A high dynamic range image reconstruction method based on exposure guide feature decoupling

Abstract

The application discloses a high dynamic range image reconstruction method based on exposure guide feature decoupling, which comprises the following core components: an encoder for projecting an LDR image to an HDR domain and extracting multi-scale features; a global spatial attention modeling module for capturing long-range spatial dependence across exposure frames; a frequency information semantic consistency module for realizing implicit alignment through semantic cross-attention, and differentially modeling and complementarily fusing low-frequency structural information and high-frequency texture information in the frequency domain; a LightMLP lightweight feature refining module for retaining local structural details under low parameter complexity; and a decoder for gradually restoring spatial resolution and reconstructing an HDR image in combination with a skip connection; the reconstructed HDR image has clearer texture and more complete structure, and is superior to most existing methods in multiple objective indicators.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image processing technology, specifically involving technologies such as deep learning and multi-exposure image fusion, especially high dynamic range image reconstruction tasks based on deep learning. Background Technology

[0002] High dynamic range (HDR) imaging technology is one of the important research directions in computer vision and image processing. It is mainly used to solve the problem that traditional image sensors have limited dynamic range and are unable to fully record large-scale brightness changes in natural scenes. For complex scenes that contain both bright and dark areas, traditional image sensors often cannot capture information from both bright and dark areas in a single exposure. This can easily lead to oversaturation and loss of detail in bright areas, as well as low signal-to-noise ratio and lack of structural information in dark areas, thus affecting the visual quality of the image and the processing effect of subsequent visual tasks.

[0003] To address the aforementioned issues, the HDR reconstruction method based on multi-exposure image fusion does not require significant modifications to existing imaging hardware. It only requires acquiring multiple low dynamic range (LDR) images under different exposure conditions in the same scene, and then fusing them through an algorithm to obtain a high dynamic range image. Therefore, it has advantages such as low cost, flexible implementation, and wide applicability, and has become an important technical route for HDR imaging.

[0004] However, in practical applications, multi-exposure HDR image reconstruction methods still face numerous technical challenges. Since multiple exposure images are typically acquired sequentially, they are susceptible to factors such as camera shake, displacement of moving objects in the scene, and changes in lighting during acquisition, leading to spatial misalignment and inconsistencies between different exposure images. Especially in dynamic scenes, directly fusing multiple exposure images can easily produce noticeable ghosting artifacts, edge blurring, and structural distortion in moving areas, severely degrading the quality and stability of the HDR reconstruction results.

[0005] To address the problem of multi-exposure image fusion in dynamic scenes, existing technologies have proposed various methods based on optical flow estimation, explicit registration, convolutional neural networks, attention mechanisms, and Transformers. Early methods often used optical flow estimation or image registration to align images from different exposures; however, due to significant brightness differences between these images, the registration results were often inaccurate. While subsequent deep learning-based methods have mitigated the impact of registration errors to some extent, existing methods mostly focus on unified feature extraction and fusion from a data-driven perspective, lacking specific modeling of the effectiveness, reliability, and physical imaging characteristics of information in different exposure images.

[0006] Furthermore, existing HDR fusion methods generally employ an exposure-insensitive feature processing approach, that is, uniformly encoding, aligning, and fusing feature information from images with different exposures, without adequately distinguishing the differences in information contribution of each exposed image in different regions and frequency components. In reality, different exposed images have significant physical complementary characteristics during the imaging process: underexposed images can usually better preserve high-frequency details such as texture and edges in bright areas, but are susceptible to noise interference in dark areas; overexposed images can better preserve overall illumination and low-frequency structural information in shadow areas, but in bright areas, high-frequency details are irrecoverably lost due to sensor saturation.

[0007] Despite significant progress made by existing methods in this field, the following challenges remain:

[0008] 1. It is difficult to effectively suppress ghosting artifacts caused by camera shake and target motion in dynamic scenes.

[0009] 2. Failure to fully utilize the physical complementarity of images exposed at different times in the frequency domain and illumination response level resulted in insufficient extraction of effective information and prominent noise and distortion problems.

[0010] 3. The model has high complexity and computational cost, and lacks explicit guidance on the direction of fusion constraints during training.

