A high dynamic imaging method based on deep learning

Abstract

This invention relates to the field of high dynamic range imaging technology and discloses a deep learning-based high dynamic range imaging method. The method includes acquiring original images with different exposure types; extracting features from the original images; using a cross-frame entanglement mechanism to perform quantum motion compensation on abnormal exposure feature maps based on normal exposure feature maps, resulting in compensated abnormal exposure feature maps; fusing features between the normal exposure feature maps and the compensated abnormal exposure feature maps using a multi-scale feature pyramid and adaptive spatial attention, resulting in fused features; and reconstructing the image using an attention-guided residual reconstruction model to obtain a high dynamic range image. This invention solves the motion offset problem between multi-exposure images through quantum motion compensation, eliminating dynamic blur; achieves detail enhancement and color correction through a multi-scale feature pyramid and spatial attention; and reduces computational complexity and improves reconstruction efficiency through a lightweight reconstruction model.

Description

Technical Field

[0001] This invention relates to the field of high dynamic range imaging technology, and in particular to a high dynamic range imaging method based on deep learning. Background Technology

[0002] The dynamic range of an image is a crucial parameter for imaging system quality. Most digital imaging systems use 8-bit quantization, which limits the grayscale and color level range of the output image, resulting in a dynamic range of only two orders of magnitude. This is far below the dynamic range of natural scenes and human visual perception, severely restricting the application of imaging systems. To extend the dynamic range, High Dynamic Range Imaging (HDR) technology is needed. This technology uses hardware and / or software methods to expand the range of the output image and accurately reproduce the details of real-world scenes.

[0003] Currently, high dynamic range (HVR) imaging methods mainly include multiple exposure HVR imaging and compressed sensing imaging. The core idea of ​​multiple exposure HVR is to capture multiple images of the same scene at different exposures and then fuse them appropriately. Through tone mapping, the resulting image should have a larger dynamic range than any image in the sequence of images with different exposures, thus generating a high dynamic range image. Compressed sensing imaging, on the other hand, uses sparse modulation to reduce the dynamic range of the signal under test and multi-pixel detection with superimposed random jitter signals to improve the effective dynamic range of optical measurements, thereby improving the quality of compressed sensing imaging when the detector bit depth is limited.

[0004] While existing methods can reproduce the true state of an image relatively well, they still have certain limitations. On the one hand, multi-exposure high dynamic range imaging requires a certain level of skill from the user, necessitating experience-based judgment based on the shooting scene and adjustment of exposure values ​​using system peripherals. Therefore, this judgment is highly susceptible to human factors, and image fusion errors are also significant, making high dynamic range imaging time-consuming and labor-intensive. On the other hand, achieving high-quality compressed sensing imaging requires high-performance sensors and acquisition equipment to obtain and process large amounts of data. This leads to problems such as system complexity, high cost, and difficulties in storage and transportation for compressed sensing imaging technology. Furthermore, although compressed sensing imaging can recover the original signal from compressed measurements, the quality of the recovery can be affected by various factors, thus limiting the applicability of compressed sensing imaging. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a high dynamic range imaging method based on deep learning, which can solve the problems of blurred dynamic scene reconstruction, reliance on manual parameter adjustment, and high hardware costs in existing high dynamic range imaging technologies, thereby reducing computational load and improving image reconstruction quality.

[0006] In a first aspect, the present invention provides a high dynamic range imaging method based on deep learning, the method comprising: Acquire raw images with different exposure types, including normal exposure and abnormal exposure; Feature extraction is performed on the original image to obtain a feature map group, which includes a normal exposure feature map and an abnormal exposure feature map; Based on the normal exposure feature map, a cross-frame entanglement mechanism is used to perform quantum motion compensation on the abnormal exposure feature map to obtain a compensated abnormal exposure feature map. By using a multi-scale feature pyramid and adaptive spatial attention, feature fusion is performed on the normal exposure feature map and the compensated abnormal exposure feature map to obtain fused features; The fused features are reconstructed using an attention-guided residual reconstruction model to obtain a high dynamic range image.

