A domain adaptation image deblurring algorithm
By employing a domain-adaptive image deblurring algorithm, a kernel estimation and deblurring model is constructed using the U-Net network and a multi-stage progressive recovery network. This model is then fine-tuned using a target domain dataset, addressing the performance differences between the source and target domains and improving the accuracy and adaptability of image deblurring.
Patent Information
- Application Number
- CN202610351377.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-21
- Publication Date
- 2026-06-19
AI Technical Summary
Existing deep learning-based motion blur removal methods exhibit significant performance differences between the source and target domains, resulting in insufficient generalization ability in real-world blurred image testing. These methods struggle to encompass all possible complex motion patterns in the real world.
A domain-adaptive image deblurring algorithm is adopted. A kernel estimation model is built based on the U-Net network. The model is pre-trained using the source domain dataset and combined with the target domain blurred image dataset to construct a support set that is distributed in the same direction as the target domain. The deblurring model is then fine-tuned to improve its performance in the target domain.
It effectively reduces the performance difference between the source and target domains, improves the performance of the deblurring algorithm in the target domain, and enhances the accuracy and adaptability of image deblurring.
Smart Images
Figure CN122243808A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing, and more specifically to a domain-adaptive image deblurring algorithm, which is particularly suitable for processing blurred images caused by object motion. Background Technology
[0002] Image deblurring is an important research direction in computer vision and image processing. Its goal is to reconstruct clear, high-quality images from blurred images caused by factors such as camera shake, object motion, or defocus. Motion blur is particularly common in dynamic shooting scenarios, such as intelligent perception in unmanned systems, and it adversely affects subsequent visual tasks (such as object detection and recognition).
[0003] Supervised motion blur removal methods based on deep learning have become mainstream. Since acquiring realistic "blurred-sharp" image pairs is extremely difficult, these methods typically rely on convolving sharp images with pre-defined, relatively ideal motion blur kernels to generate "blurred-sharp" image pairs for supervised training. However, artificially generated blur kernels cannot accurately reflect real blur information, and the motion blur priors learned by the model on the source domain dataset differ significantly from the complex and diverse motion blur patterns in the real world (target domain). This mismatch between blur patterns in the source and target domains results in high-performance deblurring models trained on the source domain dataset exhibiting severely insufficient generalization ability when tested on real blurred images.
[0004] To address the aforementioned issues, existing technologies primarily improve the performance of deblurring models by increasing the diversity of datasets. However, this method has inherent limitations: the expanded datasets cannot exhaustively represent all possible complex motion patterns in the real world. Therefore, developing a deblurring algorithm that can adapt to different real-world scenarios has become a key challenge in improving the practicality of deblurring technology. Summary of the Invention
[0005] In order to overcome the shortcomings of the prior art, the present invention aims to provide a domain-adaptive image deblurring algorithm to reduce the performance difference between the source domain and the target domain and improve the performance of the deblurring algorithm in the target domain.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] A domain-adaptive image deblurring algorithm includes the following steps;
[0008] Step 1: Construct a kernel estimation model based on the U-Net network ;
[0009] Step 2: Construct a deblurring model ;
[0010] Step 3: Utilize the source domain dataset Pre-trained kernel estimation model and deblurring model ;
[0011] Step 4: Utilize the target domain blurred image dataset Kernel estimation model Construct a support set that is identically distributed to the target domain. ;
[0012] Step 5: Utilize support sets Fine-tuning the deblurring model And obtain the updated deblurred model. ;
[0013] Step 6: Utilize the updated deblurring model described in Step 5 Acquire target domain blurred image Corresponding clear image ;
[0014] Furthermore, in step 1, a kernel estimation model is constructed based on the U-Net network. The method is as follows:
[0015] Step 1.1: Extract blurred image features using the encoder, decoder, and skip connection structure of the U-Net network. ;
[0016] Step 1.2, based on the features of the blurred image described in Step 1.1 Construct a parameterized branch for motion information in a blurred image, including: length ,angle and curvature Parameter estimation branch; where, length and angle The parameter estimation branch consists of two convolutional layers with a kernel size of 3 and a stride of 1, and a sigmoid activation function; curvature The parameter estimation branch consists of two convolutional layers with a kernel size of 3 and a stride of 1, and a Tanh activation function; and its length... ,angle and curvature The parameters of the internal network in the parameter estimation branch are not shared, as shown in the following formula;