[0011] To address the aforementioned issues, this invention proposes a high dynamic range image reconstruction method based on exposure-guided feature decoupling. This invention designs a frequency information semantic consistency module, achieving implicit alignment across exposure frames through semantic cross-attention, and performing differentiated modeling and complementary fusion of low-frequency structural information and high-frequency texture information in the frequency domain. Simultaneously, a two-level progressive reliability mask refinement strategy and a two-sample HDR contrast loss function are designed to provide physically reliable fusion constraints for model training. The specific structure is shown in the attached figure. Figure 1 As shown. Summary of the Invention

[0012] In view of this, the purpose of this application is to provide a high dynamic range image reconstruction method based on exposure-guided feature decoupling, the main process of which is shown in the attached figure. Figure 1 As shown, the main steps include: HDR domain projection and input representation construction, encoder feature extraction, global spatial attention modeling, frequency information semantic consistency module to achieve exposure-guided feature decoupling and complementary fusion, LightMLP lightweight feature refinement, and decoder reconstruction output HDR image.

[0013] The specific details of the high dynamic range image reconstruction method based on exposure-guided feature decoupling are as follows:

[0014] HDR domain projection and input representation construction:

[0015] To improve the model's robustness to different exposure differences, this invention first projects each input LDR image into the HDR domain. Let the i-th frame input LDR image be... Its corresponding exposure time is Then its HDR domain projection This can be expressed as Formula 1:

[0016]

[0017] Wherein, γ is the gamma coefficient, a positive real number, preferably 2.2. Subsequently, the original LDR image Li is stitched together with the corresponding HDR domain projection Hi along the channel dimension to obtain a 6-channel input representation Ii for each frame. That is, each frame retains both the brightness and color representation in the original LDR space and introduces the irradiance prior from the HDR domain, thereby improving the adaptability of subsequent feature extraction to exposure changes.

[0018] Encoder and shallow feature extraction:

[0019] The k 6-channel input images are fed into an encoder for feature extraction. The encoder includes multiple convolutional residual blocks for progressively extracting multi-scale representations and compressing spatial resolution. The encoder consists of three downsampled convolutional residual blocks and one non-downsampled feature refinement block. The first downsampled block uses a 3×3 convolutional kernel with a stride of 2, has 6 input channels and 64 output channels, and includes a batch normalization layer and a LeakyReLU activation function; the second downsampled block uses a 3×3 convolutional kernel with a stride of 2, expanding the number of channels from 64 to 128; the third downsampled block uses a 3×3 convolutional kernel with a stride of 2, expanding the number of channels from 128 to 256; the feature refinement block uses a 3×3 convolutional kernel with a stride of 1, further refining the 256-dimensional features to output deep semantic features. After the above processing, the spatial resolution is downsampled to 1 / 8 of the original image. Simultaneously, the encoder retains shallow features for subsequent skip connections to supplement texture and edge details.

[0020] Shallow features were extracted from each frame. This invention fuses shallow features from each frame and performs initial interaction through a multilayer perceptron (MLP) to form initial fused features for global spatial modeling. This design can introduce cross-frame contextual information while preserving local information of each exposure image, laying the foundation for subsequent long-range dependency modeling and exposure-aware feature separation.

[0021] Global Spatial Attention Modeling Module (G-SAB):

[0022] To capture the long-range spatial dependencies between different exposure frames, this invention sets up a global spatial attention block (G-SAB) after the encoder, the structure of which is shown in the attached figure. Figure 2As shown. This module employs a window-based multi-head self-attention mechanism to globally model the fused features. The input features are first divided into a token sequence via patching, and then sequentially processed through layer normalization, window-based multi-head self-attention, and a point-by-point multilayer perceptron to enhance the spatial correlation modeling capability between different exposure frames. Its forward propagation can be expressed as Equations 2 and 3:

[0023]

[0024]

[0025] Where LN represents layer normalization, W-MSA represents window multi-head self-attention, and MLP represents multilayer perceptron. The window size is 8×8, the number of multi-heads is 8, the feature dimension is 256, the MLP expansion ratio is 4, and the number of G-SAB stacking layers L is preferably 6. Through the above design, the model's adaptability to large-scale motion, complex occlusion, and brightness changes can be enhanced without significantly increasing the computational burden.