[0007] Furthermore, the step of extracting features from the original image to obtain a feature map set includes: By using several independent convolutional encoders, parallel feature extraction is performed on the original images with different exposure types to obtain normal exposure feature maps and abnormal exposure feature maps. The abnormal exposure feature maps include underexposure feature maps and overexposure feature maps.

[0008] Furthermore, the convolutional encoder consists of a single-layer 3×3 convolution, with both stride and padding configured to be 1.

[0009] Further, the step of performing quantum motion compensation on the abnormal exposure feature map based on the normal exposure feature map using a cross-frame entanglement mechanism to obtain the compensated abnormal exposure feature map includes: The normal exposure feature map is used as a reference feature, and the abnormal exposure feature map is used as a feature to be compensated. The abnormal exposure feature map includes underexposed feature map and overexposed feature map. The reference feature and the feature to be compensated are enhanced, and the enhanced reference feature and the enhanced feature to be compensated are spliced ​​and quantum entangled to generate the motion field of the enhanced feature to be compensated. The motion field is subjected to offset constraints and grid-based linear sampling to obtain a compensated abnormal exposure feature map, which includes a compensated underexposure feature map and a compensated overexposure feature map.

[0010] Furthermore, the step of performing feature enhancement on the reference feature and the feature to be compensated includes: A 3×3 two-dimensional convolutional layer is used to enhance the reference features and the features to be compensated.

[0011] Furthermore, the step of splicing and quantum entanglement between the enhanced reference features and the enhanced features to be compensated to generate the motion field of the enhanced features to be compensated includes: According to the channel dimension, the enhanced reference feature and the enhanced feature to be compensated are spliced ​​together to obtain the spliced ​​feature; The spliced ​​features are subjected to dimensionality reduction, activation, and convolution to generate the motion field of the enhanced features to be compensated.

[0012] Furthermore, the step of applying offset constraints and grid-based linear sampling to the motion field to obtain the compensated abnormal exposure feature map includes: The offset range of the motion field is limited by action boundary constraints, and the mesh is divided within the limited offset range. Linear sampling is performed based on a grid, and boundary pixels are filled using a preset filling strategy to obtain a compensated abnormal exposure feature map.

[0013] Furthermore, the step of fusing features between the normal exposure feature map and the compensated abnormal exposure feature map using a multi-scale feature pyramid and adaptive spatial attention to obtain fused features includes: The compensated abnormal exposure feature map is enhanced, and then stitched with the normal exposure feature map to obtain interactive features; Multi-scale convolution and global average pooling are performed in parallel on the interactive features, and the output results are batch normalized and concatenated to obtain the initial fused features. The initial fused features are convolved and activated to generate a spatial attention map, which is then normalized and dimensionality reduced to obtain the fused features.

[0014] Furthermore, the step of fusing features between the normal exposure feature map and the compensated abnormal exposure feature map using a multi-scale feature pyramid and adaptive spatial attention to obtain the fused features further includes: The fusion feature between the normal exposure feature map and the compensated underexposure feature map is taken as the first fusion feature, and the fusion feature between the normal exposure feature map and the compensated overexposure feature map is taken as the second fusion feature; The first fusion feature and the second fusion feature are concatenated and dimensionality reduced to obtain the third fusion feature.

[0015] Furthermore, the residual reconstruction model includes multiple cascaded residual blocks, convolutional layers, and normalization layers; The residual block comprises a series of convolutional sub-modules and a channel attention sub-module; The convolutional submodule includes a first sub-convolutional layer, a first activation layer, a second sub-convolutional layer, and a global pooling layer; the channel attention submodule includes a third sub-convolutional layer, a second activation layer, a fourth sub-convolutional layer, and a third activation layer.