[0017]
[0018]
[0019]
[0020] Step 1.3, using the length described in step 1.2 ,angle and curvature Parameters construct fuzzy kernel The details are as follows;
[0021]
[0022] in, This is an indicator function, which is used in the fuzzy kernel. The center position is 1, and the other positions are 0; for The fuzzy kernel parameterization function is as follows:
[0023]
[0024] in, Indicates the size of the fuzzy kernel. Used to map motion information to trajectory points The quantity, based on length ,angle and curvature Mapping fuzzy motion information to The formula is shown below:
[0025]
[0026] in, Indicates in A discrete sequence sampled at linear intervals within an interval is used to approximate the trajectory of an object in the horizontal direction, and combined with... , and Perform trajectory point mapping;
[0027] Furthermore, in step 2, a deblurring model is constructed. The method is as follows:
[0028] Step 2.1: Construct a deblurring model using the Multi-Stage Progressive Recovery Network (MPRNet). The basic infrastructure; The model refines the defuzzified results step by step through a multi-stage feature extraction and fusion mechanism.
[0029] Furthermore, in step 3, the source domain dataset is utilized. Pre-trained kernel estimation model and deblurring model The method is as follows:
[0030] Step 3.1, Construct the kernel estimation model loss function The details are as follows;
[0031]
[0032] in, Represents the source domain dataset Clear images in Represents the source domain dataset The blurred image in Indicates the convolution operation;
[0033] Step 3.2, according to the loss function described in step 3.1 Training kernel estimation model The details are as follows:
[0034]
[0035] in, This represents the learning rate of the kernel estimation model. This represents the number of images processed per batch by the kernel estimation model;
[0036] Step 3.3, Constructing a Deblurring Model loss function The details are as follows:
[0037]
[0038]
[0039]
[0040]
[0041] in, For weight parameters, Indicates the Laplacian operator. This represents the VGG-Net feature extractor;
[0042] Step 3.4, according to the loss function described in step 3.3 Training the deblurring model The details are as follows:
[0043]
[0044] in, Deblurring model The learning rate Deblurring model Number of images processed per batch;
[0045] Furthermore, in step 4, the target domain blurred image dataset is utilized. Kernel estimation model Construct a support set that is identically distributed to the target domain. The method is as follows:
[0046] Step 4.1: Construct a sharpness detection module; use the sharpness detection module to acquire the blurred image of the target domain. Chinese pseudo-clear image patches and blurred image patches First, the original image Perform geometric transformations to the left, right, up, and down, as follows:
[0047]
[0048]
[0049]
[0050]
[0051] in, Original image The image after the leftward translation operation, Original image The image after being shifted to the right. Original image The image after the upward translation operation, Original image The image after being shifted downwards; then, the original image... and images after four geometric transformations , , and The peak signal-to-noise ratio (PSNR) is calculated for each sliding window, as shown below:
[0052]
[0053] in, Represents the original image The image after geometric transformation The mean squared error function is used; finally, the images after the four geometric transformations in each sliding window are compared with the original image. The sum of the peak signal-to-noise ratios (PSNR) and the original image corresponding to the sliding window with the minimum sum of PSNR are calculated. This is denoted as a pseudo-clear image block. The original image corresponding to the sliding window with the largest sum of peak signal-to-noise ratio. Denote as blurred image patch ;
[0054] Step 4.2: Utilize the pseudo-sharp image patch dataset Blurred image patch dataset and target domain blurred image dataset Fine-tuning the kernel estimation model This prompts the kernel estimation model The motion blur information that is distributed in the same way as the target domain is extracted, as shown below:
[0055]
[0056] in, Indicates the learning rate. The loss represents the data being identically distributed, as follows:
[0057]
[0058] in, This represents the loss of the Fast Fourier Transform. The discrete wavelet transform loss is represented as follows:
[0059]
[0060]
[0061] Step 4.3: Utilize the fine-tuned kernel estimation model Acquiring and pseudo-sharp image patches Corresponding to and with the blurred image in the target domain Identically distributed pseudo-blurred image patches And utilize pseudo-clear image patches and pseudo-blurred image patches Building support sets Among them, pseudo-blurred image blocks The calculation is as follows:
[0062]
[0063] Furthermore, step 5 utilizes support sets. Fine-tuning the deblurring model And obtain the updated deblurred model. The method is as follows:
[0064] Using support sets By combining steps 3.3 and 3.4 above, fine-tuning the deblurring model ;
[0065] Step 6: Utilize the updated deblurring model described in Step 5 Acquire target domain blurred image Corresponding clear image The method is as follows:
[0066] Attached Figure Description
[0067] Figure 1 Flowchart of the domain-adaptive image deblurring algorithm.