[0026] Frequency Information Semantic Consistency Module (FISC):

[0027] To address the noise amplification, detail loss, and ghosting issues caused by non-perceptual fusion in existing methods, this invention designs a Frequency Information Semantic Consistency (FISC) module after G-SAB, the structure of which is shown in the attached figure. Figure 3 As shown. The core of this module lies in: on the one hand, using cross-exposure semantic correlation to achieve implicit alignment, and on the other hand, performing differentiated modeling of low-frequency structural information and high-frequency texture information in the frequency domain to achieve feature decoupling and complementary fusion of exposure perception.

[0028] First, the G-SAB output features are divided into k subgroups, and residual associations are performed with the shallow features of the corresponding exposure frames to obtain the linked features of each exposure frame, as shown in Equation 4:

[0029]

[0030] This design can re-inject shallow detail information of each exposure frame into the deep semantic representation, enhancing the ability of subsequent features to preserve exposure attributes.

[0031] Secondly, for non-reference frames, this invention predicts a reliable attention map through an attention generation submodule. This submodule includes two layers of 3×3 convolutions and a Sigmoid activation, with an output value ranging from 0 to 1, used to indicate the effective regions in the current non-reference frame that can be used for fusion and the misaligned regions that need to be suppressed. The attention map can pre-identify regions affected by motion misalignment, saturation distortion, or noise contamination before feature fusion, thereby suppressing the propagation of ghost artifacts.

[0032] Furthermore, this invention introduces a semantic cross-attention module (S-CA), the structure of which is shown in the attached figure. Figure 4 As shown in Equation 5, a query vector Q is generated using the link features of the reference frame, and a key vector K and a value vector V are generated using the features of the non-reference frame. Explicit semantic alignment across frames is achieved through a semantic relevance matrix.

[0033]

[0034] in, The key vector dimension is preferably 32. Unlike traditional explicit optical flow or registration methods, this invention does not require precise pixel-level alignment of the input image. Instead, it achieves robust interaction in the feature space through semantic relevance. For misaligned ghosting regions, since they are semantically inconsistent with the reference frame, the corresponding attention weights will approach zero, thus being automatically suppressed.

[0035] After completing cross-exposure semantic interaction, this invention further performs frequency domain decomposition, dividing the aggregated features into low-frequency and high-frequency branches. The low-frequency branch is mainly responsible for integrating stable illumination and structural information, preferably extracting relatively complete low-frequency structures in the shadow areas of overexposed images; the high-frequency branch is mainly responsible for enhancing details such as texture and edges, preferably extracting high-frequency information retained in the highlight areas of underexposed images. By fusing the outputs of the low-frequency and high-frequency branches through a dynamic weighting mechanism, the model can adaptively select the most reliable information source based on the physical characteristics of images with different exposures, thereby improving the structural integrity and texture fidelity of HDR images.

[0036] LightMLP Lightweight Feature Refinement Module:

[0037] The fusion results are fed into the LightMLP module for lightweight refinement, the structure of which is shown in the attached figure. Figure 5 As shown. The LightMLP includes 1×1 convolutions for channel blending, 3×3 depthwise separable convolutions for local spatial context modeling, and residual connections to preserve local structural details with low parameter complexity. The parameter complexity of LightMLP is approximately 2C², where C is the channel dimension. The depthwise separable convolutions compensate for the shortcomings of the attention mechanism in modeling local spatial relationships.

[0038] Decoder and HDR image output:

[0039] The fused features processed by the FISC module are input into the decoder to progressively restore spatial resolution and reconstruct the HDR image. The decoder consists of multiple upsampled convolutional residual blocks, with the number of channels in each layer set sequentially to 256, 128, 64, 32, and 3, and employs bilinear interpolation or deconvolution to achieve progressive upsampling. During decoding, skip connections from shallow features of the encoder are introduced to compensate for edge blurring and detail loss that may occur when relying solely on deep semantic features. The final output is a 3-channel linear domain HDR image.

[0040] Two-stage progressive reliability mask refinement strategy:

[0041] To improve the effectiveness of contrast constraints during the training phase, this invention further designs a two-stage progressive reliability mask generation strategy, the process of which is shown in the attached figure. Figure 6 As shown, this is used to filter physically reliable regions suitable for applying contrastive learning constraints.