[0016] This invention provides a high dynamic range imaging method based on deep learning. Employing an end-to-end deep learning architecture, in the feature extraction stage, a three-way parallel convolutional encoder extracts features from raw images with different exposure types, providing a unified feature representation foundation for subsequent processing. In the motion compensation stage, a cross-frame entanglement mechanism performs quantum motion compensation on abnormally exposed raw images, effectively solving the motion offset problem between multi-exposure images, eliminating dynamic blur, and achieving precise alignment of feature displacements in dynamic scenes. In the feature fusion stage, a multi-scale feature pyramid and adaptive spatial attention enable adaptive feature fusion, detail enhancement, and color correction, effectively expanding the dynamic range. In the image reconstruction stage, a lightweight reconstruction model enhances details, achieving high-quality HDR image reconstruction. This invention achieves high-quality image reconstruction with a single forward computation, effectively reducing computational complexity and thus improving reconstruction efficiency. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the high dynamic range imaging method based on deep learning in an embodiment of the present invention; Figure 2 This is a schematic diagram of the architecture of the quantum motion compensation module in an embodiment of the present invention; Figure 3 This is a schematic diagram of the feature fusion module architecture in an embodiment of the present invention; Figure 4 This is a schematic diagram of the architecture of the residual block in the residual reconstruction model in this embodiment of the invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0019] Please see Figure 1 The first embodiment of the present invention proposes a high dynamic range imaging method based on deep learning, including steps S10 to S50: Step S10: Obtain original images with different exposure types, including normal exposure and abnormal exposure; Step S20: Extract features from the original image to obtain a feature map group, which includes a normal exposure feature map and an abnormal exposure feature map. Step S30: Based on the normal exposure feature map, a cross-frame entanglement mechanism is used to perform quantum motion compensation on the abnormal exposure feature map to obtain a compensated abnormal exposure feature map. Step S40: By using a multi-scale feature pyramid and adaptive spatial attention, feature fusion is performed on the normal exposure feature map and the compensated abnormal exposure feature map to obtain fused features; Step S50: The fused features are reconstructed using an attention-guided residual reconstruction model to obtain a high dynamic range image.

[0020] This invention provides an end-to-end imaging method for high dynamic range imaging based on quantum motion compensation and multi-scale interaction. The invention employs a three-frame collaborative processing mechanism, using images of different exposure types acquired in a single shot as input. It achieves adaptive feature fusion through feature extraction, motion compensation, and feature interaction, and adds details through a lightweight reconstruction model, thereby achieving high dynamic range imaging with low computational cost and high quality.

[0021] First, feature extraction is performed on the original images to be reconstructed. In this embodiment, the original images include normally exposed images and abnormally exposed images. Abnormally exposed images include underexposed and overexposed images. For these three types of images, this embodiment uses three independent convolutional encoders to perform parallel feature extraction on normally exposed, underexposed, and overexposed images, thereby generating feature maps in a unified representation space, including normally exposed feature maps and abnormally exposed feature maps. The abnormally exposed feature maps include underexposed and overexposed feature maps. Preferably, each convolutional encoder consists of a single-layer 3×3 convolution with a stride of 1 and padding of 1. Assuming each image has a shape of (B, 3, H, W), where B represents the batch size, 3 represents the number of channels, H represents the height, and W represents the width, the convolutional encoder provided in this embodiment expands the input channels from 3 to 64 channels while maintaining the spatial resolution of the input image. This process captures basic texture features through local receptive fields and establishes illumination-invariant representations, ultimately outputting three sets of 64-channel feature maps: underexposed feature map f1, normally exposed feature map f2, and overexposed feature map f3, thereby providing a unified feature representation basis for subsequent processing.

[0022] For the extracted multiple feature maps, using the normal exposure feature map f2 as the reference feature, this embodiment employs a cross-frame entanglement mechanism to perform quantum motion compensation on the abnormal exposure feature maps (including f1 and f3) through the reference feature. The specific steps include: The normal exposure feature map is used as a reference feature, and the abnormal exposure feature map is used as a feature to be compensated. The abnormal exposure feature map includes underexposed feature map and overexposed feature map. The reference feature and the feature to be compensated are enhanced, and the enhanced reference feature and the enhanced feature to be compensated are spliced ​​and quantum entangled to generate the motion field of the enhanced feature to be compensated. The motion field is subjected to offset constraints and grid-based linear sampling to obtain a compensated abnormal exposure feature map, which includes a compensated underexposure feature map and a compensated overexposure feature map.

[0023] In this embodiment, motion compensation for underexposed and overexposed feature maps is performed in parallel using two independent quantum motion compensation modules. Specifically, the normally exposed feature map f2 is used as a reference feature, and both the underexposed feature map f1 and the overexposed feature map f3 are used as features to be compensated. Then, f1 and f2 are input into one quantum motion compensation module, and f2 and f3 are input into another quantum motion compensation module. Motion compensation for the features to be compensated is achieved through the parallel processing of the two independent quantum motion compensation modules. Preferably, the architecture of the two quantum motion compensation modules is consistent. The following is based on... Figure 2 The schematic diagram of the quantum motion compensation module shown below illustrates the motion compensation process in detail, taking the motion compensation of reference feature f2 and feature f1 to be compensated as an example.

[0024] In quantum motion compensation, f1 and f2 are first convolved to enhance motion-sensitive features. A 3×3 two-dimensional convolutional layer (Conv2d 3×3) is used for this convolution. Then, the enhanced f1 and f2 are concatenated along the channel dimension to obtain concatenated features. These concatenated features are then quantum-entangled to generate the motion field of the enhanced f1. The quantum entanglement steps include: According to the channel dimension, the enhanced reference feature and the enhanced feature to be compensated are spliced ​​together to obtain the spliced ​​feature; The spliced ​​features are subjected to dimensionality reduction, activation, and convolution to generate the motion field of the enhanced features to be compensated.

[0025] In this embodiment, 128→64 channel dimensionality reduction compression is first achieved through 1×1 two-dimensional convolution (Conv2d 1×1), and the compressed features are activated using an activation function (LeakyReLU). Finally, a 2-channel motion field is generated using 5×5 two-dimensional large-kernel convolution (Conv2d 5×5). This can be understood as generating the motion field f1 through quantum entanglement of f1 and f2, which is a motion field compensated for by f2. For the generated motion field, the compensated f1 can be obtained through offset constraints and linear sampling. The specific steps include: The offset range of the motion field is limited by action boundary constraints, and the mesh is divided within the limited offset range. Linear sampling is performed based on a grid, and boundary pixels are filled using a preset filling strategy to obtain a compensated abnormal exposure feature map.

[0026] In this embodiment, the motion field is subjected to action boundary constraints, mapping it within a preset pixel range. Preferably, the motion field is constrained using the function tanh, and the offset range is limited to [-2, 2] pixels by multiplying tanh by 2. The motion field after boundary constraints is then divided into grids. Based on the grids, sub-pixel-level precise deformation is achieved through dynamic sampling. For the dynamically sampled feature map, the boundary pixels are processed using a predicted filling strategy to obtain an underexposed feature map f1 that is perfectly aligned with the reference feature f2, i.e., the compensated underexposed feature map f1. Preferably, the dynamic sampling uses a differentiable bilinear sampling technique, and the filling strategy can use a zero-filling strategy or a reflection-filling strategy to process edge information to maintain image size consistency. It should be noted that the bilinear sampling technique and filling strategy used in this embodiment can refer to the steps of conventional sampling techniques and filling strategies, and will not be repeated here.

[0027] In this embodiment, through a parallel quantum motion compensation module, underexposed feature map f1 and overexposed feature map f3 that are perfectly aligned with reference feature f2 can be obtained. By motion compensation of the underexposed and overexposed feature maps, motion offset artifacts between multi-exposure sequences can be eliminated, thus laying a geometric consistency foundation for subsequent feature fusion.

[0028] After obtaining the normal exposure feature map f2, the compensated underexposure feature map f1, and the compensated overexposure feature map f3, feature fusion based on feature stitching can be performed on these three feature maps to obtain the fused feature.

[0029] In a preferred embodiment, to effectively expand the dynamic range of the fused features and improve feature performance, this embodiment employs a feature fusion module based on a multi-scale feature pyramid and adaptive spatial attention. This module fuses the normal exposure feature map and the compensated abnormal exposure feature map. The feature fusion module has a dual-branch structure: one branch fuses the normal exposure feature map and the compensated underexposure feature map, while the other branch fuses the normal exposure feature map and the compensated overexposure feature map. The fused features output by the feature fusion module, i.e., the fused features from both branches, are then concatenated and dimensionality reduced to obtain the final fused features. The specific steps for each branch to fuse the normal exposure feature map and the compensated abnormal exposure feature map include: The compensated abnormal exposure feature map is enhanced, and then stitched with the normal exposure feature map to obtain interactive features; Multi-scale convolution and global average pooling are performed in parallel on the interactive features, and the output results are batch normalized and concatenated to obtain the initial fused features. The initial fused features are convolved and activated to generate a spatial attention map, which is then normalized and dimensionality reduced to obtain the fused features.

[0030] In this embodiment, the fusion steps of the two feature maps are explained using the normally exposed feature map f2 and the compensated underexposed feature map f1 as examples. Please refer to [link to relevant documentation]. Figure 3 The schematic diagram of the feature fusion module shown first employs convolutional layers and activation functions to enhance the features of the compensated underexposed feature map f1. Preferably, a two-layer convolution and activation is used for feature enhancement, with each layer employing a 3×3 two-dimensional convolutional layer (Conv2d 3×3) and the LeakyReLU activation function. The enhanced and compensated underexposed feature map f1 is then concatenated with the normally exposed feature map f2 to obtain the interactive features.

[0031] The interactive features are then input into a multi-scale perceptual pyramid for multi-scale feature extraction. This pyramid consists of multiple convolutional layers and global average pooling layers of different scales arranged side-by-side. Preferably, a parallel branching structure of 3×3, 5×5, and 7×7 two-dimensional convolutions (Conv2d) and two-dimensional global average pooling (AdaptiveAvgpool2d) is used. Batch normalization (BN) is applied to the outputs of the convolutional branches, while dimensionality reduction is performed on the outputs of the pooling branches using 1×1 two-dimensional convolutions (Conv2d 1×1). The outputs of the batch-normalized branches and the dimensionality-reduced branches are then concatenated to obtain an initial 128-channel fused feature.

[0032] For the initial fusion features, a 3×3 two-dimensional convolution, batch normalization, and activation using the Sigmoid function (Conv2d+BN+Sigmoid) were used to generate a spatial attention map. Finally, the weights were normalized by the Softmax function to achieve adaptive weighted fusion, and a 64-channel fusion feature was output by dimensionality reduction through a 1×1 two-dimensional convolution (Conv2d 1×1).

[0033] The above steps can be used to fuse the normal exposure feature map and the compensated abnormal exposure feature map to obtain the fused feature. The fused feature generated by fusing the normal exposure feature map and the compensated underexposure feature map optimizes the shadow performance, while the fused feature generated by fusing the normal exposure feature map and the compensated overexposure feature map optimizes the highlight performance, thereby achieving an effective expansion of the dynamic range.

[0034] The two fused features obtained from parallel feature fusion processing need to be further fused. Specifically, the fused feature between the normal exposure feature map and the compensated underexposure feature map is taken as the first fused feature, and the fused feature between the normal exposure feature map and the compensated overexposure feature map is taken as the second fused feature. The first and second fused features are concatenated along the channel dimension to obtain a 128-channel composite feature. Then, a linear transformation is performed through a 1×1 convolution kernel to reduce the dimensionality, and the output is a compact aggregated 64-channel third fused feature, which is the final fused feature. This fused feature integrates visual information across the entire dynamic range, providing a high information density input for image reconstruction.