[0068] Figure 2 The structure diagram of the kernel estimation network model is shown.
[0069] Figure 3 This is a diagram of the deblurred network model structure. Detailed Implementation
[0070] The detailed description and technology of the invention are illustrated below in conjunction with the accompanying drawings. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the protection scope of the present invention.
[0071] The flowchart of this invention is as follows Figure 1 As shown, the specific steps include:
[0072] Step 1: Construct a kernel estimation model based on the U-Net network Kernel estimation model Network structure such as Figure 2 As shown, the specific method is as follows:
[0073] Step 1.1: Extract blurred image features using the encoder, decoder, and skip connection structure of the U-Net network. ;
[0074] Step 1.2, based on the blurred image features described in Step 1.1 Construct a parameterized branch for motion information in a blurred image, including: length ,angle and curvature Parameter estimation branch; where, length and angle The parameter estimation branch consists of two convolutional layers with a kernel size of 3 and a stride of 1, and a sigmoid activation function; curvature The parameter estimation branch consists of two convolutional layers with a kernel size of 3 and a stride of 1, and a Tanh activation function; and its length... ,angle and curvature The parameters within the parameter estimation branch are not shared. The specific formula is shown below:
[0075]
[0076]
[0077]
[0078] Step 1.3, using the length described in step 1.2 ,angle and curvature Parameters construct fuzzy kernel The details are as follows:
[0079]
[0080] in, This is an indicator function, which is used in the fuzzy kernel. The center position is 1, and the other positions are 0; for The fuzzy kernel parameterization function is as follows:
[0081]
[0082] in, Indicates the size of the fuzzy kernel. Used to map motion information to trajectory points The quantity.
[0083] Furthermore, based on length ,angle and curvature Mapping fuzzy motion information to The formula is shown below:
[0084]
[0085] in, Indicates in A discrete sequence sampled at linear intervals within an interval is used to approximate the trajectory of an object in the horizontal direction, and combined with... , and Perform trajectory point mapping;
[0086] Step 2: Construct a deblurring model Defuzzing model Network structure such as Figure 3 As shown, the specific method is as follows:
[0087] Constructing a deblurring model using the multi-stage progressive recovery network MPRNet Network architecture; The model refines the defuzzified results step by step through a multi-stage feature extraction and fusion mechanism.