[0042] First, construct the first-level mask. Calculate the global brightness mean from the ground-value HDR image. and standard deviation The threshold selection is then performed by combining pixel brightness and local contrast. The first-level mask can be expressed as Formula 6:

[0043]

[0044] in, This represents a measure of low brightness at pixel x. λ represents a measure of local high contrast, and is the shadow threshold sensitivity coefficient, preferably 1.0. The contrast threshold is preferably 0.1. This mask prioritizes preserving low-brightness shadow areas suitable for recovering structure from overexposed frames, as well as high-contrast areas containing rich edge textures.

[0045] Based on the first-level mask, this invention further constructs a second-level mask. The pre-trained VGG-16 network is used to extract perceptual features between candidate regions and ground truth regions, and the feature distance is calculated. When this distance exceeds the texture threshold When this happens, the corresponding candidate region is removed from the mask. A second-level mask can be represented by Formula 7:

[0046]

[0047] in, The optimal value is 0.5. Through a two-level mask progressive filtering, the effective area of ​​contrastive learning can be constrained from both the dimensions of physical reliability and texture realism.

[0048] Model training methods:

[0049] This invention combines reconstruction loss and two-sample HDR contrast loss to simultaneously optimize the brightness restoration, structural consistency, and exposure complement fusion capabilities of HDR images.

[0050] Considering the heavy-tailed characteristic of the HDR linear domain brightness distribution, this invention first performs μ-law compression on the ground truth HDR image H and the predicted HDR image, as shown in Equation 8:

[0051]

[0052] Where μ is a positive real number, preferably 5000. Then, it is defined in the compression domain. Reconstruction loss, as shown in Formula 9:

[0053]

[0054] This loss can create a more balanced gradient between highlight and shadow areas, which helps improve the overall reconstruction stability and visual consistency.

[0055] Refining the mask Guided by this principle, the present invention constructs dual positive and dual negative samples for HDR contrast learning. Positive samples include: one class emphasizing detail recovery in low-exposure images, and another class emphasizing the preservation of highlight gradients and saturation edges in high-exposure images. Negative samples include: a class of noise artifact variants formed by injecting Gaussian perturbations into high-risk highlight regions, and a class of structural mismatch variants constructed through irrelevant mask reconstruction. The dual positive NT-Xent type contrast target can be expressed as Equation 10:

[0056]

[0057] Where sim(·) is the cosine similarity and τ is the temperature parameter, preferably 0.07.

[0058] The overall training objective of this invention is defined by Formula 11:

[0059]

[0060] in, and These are non-negative weight coefficients, preferably 1.0 and 0.13 respectively. The contrast branch is only enabled during the training phase and requires no additional computational overhead during the inference phase.

[0061] Beneficial effects: Compared with existing multi-exposure HDR image reconstruction methods, the present invention has the following significant advantages:

[0062] 1. This invention improves the robustness of the model to exposure differences by projecting images with different exposures into the HDR domain and constructing an exposure-enhanced input representation, enabling subsequent feature extraction to more effectively adapt to brightness changes under different exposure conditions.

[0063] 2. This invention combines global spatial attention with semantic cross-attention to achieve robust cross-frame semantic alignment without explicit optical flow registration, effectively suppressing ghost artifacts in dynamic scenes and avoiding estimation errors in traditional registration methods under large brightness differences.

[0064] 3. This invention uses a frequency domain decomposition mechanism to model and complementarily fuse low-frequency structural information in overexposed images and high-frequency texture information in underexposed images, respectively, which significantly improves the structural consistency and texture fidelity of the reconstructed images and realizes the utilization of differentiated information in exposure perception.

[0065] 4. This invention provides physically reliable and texture-consistent constraint directions for model training through a two-level reliability mask and a two-sample HDR contrast loss, thereby enhancing the recovery stability of difficult regions.

[0066] 5. The comparison branch of this invention only participates in optimization during the training phase and does not increase the additional overhead during the inference phase, which is beneficial for practical deployment and application.

[0067] 6. The experimental results in the specific embodiments described below confirm the effectiveness and superiority of the present invention. Attached Figure Description

[0068] Figure 1 This is a general framework diagram of a multi-exposure HDR image reconstruction method based on exposure-guided feature decoupling;

[0069] Figure 2 This is a structural diagram of the Global Spatial Attention Modeling Module (G-SAB);

[0070] Figure 3 This is a structural diagram of the Frequency Information Semantic Consistency Module (FISC);

[0071] Figure 4 This is a structural diagram of the Semantic Cross-Attention Module (S-CA);

[0072] Figure 5 This is a structural diagram of the LightMLP lightweight feature refining module;

[0073] Figure 6 This is a schematic diagram of the two-level reliability mask generation strategy.

[0074] Figure 7 It is a subjective quality comparison of scene images based on the Kalantari dataset;

[0075] Figure 8 This is a scenario example using the Tel dataset. Detailed Implementation

[0076] This invention constructs a high dynamic range image reconstruction network (EFD-HDRNet) based on exposure-guided feature decoupling, the flowchart of which is attached. Figure 1 As shown, the specific steps include:

[0077] Step 1: Acquire k low dynamic range (LDR) images (preferably k = 3) under different exposure conditions in the same scene, and select a medium exposure image as the reference frame. Project each LDR image into the HDR domain according to Formula 1, and stitch it with the original LDR image in the channel dimension to obtain a 6-channel input representation for each frame.

[0078] Step 2: Input the 6-channel images of each frame separately into the encoder for feature extraction. The encoder consists of 3 downsampled convolutional residual blocks (stride of 2, channel counts of 6→64→128→256 respectively) and 1 non-downsampled feature refinement block, with spatial resolution downsampled to 1 / 8 of the original image. Shallow features from each frame are fused and initially interacted through a multilayer perceptron to form initial fused features. Step 3: Extract shallow features from the input low-resolution image using the convolutional encoder. Then, the extracted image features and the mixed prior obtained in Step 2 are fed into the multimodal feature fusion module. This module is built based on a bidirectional cross-attention mechanism, allowing mutual correction and guidance between prior information and image features.

[0079] Step 3: Input the initial fused features into the global spatial attention modeling module G-SAB, and perform global spatial modeling through a window-based multi-head self-attention mechanism (window size 8×8, number of multi-heads 8, feature dimension 256, MLP expansion ratio 4, number of stacking layers 6).

[0080] Step 4: Input the G-SAB output features into the Frequency Information Semantic Consistency (FISC) module. First, the features are segmented into k subgroups and associated with the shallow feature residuals; for non-reference frames, a reliability attention map is predicted through the attention generation submodule; the semantic cross-attention module S-CA generates Q for reference frames and K and V for non-reference frames to achieve cross-frame semantic alignment; then, in the frequency domain, the features are divided into low-frequency and high-frequency branches, processed separately, and fused through dynamic weighting.

[0081] Step 5: Feed the fusion result into the LightMLP module for lightweight refinement. This module achieves channel blending through 1×1 convolution, local spatial context modeling through 3×3 depthwise separable convolution, and has residual connections.

[0082] Step 6: Input the refined fused features into the decoder, combine them with the skip connections of the shallow features of the encoder, upsample step by step and reconstruct the output 3-channel linear domain HDR image.

[0083] Training Implementation

[0084] This invention uses two representative publicly available dynamic multi-exposure HDR datasets for training and validation. The first is the Kalantari et al. dataset, which contains 74 training scenes and 15 test scenes, featuring significant foreground motion and relatively complex background motion. The second is the Tel et al. dataset, which contains 108 training scenes and 36 test scenes, characterized primarily by significant foreground motion and / or camera shake. During training, the teacher model is trained for 500 epochs using the Adam optimizer, with a batch size of 64. The initial learning rate is set to 0.001, decaying to 0.2 times the original rate every 100 epochs. The student model also uses the Adam optimizer, trained for 200 epochs, maintaining a batch size of 64. Its initial learning rate is 0.001, decaying to 0.2 times the original rate every 40 epochs.

[0085] For data preprocessing, during the training phase, the input image was offline cropped into 128×128 image patches with a cropping stride of 64. Both the validation and testing phases were performed at the original full resolution. During training, data augmentation operations such as random horizontal flipping, random vertical flipping, and random 90° rotation were implemented, and geometric transformations were simultaneously applied to multiple exposure images of the same scene.

[0086] Regarding training hyperparameters, the model of this invention is trained for 500 epochs, with a batch size of 4, and the optimizer is Adam (β1=0.9, β2=0.999, ...). The initial learning rate is set to... The weighting coefficient of the reconstruction term in the total loss function decays once every 100 epochs. =1.0, weight coefficient of the comparison item =0.13. The experiment was implemented using PyTorch and completed on a single NVIDIA RTX 3080 Ti GPU.

[0087] Test Implementation

[0088] To comprehensively evaluate the performance of this invention in HDR image reconstruction, four complementary metrics are used for integrated evaluation: PSNR-1 (peak signal-to-noise ratio in the linear domain), SSIM-1 (structural similarity in the linear domain), PSNR-μ (peak signal-to-noise ratio in the μ-law tone mapping domain), and SSIM-μ (structural similarity in the μ-law tone mapping domain). Higher values ​​for all four metrics indicate better performance.

[0089] This invention selects several representative multi-exposure HDR image reconstruction methods as comparison objects, including Sen (TOG'12), Hu (CVPR'13), Kalantari (TOG'17), AHDRNet (CVPR'19), PAMNet (ACCV'20), NHDRRnet (TIP'20), HDR-GAN (TIP'21), FHDRNet (CVIU'24), CA-ViT (ECCV'22), and SCTNet (ICCV'23).

[0090] Table 1 presents the quantitative comparison results between the present invention and existing technologies on the Kalantari et al. dataset. The present invention achieves the best results in all three metrics: PSNR-1, SSIM-1, and PSNR-μ, reaching 42.50, 0.9899, ​​and 44.53, respectively. This indicates that the present invention can not only improve the reconstruction accuracy of HDR images in linear irradiance space, but also maintain better brightness consistency and detail fidelity at the perceptual level.

[0091] Table 1. Quantitative comparison of this invention with other methods on the Kalantari et al. dataset.

[0092]

[0093] Table 2 presents the quantitative comparison results on the Tel et al. dataset. This invention achieves optimal results in the three metrics of SSIM-1, PSNR-μ, and SSIM-μ. Although there is a very small difference in PSNR-1 compared to the optimal method (47.49 vs 47.51), it shows a more significant advantage in the tone mapping domain metric, which better reflects perceived quality.

[0094] Table 2. Quantitative comparison of this invention with other methods on the Tel et al. dataset.

[0095]

[0096] The above experimental results fully demonstrate that the high dynamic range image reconstruction method based on exposure-guided feature decoupling proposed in this invention has excellent reconstruction quality, robustness and generalization ability in dynamic scenes, and is superior to or close to the existing best methods in many objective evaluation indicators.

[0097] This invention provides a high dynamic range image reconstruction method based on exposure-guided feature decoupling, relating to the fields of image processing and deep learning. The method can run on a computer system, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When executed, the computer program implements the high dynamic range image reconstruction process based on exposure-guided feature decoupling.

Claims

1. A high dynamic range image reconstruction method based on exposure-guided feature decoupling, characterized in that, The process includes the following steps: First, k low dynamic range images with different exposure conditions captured in the same scene are projected into the HDR domain and stitched with the original LDR image in the channel dimension to obtain a multi-channel input representation for each frame; then, shallow and deep features of each frame are extracted by an encoder; and finally, long-range spatial dependency modeling is performed on the fused features by a global spatial attention modeling module. Then, through the frequency information semantic consistency module, implicit alignment across exposure frames is achieved by using semantic cross-attention. In the frequency domain, features are divided into low-frequency structure branches and high-frequency texture branches for differential modeling and dynamic weighted fusion. After LightMLP lightweight feature refinement, the spatial resolution is gradually restored by the decoder combined with skip connections, and finally the reconstructed HDR image is output. The frequency information semantic consistency module first divides the output features of the global spatial attention modeling module into k subgroups, and then performs residual correlation with the shallow features of the corresponding exposure frames to obtain linked features. For non-reference frames, the attention generation submodule predicts a reliability attention map to indicate the effective areas that can be used for fusion and the misaligned areas that need to be suppressed. Through the semantic cross-attention module, a query vector Q is generated from the linked features of the reference frame, and a key vector K and a value vector V are generated from the features of the non-reference frame. Cross-frame explicit semantic alignment is achieved through the semantic correlation matrix. Then, the aggregated features are divided into low-frequency branches and high-frequency branches. The low-frequency branches integrate stable illumination and structural information, while the high-frequency branches enhance texture and edge details. Fusion is performed through a dynamic weighting mechanism. The training process introduces a two-level progressive reliability mask refinement strategy and a two-sample HDR contrast loss function. The first-level mask is thresholded based on the global brightness mean, standard deviation, and local contrast. The second-level mask further removes texture degradation areas by extracting perceptual feature distances using a pre-trained VGG-16 network based on the first-level mask. The two-sample HDR contrast loss constructs two positive samples and two negative samples for contrast learning. The positive samples include samples that emphasize the restoration of details in low-exposure images and samples that emphasize the preservation of highlight gradients in high-exposure images. The negative samples include noise artifact variants and structure mismatch variants.

2. The high dynamic range image reconstruction method based on exposure-guided feature decoupling according to claim 1, characterized in that, Includes the following steps: Step 1: Obtain k low dynamic range images of the same scene under different exposure conditions, where k is a positive integer not less than 2; A medium-exposure image is selected as the reference frame; each LDR image is projected into the HDR domain and stitched with the original LDR image in the channel dimension to obtain a 6-channel input representation for each frame; Step 2: Input each frame of 6-channel image into the encoder for feature extraction to obtain shallow and deep features; The shallow features of each frame are fused and initially interacted through a multilayer perceptron to form initial fused features; Step 3: Input the initial fused features into the global spatial attention modeling module, and perform global modeling through a window-based multi-head self-attention mechanism to enhance the ability to model spatial correlation between different exposure frames; Step 4: Input the output features of the global spatial attention modeling module into the frequency information semantic consistency module, divide them into k subgroups and associate them with shallow feature residuals, achieve cross-frame alignment through semantic cross-attention, and perform low-frequency and high-frequency differential modeling and dynamic weighted fusion in the frequency domain. Step 5: Feed the fusion result into the LightMLP module for lightweight feature refinement; Step 6: Input the refined fused features into the decoder, combine them with the skip connections of the shallow features of the encoder, upsample step by step and reconstruct the output 3-channel linear domain HDR image.

3. The high dynamic range image reconstruction method based on exposure-guided feature decoupling according to claim 1, characterized in that, Specific steps for step 1: Step 1-1, let the input LDR image of the i-th frame be... Its corresponding exposure time is Then its HDR domain projection for: , where γ is the gamma coefficient; Steps 1-2: Extract the original LDR image. With corresponding HDR domain projection The data is concatenated along the channel dimension to obtain a 6-channel input representation for each frame. It retains the brightness and color expression in the original LDR space, while introducing the irradiance prior in the HDR domain.

4. The high dynamic range image reconstruction method based on exposure-guided feature decoupling according to claim 1, characterized in that, The specific structure of the encoder described in step 2: The encoder consists of three downsampled convolutional residual blocks and one non-downsampled feature refinement block. The first downsampled block uses a 3×3 convolutional kernel with a stride of 2, has 6 input channels and 64 output channels, and includes a batch normalization layer and a LeakyReLU activation function. The second downsampled block uses a 3×3 convolutional kernel with a stride of 2, expanding the number of channels from 64 to 128. The third downsampled block uses a 3×3 convolutional kernel with a stride of 2, expanding the number of channels from 128 to 256. The feature refinement block uses a 3×3 convolutional kernel with a stride of 1 to further refine the 256-dimensional features. After the above processing, the spatial resolution is downsampled to 1 / 8 of the original image. The encoder also retains shallow features for subsequent skip connections.

5. The high dynamic range image reconstruction method based on exposure-guided feature decoupling according to claim 1, characterized in that, The specific operations of the global spatial attention modeling module described in step 3 are as follows: The input features are first divided into a patch to generate a token sequence, and then passed through layer normalization, multi-head window self-attention, and point-by-point multilayer perceptron; its forward propagation is as follows: , Where LN represents layer normalization, W-MSA represents window multi-head self-attention, and MLP represents multilayer perceptron.

6. The high dynamic range image reconstruction method based on exposure-guided feature decoupling according to claim 1, characterized in that, The specific operations of the frequency information semantic consistency module in step 4 are as follows: Step 4-1. Divide the output features of the global spatial attention modeling module into k subgroups, and perform residual association with the shallow features of the corresponding exposure frames to obtain the linked features of each exposure frame. ; Step 4-2. For non-reference frames, predict a reliability attention map using an attention generation submodule that includes two layers of 3×3 convolutions and Sigmoid activation. The output value ranges from 0 to 1 and is used to indicate the effective region and the misaligned region that needs to be suppressed. Step 4-3. Using the semantic cross-attention module, a query vector Q is generated based on the link features of the reference frame, and a key vector K and a value vector V are generated based on the features of the non-reference frames. The semantic relevance matrix is ​​then calculated to achieve cross-frame semantic alignment. The dimension of the key vector; Step 4-4. Divide the aggregated features into a low-frequency branch and a high-frequency branch; the low-frequency branch extracts low-frequency structural information in the shadow area of ​​the overexposed image; the high-frequency branch extracts high-frequency texture information in the highlight area of ​​the underexposed image; and the outputs of the two branches are fused through a dynamic weighting mechanism.

7. The high dynamic range image reconstruction method based on exposure-guided feature decoupling according to claim 1, characterized in that, The LightMLP module includes 1×1 convolutions for channel blending, 3×3 depthwise separable convolutions for local spatial context modeling, and residual connections, with a parameter complexity of approximately [missing information]. , where C is the channel dimension.

8. The high dynamic range image reconstruction method based on exposure-guided feature decoupling according to claim 1, characterized in that, The specific operation of the two-level progressive reliability mask refinement strategy is as follows: Step A-1. Construct the first-level mask Calculate the global brightness mean from a ground truth HDR image and standard deviation And combine pixel brightness and local contrast for threshold filtering. This is the shadow threshold sensitivity coefficient. Contrast threshold Step A-2. Construct a secondary mask The pre-trained VGG-16 network is used to extract perceptual features between candidate regions and ground truth regions, and the feature distance is calculated. When the distance exceeds the texture threshold The corresponding area will be removed from the mask.

9. A high dynamic range image reconstruction method based on exposure-guided feature decoupling according to claim 1, characterized in that, The specific operation of the training method is as follows: Step B-1. Perform μ-law compression on the ground truth HDR image H and the predicted HDR image: Where μ is 5000, defined in the compression domain. Reconstruction losses; Step B-2. Refining the mask Guided by this method, dual positive and dual negative samples are constructed for HDR contrast learning. Positive samples include those emphasizing detail recovery in low-exposure images and those emphasizing specular gradient preservation in high-exposure images. Negative samples include noise artifact variants formed by injecting Gaussian perturbations into high-risk specular regions and structural mismatch variants formed by irrelevant mask reconstruction. Dual positive NT-Xent type contrast targets are used, where τ is a temperature parameter. Step B-3. The overall training objective is... The comparison branch is only enabled during the training phase and requires no additional computational overhead during the inference phase.

10. An image processing system for implementing the method of any one of claims 1 to 9, characterized in that, The system includes: The feature extraction module is used to receive k LDR input images with different exposure values, perform HDR domain projection and channel stitching on each input image, and extract shallow and deep features through the encoder; The global spatial modeling module is used to perform window-based multi-head self-attention modeling on the fused features of each exposure frame to establish global spatial correlation and long-range dependency across exposure frames. The frequency information semantic consistency module is used to perform feature association, reliability attention map generation, semantic cross-attention alignment, low-frequency and high-frequency branch decomposition, dynamic weight fusion and LightMLP refinement, thereby achieving exposure-guided feature decoupling and complementary fusion. The decoding and reconstruction module is used to combine the fused deep features with shallow skip connection features, upsample step by step, and reconstruct the output HDR image. The training constraint module is used to apply μ-law compression reconstruction loss and two-sample HDR contrast loss to the prediction results during the training phase based on a two-level progressive reliability mask.

Images