[0035] In the image reconstruction stage, this embodiment employs an attention-guided residual reconstruction model to reconstruct the image from the input fused features. The residual reconstruction model consists of three cascaded residual blocks, which form an effective receptive field through stacking. It models the mapping relationship between illumination and color in depth. Each residual block consists of cascaded convolutional submodules and channel attention submodules, and its structure is as follows: Figure 4As shown, in this embodiment, the convolutional submodule includes a first sub-convolutional layer, a first activation layer, a second sub-convolutional layer, and a global pooling layer. The channel attention submodule includes a third sub-convolutional layer, a second activation layer, a fourth sub-convolutional layer, and a third activation layer. Specifically, the convolutional submodule contains a convolutional path of 3×3 convolution (3×3 Conv), ReLU activation (ReLU), and 3×3 convolution (3×3 Conv). After convolution processing of the input fused features through this convolutional path, global pooling is performed through the Global Avgpool layer. For the features after global pooling, convolutional dimensionality reduction and channel weighting are performed through the channel attention submodule. In the channel attention submodule, after feature dimensionality reduction through 1×1 convolution (1×1 Conv), ReLU activation (ReLU), and dimensionality increase through 1×1 convolution (1×1 Conv), channel weights are generated through the activation function (Sigmoid). The residual connection between the channel attention weighted output and the original input can prevent gradient vanishing.

[0036] Based on the above residual block architecture, the fusion features are processed by three cascaded residual blocks to form an effective receptive field, realize depth modeling of illumination and color mapping, and finally map the 64 channels to the RGB space through 3×3 convolution, and normalize the pixel values ​​to the range of [0,1] by sigmoid activation, thereby outputting the reconstructed high dynamic range (HDR) image.

[0037] In summary, the embodiments of this invention propose a high dynamic range imaging method based on deep learning. The method acquires original images with different exposure types, including normal and abnormal exposure. Features are extracted from the original images to obtain a feature map set, which includes normal and abnormal exposure feature maps. Based on the normal exposure feature map, a cross-frame entanglement mechanism is used to perform quantum motion compensation on the abnormal exposure feature map to obtain a compensated abnormal exposure feature map. Features are fused between the normal exposure feature map and the compensated abnormal exposure feature map using a multi-scale feature pyramid and adaptive spatial attention to obtain fused features. Finally, an attention-guided residual reconstruction model is used to reconstruct the image using the fused features to obtain a high dynamic range image. In the feature extraction stage, this invention uses a three-way parallel convolutional encoder to extract features from original images with different exposure types, providing a unified feature representation foundation for subsequent processing. In the motion compensation stage, a cross-frame entanglement mechanism is used to perform quantum motion compensation on abnormally exposed original images, effectively solving the motion offset problem between multi-exposure images, eliminating dynamic blur, and achieving precise alignment of feature displacements in dynamic scenes. In the feature fusion stage, a multi-scale feature pyramid and adaptive spatial attention are used to achieve adaptive feature fusion, detail enhancement, and color correction, effectively expanding the dynamic range. In the image reconstruction stage, a lightweight reconstruction model is used to enhance details, enabling high-quality HDR image reconstruction. This invention achieves high-quality image reconstruction with a single forward computation, effectively reducing computational complexity and thus improving reconstruction efficiency.

[0038] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the various embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as these combinations of technical features do not contradict each other, they should be considered within the scope of this specification.

[0039] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the technical principles of this invention, and these improvements and substitutions should also be considered within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the scope of the claims.

Claims

1. A high dynamic range imaging method based on deep learning, characterized in that, include: Acquire raw images with different exposure types, including normal exposure and abnormal exposure; Feature extraction is performed on the original image to obtain a feature map group, which includes a normal exposure feature map and an abnormal exposure feature map; Based on the normal exposure feature map, a cross-frame entanglement mechanism is used to perform quantum motion compensation on the abnormal exposure feature map to obtain a compensated abnormal exposure feature map. By using a multi-scale feature pyramid and adaptive spatial attention, feature fusion is performed on the normal exposure feature map and the compensated abnormal exposure feature map to obtain fused features; The fused features are reconstructed using an attention-guided residual reconstruction model to obtain a high dynamic range image.

2. The high dynamic range imaging method based on deep learning according to claim 1, characterized in that, The step of extracting features from the original image to obtain a feature map set includes: By using several independent convolutional encoders, parallel feature extraction is performed on the original images with different exposure types to obtain normal exposure feature maps and abnormal exposure feature maps. The abnormal exposure feature maps include underexposure feature maps and overexposure feature maps.

3. The high dynamic range imaging method based on deep learning according to claim 2, characterized in that, The convolutional encoder consists of a single-layer 3×3 convolution, with both stride and padding configured to 1.

4. The high dynamic range imaging method based on deep learning according to claim 1, characterized in that, The step of performing quantum motion compensation on the abnormal exposure feature map based on the normal exposure feature map using a cross-frame entanglement mechanism to obtain the compensated abnormal exposure feature map includes: The normal exposure feature map is used as a reference feature, and the abnormal exposure feature map is used as a feature to be compensated. The abnormal exposure feature map includes underexposed feature map and overexposed feature map. The reference feature and the feature to be compensated are enhanced, and the enhanced reference feature and the enhanced feature to be compensated are spliced ​​and quantum entangled to generate the motion field of the enhanced feature to be compensated. The motion field is subjected to offset constraints and grid-based linear sampling to obtain a compensated abnormal exposure feature map, which includes a compensated underexposure feature map and a compensated overexposure feature map.

5. The high dynamic range imaging method based on deep learning according to claim 4, characterized in that, The step of performing feature enhancement on the reference feature and the feature to be compensated includes: A 3×3 two-dimensional convolutional layer is used to enhance the reference features and the features to be compensated.

6. The high dynamic range imaging method based on deep learning according to claim 4, characterized in that, The step of splicing and quantum entanglement between the enhanced reference features and the enhanced features to be compensated to generate the motion field of the enhanced features to be compensated includes: According to the channel dimension, the enhanced reference feature and the enhanced feature to be compensated are spliced ​​together to obtain the spliced ​​feature; The spliced ​​features are subjected to dimensionality reduction, activation, and convolution to generate an enhanced motion field for the features to be compensated.

7. The high dynamic range imaging method based on deep learning according to claim 4, characterized in that, The step of applying offset constraints and grid-based linear sampling to the motion field to obtain the compensated abnormal exposure feature map includes: The offset range of the motion field is limited by action boundary constraints, and the mesh is divided within the limited offset range. Linear sampling is performed based on a grid, and boundary pixels are filled using a preset filling strategy to obtain a compensated abnormal exposure feature map.

8. The high dynamic range imaging method based on deep learning according to claim 4, characterized in that, The step of fusing the normal exposure feature map and the compensated abnormal exposure feature map using a multi-scale feature pyramid and adaptive spatial attention to obtain the fused features includes: The compensated abnormal exposure feature map is enhanced, and then stitched with the normal exposure feature map to obtain interactive features; Multi-scale convolution and global average pooling are performed in parallel on the interactive features, and the output results are batch normalized and concatenated to obtain the initial fused features. The initial fused features are convolved and activated to generate a spatial attention map, which is then normalized and dimensionality reduced to obtain the fused features.

9. The high dynamic range imaging method based on deep learning according to claim 8, characterized in that, The step of fusing the normal exposure feature map and the compensated abnormal exposure feature map using a multi-scale feature pyramid and adaptive spatial attention to obtain the fused features further includes: The fusion feature between the normal exposure feature map and the compensated underexposure feature map is taken as the first fusion feature, and the fusion feature between the normal exposure feature map and the compensated overexposure feature map is taken as the second fusion feature; The first fusion feature and the second fusion feature are concatenated and dimensionality reduced to obtain the third fusion feature.

10. The high dynamic range imaging method based on deep learning according to claim 1, characterized in that, The residual reconstruction model includes multiple cascaded residual blocks, convolutional layers, and normalization layers; The residual block comprises a series of convolutional sub-modules and a channel attention sub-module; The convolutional submodule includes a first sub-convolutional layer, a first activation layer, a second sub-convolutional layer, and a global pooling layer; the channel attention submodule includes a third sub-convolutional layer, a second activation layer, a fourth sub-convolutional layer, and a third activation layer.

Images