[0088] Step 3: Utilize the source domain dataset Pre-trained kernel estimation model and deblurring model The specific method is as follows:
[0089] Step 3.1, Construct the kernel estimation model loss function The details are as follows;
[0090]
[0091] in, Represents the source domain dataset Clear images in Represents the source domain dataset The blurred image in Indicates the convolution operation;
[0092] Step 3.2, according to the loss function described in step 3.1 Training kernel estimation model The details are as follows:
[0093]
[0094] in, This represents the learning rate of the kernel estimation model. This represents the number of images processed per batch by the kernel estimation model;
[0095] Step 3.3, Constructing a Deblurring Model loss function The details are as follows:
[0096]
[0097]
[0098]
[0099]
[0100] in, For weight parameters, Indicates the Laplacian operator. This represents the VGG-Net feature extractor;
[0101] Step 3.4, according to the loss function described in step 3.3 Training the deblurring model The details are as follows:
[0102]
[0103] in, Deblurring model The learning rate Deblurring model Number of images processed per batch;
[0104] Step 4: Utilize the target domain blurred image dataset Kernel estimation model Construct a support set that is identically distributed to the target domain. The specific method is as follows:
[0105] Step 4.1: Construct a sharpness detection module; use the sharpness detection module to acquire the blurred image of the target domain. Chinese pseudo-clear image patches and blurred image patches ;
[0106] Furthermore, the original image Perform geometric transformations to the left, right, up, and down, as follows:
[0107]
[0108]
[0109]
[0110]
[0111] in, Original image The image after the leftward translation operation, Original image The image after being shifted to the right. Original image The image after the upward translation operation, Original image The image after being shifted downwards; then, the original image... and images after four geometric transformations , , and The peak signal-to-noise ratio (PSNR) is calculated for each sliding window, as shown below:
[0112]
[0113] in, Represents the original image The image after geometric transformation to the left, right, up, or down. The mean squared error function is used; finally, the images after the four geometric transformations in each sliding window are compared with the original image. The sum of the peak signal-to-noise ratios (PSNR) and the original image corresponding to the sliding window with the minimum sum of PSNR. This is denoted as a pseudo-clear image block. The original image corresponding to the sliding window with the largest sum of peak signal-to-noise ratio. Denote as blurred image patch ;
[0114] Step 4.2: Utilize the pseudo-sharp image patch dataset Blurred image patch dataset and target domain dataset Fine-tuning the kernel estimation model This enables the kernel estimation model. The motion blur information that is distributed in the same way as the target domain is extracted, as shown below:
[0115]
[0116] in, Indicates the learning rate. The loss represents the data being identically distributed, as shown below:
[0117]
[0118] in, This represents the loss of the Fast Fourier Transform. The discrete wavelet transform loss is represented as follows:
[0119]
[0120]
[0121] Step 4.3: Utilize the fine-tuned kernel estimation model Acquiring and pseudo-sharp image patches Corresponding to and with the blurred image in the target domain Identically distributed pseudo-blurred image patches Among them, pseudo-blurred image blocks The calculation is as follows:
[0122]
[0123] Furthermore, utilizing pseudo-sharp image patches and pseudo-blurred image patches Building support sets ;
[0124] Step 5: Utilize support sets Fine-tuning the deblurring model And obtain the updated deblurred model. ;
[0125] Furthermore, utilizing support sets By combining steps 3.3 and 3.4 above, fine-tuning the deblurring model ;
[0126] Step 6: Utilize the updated deblurring model described in Step 5 Acquire target domain blurred image Corresponding clear image The specific method is as follows:
[0127]
Claims
1. A domain-adaptive image deblurring algorithm, characterized in that, Includes the following steps: Step 1: Construct a kernel estimation model based on the U-Net network ; Step 2: Constructing a deblurring model ; Step 3: Utilize the source domain dataset Pre-trained kernel estimation model and deblurring model ; Step 4: Utilize the target domain blurred image dataset Kernel estimation model Construct a support set that is identically distributed to the target domain. ; Step 5: Utilize support sets Fine-tuning the deblurring model And obtain the updated deblurred model. ; Step Six: Utilize the updated deblurring model described in Step Five. Acquire target domain blurred image Corresponding clear image .
2. The domain-adaptive image deblurring algorithm according to claim 1, characterized in that, In step one, a kernel estimation model is constructed based on the U-Net network. The method is as follows: Step 1.1: Extract blurred image features using the encoder, decoder, and skip connection structure of the U-Net network. ; Step 1.2, based on the features of the blurred image described in Step 1.1 Construct a parameterized branch for motion information in a blurred image, including: length ,angle and curvature Parameter estimation branch; where, length and angle The parameter estimation branch consists of two convolutional layers with a kernel size of 3 and a stride of 1, and a sigmoid activation function; curvature The parameter estimation branch consists of two convolutional layers with a kernel size of 3 and a stride of 1, and a Tanh activation function; and its length... ,angle and curvature The parameters within the parameter estimation branch are not shared. The specific formula is shown below: Step 1.3, using the length described in step 1.2 ,angle and curvature Parameters construct fuzzy kernel The details are as follows: in, This is an indicator function, which is used in the fuzzy kernel. The center position is 1, and the other positions are 0; for The fuzzy kernel parameterization function is as follows: in, Indicates the size of the fuzzy kernel. Used to map motion information to trajectory points The quantity. Furthermore, based on length... ,angle and curvature Mapping fuzzy motion information to The formula is shown below: in, Indicates in A discrete sequence sampled at linear intervals within an interval is used to approximate the trajectory of an object in the horizontal direction, and combined with... , and Perform trajectory point mapping.
