232: Channel attention weights are generated using convolutional layers and the sigmoid activation function; S233: Multiply the channel attention weights element-wise with the input features to generate the weighted feature X. global Its formula is: X global =X⊙C global ; S24: Input the output features of the global context module into the basic semantic coding module and the progressive detail coding module respectively to generate basic semantic features and detail texture features.

3. The infrared and visible light image fusion method based on a progressive semantic-guided hierarchical network according to claim 1, characterized in that, The implementation steps of the basic semantic encoding module in step three are as follows: S31: Dynamic weights are generated through multi-level information distillation branches, focusing on information-dense regions, and its output feature X distill The calculation formula is: X distill =σ(Conv 1×1 (GELU(Conv 1×1 (X))))⊙X; S32: High-frequency residuals are extracted and edge texture information is enhanced through a high-frequency compensation branch, and its output features The calculation formula is: The specific steps are as follows: S321: Extract low-frequency components using 3×3 average pooling; S322: Calculate the residual between the input features and the low-frequency components as the high-frequency signal; S323: High-frequency signals are enhanced sequentially by 3×3 convolution, ReLU activation, 3×3 convolution, and Tanh activation; S33: Redundant features are suppressed through dynamic sparse activation branches, and its output feature X sparse The calculation formula is: X sparse =σ(Conv 1×1 (ReLU(Conv 1×1 (GAP(X)))))⊙X; S34: The output features of multi-level information distillation, high-frequency compensation, and dynamic sparse activation are fused. The fusion formula is as follows: S35: Perform self-attention and feedforward network computations on the fused features to generate high-level semantic features.

4. The infrared and visible light image fusion method based on a progressive semantic-guided hierarchical network according to claim 1, characterized in that, The implementation steps of the progressive detail coding module in step four are as follows: S41: Input features X∈R 64×H×W The feature is divided into two sub-features z1 and z2, and the formula is: z1, z2 = split(X); S42: Perform inverse residual block and dilated convolution on z1 through the main path (φ), and perform standard convolution on z2 through the auxiliary path (ψ). The formula is: φ(z1)=Conv d＝2 (IRB(z1)), ψ(z2)=Conv d＝1 (IRB(z2)); S43: Dynamically weighted interaction using the Sigmoid gating mechanism, the formula is: z1←z1+σ(Conv 1×1 (z z ))⊙φ(z1), z2←z2+σ(Conv 1×1 (z1))⊙ψ(z2); S44: After concatenating the outputs of all nodes across layers, generate detailed features X through a 1×1 convolution. detail Its formula is: X detail =Conv 1×1 ([z1 (1) z2 (1) ;...;z1 (L) z2 (L) The specific steps are as follows: S441: The input features are processed by inverse residual blocks and dilated convolution through the main path to expand the receptive field; S442: Standard convolution processing is performed on the input features through an auxiliary path to preserve local texture details; S443: By dynamically adjusting the feature weights of the main road and auxiliary road through a gating mechanism, the gradual integration of details is achieved.

5. The infrared and visible light image fusion method based on a progressive semantic-guided hierarchical network according to claim 1, characterized in that, The implementation steps for cross-modal feature fusion in step five are as follows: S51: Weighted fusion of the basic semantic features of infrared and visible light images to generate fused basic semantic features. The formula is: S52: Weighted fusion of detail texture features from infrared and visible light images to generate fused detail features. The formula is: S53: Input the fused basic semantic features and detailed texture features into the decoder to generate the fused image data. Fuse Its formula is:

6. The infrared and visible light image fusion method based on a progressive semantic-guided hierarchical network according to claim 1, characterized in that, The implementation steps of the two-stage training strategy in step seven are as follows: S71: The first stage involves independently training the encoder and decoder, ensuring the accuracy of feature extraction through reconstruction loss L1, where L1 = L ssim +L mse +L contrastive +L mmd +L perceptual L perceptual For the perceptual loss measured by a pre-trained VGG16 network, L ssim For structural similarity loss, L mse For mean square error loss, L contrastive To compare the losses, L mmd This represents the loss due to the maximum mean difference. S72: The second-stage joint optimization feature fusion module optimizes the feature interaction process through the fusion loss L2, where L2 = L fusion +L cc L fusion As a fusion loss, L1 loss based on image brightness and gradient is used to promote cross-modal feature fusion. cc For correlation coefficient loss; S73: Employs a cosine annealing learning rate scheduler to dynamically adjust the learning rate, improving training efficiency and model performance.

Images