3. The domain-adaptive image deblurring algorithm according to claim 1, characterized in that, In step two, a deblurring model is constructed. The method is as follows: Constructing a deblurring model using the multi-stage progressive recovery network MPRNet Network architecture; The model refines the defuzzified results step by step through a multi-stage feature extraction and fusion mechanism.
4. The domain-adaptive image deblurring algorithm according to claim 1, characterized in that, Step three utilizes the source domain dataset Pre-trained kernel estimation model and deblurring model The method is as follows: Step 3.1, Construct the kernel estimation model loss function The details are as follows; in, Represents the source domain dataset Clear images in Represents the source domain dataset The blurred image in Indicates the convolution operation; Step 3.2, according to the loss function described in step 3.1 Training kernel estimation model The details are as follows: in, This represents the learning rate of the kernel estimation model. This represents the number of images processed per batch by the kernel estimation model; Step 3.3, Constructing a Deblurring Model loss function The details are as follows: in, For weight parameters, Indicates the Laplacian operator. This represents the VGG-Net feature extractor; Step 3.4, according to the loss function described in step 3.3 Training the deblurring model The details are as follows: in, Deblurring model The learning rate Deblurring model Number of images processed per batch.
5. The domain-adaptive image deblurring algorithm according to claim 1, characterized in that, In step four, the target domain blurred image dataset is used. Kernel estimation model Construct a support set that is identically distributed to the target domain. The method is as follows: Step 4.1: Construct a sharpness detection module; use the sharpness detection module to acquire the blurred image of the target domain. Chinese pseudo-clear image patches and blurred image patches First, the original image Perform geometric transformations to the left, right, up, and down, as follows: in, Original image The image after the leftward translation operation, Original image The image after being shifted to the right. Original image The image after the upward translation operation, Original image The image after being shifted downwards; then, the original image... and images after four geometric transformations , , and The peak signal-to-noise ratio (PSNR) is calculated for each sliding window, as shown below: in, Represents the original image The image after geometric transformation The mean squared error function is used; finally, the images after the four geometric transformations in each sliding window are compared with the original image. The sum of the peak signal-to-noise ratios (PSNR) and the original image corresponding to the sliding window with the minimum sum of PSNR. This is denoted as a pseudo-clear image block. The original image corresponding to the sliding window with the largest sum of peak signal-to-noise ratio. Denote as blurred image patch ; Step 4.2: Utilize the pseudo-sharp image patch dataset Blurred image patch dataset and target domain blurred image dataset Fine-tuning the kernel estimation model This prompts the kernel estimation model The motion blur information that is distributed in the same way as the target domain is extracted, as shown below: in, Indicates the learning rate. The loss represents the data being identically distributed, as follows: in, This represents the loss of the Fast Fourier Transform. The discrete wavelet transform loss is represented as follows: Step 4.3: Utilize the fine-tuned kernel estimation model Acquiring and pseudo-sharp image patches Corresponding to and with the blurred image in the target domain Identically distributed pseudo-blurred image patches And utilize pseudo-clear image patches and pseudo-blurred image patches Building support sets Among them, pseudo-blurred image blocks The calculation is as follows:
6. The domain-adaptive image deblurring algorithm according to claim 1, characterized in that, Step five utilizes support sets Fine-tuning the deblurring model And obtain the updated deblurred model. The method is as follows: Using support sets By combining steps 3.3 and 3.4 above, fine-tuning the deblurring model .
7. The domain-adaptive image deblurring algorithm according to claim 1, characterized in that, The updated deblurring model is used in step six. Acquire target domain blurred image Corresponding clear image The method is as follows: