Image deblurring method based on amplitude phase and time domain channel fusion network

Abstract

The invention provides an image deblurring method based on amplitude, phase and time domain channel fusion, a network adopts a U-shaped architecture, an adaptive amplitude and phase compensation strategy is introduced, and information weights of different frequency bands are dynamically adjusted on a channel level; the method comprises two stages and five steps: a training stage: S1: in the training stage, obtaining an image deblurring data set and preprocessing an image; s2, constructing an image deblurring model based on amplitude phase and time domain channel fusion; s3, optimizing a loss function; s4, training and constructing an image deblurring type based on amplitude phase and time domain channel fusion; and S5, performing performance evaluation on the training model in the test stage. According to the method, the time domain and frequency domain feature enhancement strategies are combined, the image definition is recovered, meanwhile, the influence of low resolution and blurring is effectively reduced, and therefore the overall performance of a deblurring task is improved.

Description

Technical Field

[0001] The present invention relates to an image deblurring method based on the fusion of amplitude, phase and time domain channels. The method is applied to the enhancement task of low-resolution blurred images in unmanned aerial vehicle vision systems. The present invention provides a deblurring method that utilizes the image deblurring method based on the fusion of amplitude, phase and time domain channels, combined with multi-scale frequency domain decomposition of amplitude and phase information, and combined with a time-frequency domain channel enhancement module to achieve dynamic balance between global and local information. Background Art

[0002] During high-speed flight, dynamic factors such as rotation and vibration can cause image blur. Furthermore, in drone vision systems, the images output by small sensors generally have low resolution. The combination of these two factors creates blurry low-resolution images.

[0003] Traditional methods for image deblurring are primarily categorized into two types: time-domain and frequency-domain enhancement. Time-domain methods enhance image clarity through linear combinations or second-order differential calculations. Frequency-domain methods employ techniques such as Gaussian difference and harmonic enhancement to improve image smoothness and texture. However, low-resolution images suffer from a lack of high-frequency information, making it difficult to discern fine structures. Consequently, traditional methods are limited in their ability to restore complex backgrounds and details in low-resolution images.

[0004] In recent years, deep learning has made significant progress in image deblurring. However, existing deep learning methods suffer from deficiencies in low-resolution scenarios, such as insufficient feature extraction and poor blur suppression. This patent overcomes the dual limitations of motion blur and resolution degradation by combining multi-scale feature enhancement, time-frequency information fusion, and an adaptive attention mechanism, improving the restoration quality of low-resolution blurred images. Summary of the Invention

[0005] This paper provides an image deblurring method based on the fusion of amplitude, phase, and time domain channels. This method uses a Multi-Frequency Channel Mixing Network (MFCMN) to extract amplitude and phase information from the input image using a fast Fourier transform (FFT) and introduces an adaptive phase compensation strategy. This method dynamically adjusts the weights of information in different frequency bands at the channel level, optimizing the balance between global context and local detail information, and enhancing low-resolution image restoration.

[0006] like Figure 1As shown, the amplitude phase and time domain channel fusion network MFCMN adopts a U-shaped network architecture (UNet) as a whole, and generates multi-scale features by symmetrically improved time-frequency encoder and decoder modules TFBLOCK (Time-frequency encoder-decoder block) of three scales. The first layer corresponds to the original scale of the blurred image, the second layer corresponds to the scale after 0.5 times downsampling, and the third layer corresponds to the scale after 0.25 times downsampling. Each level contains corresponding encoding and decoding processes, aiming to obtain multi-scale low-resolution image features, and the U-shaped structure can maintain a lightweight design and reduce the overall complexity of the network. In the present invention, the phase-amplitude enhancement module MFFN (Multiscale Frequency Fusion Network) module is added to the time-frequency encoder / decoder feature fusion path. Through the auxiliary optimization of the frequency domain amplitude and phase information, the contrast and high-frequency structure details of the image are enhanced from the amplitude and phase perspectives, and the time-frequency domain channel enhancement module TF-SCA (Times Frequency Self-Channel The Attention module enhances the flow of channel information through fully connected mapping and assigns different channel weights during multi-scale feature extraction to retain important feature information. Finally, the TFLS (Time Frequency Loss) joint loss function combines amplitude and phase losses in the time and frequency domains to enhance the ability to recover image details. The encoder extracts multi-scale features and gradually downsamples them. The decoder reconstructs the image by upsampling and combining skip links to fuse information at different scales. This achieves efficient restoration of blurred images, enhancing detail clarity and structural integrity.

[0007] like Figure 2 As shown in the flowchart, the technical solution of the present invention is divided into two stages and the following steps:

[0008] Training phase:

[0009] S1: In the training phase, the image deblurring dataset is obtained and the images are preprocessed;

[0010] S2: Construct an image deblurring model based on the fusion of amplitude, phase and time domain channels;

[0011] S3: optimize the loss function;

[0012] S4: Training and constructing image deblurring model based on amplitude phase and time domain channel fusion;

[0013] Testing phase:

[0014] S5: The performance of the trained model is evaluated in the testing phase.

[0015] The step S1 specifically includes the following steps:

[0016] S1.1: Obtain an image deblurring dataset, preprocess the image deblurring dataset, and obtain a training set of the image deblurring dataset;

[0017] S1.2: Preferably, the image deblurring dataset is collected from low-resolution blur datasets such as motion blur and drone aerial photography.

[0018] The step S2 specifically includes the following steps:

[0019] S2.1: Introducing a U-shaped network model, replacing the original codec with a time-frequency codec in the backbone network of the original network model. Subsequently, a time-frequency channel enhancement module and a phase-amplitude enhancement module are added to the time-frequency codec to enhance the model's multi-scale feature extraction capabilities.

[0020] S2.2: Reference Figure 3 The first layer of the time-frequency encoder / decoder (TFBLOCK) consists of two MFFN modules, a TF-SCA module, an LN, and a convolutional layer. The module contains 6 convolutional layers and 2 residual blocks. The overall three-layer deep encoding and decoding architecture is adopted. A convolutional layer with a stride of 2 is inserted between every two encoding layers to downsample the feature map. The decoder uses Pixel-Shuffle (PS) for upsampling at the corresponding position. It also contains 6 convolutional layers and 2 residual blocks. In addition, between the symmetric layers of the time-frequency encoder and decoder, additional skip connections are introduced to fuse feature information, thereby enhancing feature expression capabilities.

[0021] S2.3: If Figure 4 The phase-amplitude enhancement module MFFN shown in the figure adds a branch for extracting the phase-amplitude features of the image in the frequency domain. The branch processing process of the phase-amplitude enhancement is as follows:

[0022] First, the fast Fourier transform (FFT) is used to perform multi-scale frequency domain decomposition on the input image to extract the amplitude and phase information of different frequency bands, which can be expressed as follows: In the formula Indicates setting a patch window on the channel to reduce the amount of data processed each time; F represents the fast Fourier transform (FFT), which enhances the extracted frequency domain amplitude and phase information separately, especially in the high-frequency part, by enhancing the amplitude information. The formula is as follows:

[0023]

[0024]

[0025] At the same time, the fuzzy phase is decomposed into the real phase Compared with the motion phase Δθ, the compensation of Δθ is implicitly learned through the loss function, which enhances the effective edge and texture information during the up and down sampling of the image. The phase in the frequency domain represents the phase offset or delay of the signal at each frequency. The phase can be calculated using the inverse tangent function. The phase formula is as follows:

[0026]

[0027]

[0028] Among them, among them, represents the real part of the complex number X(f), Represents the imaginary part of the complex number X(f). The amplitude data is then enhanced by softmax. , atan2 is the inverse tangent function, which returns the angle of the tangent value of a given point. In order to further enhance the flow of information, the module also introduces an adaptive phase compensation strategy to enhance the information of different frequency bands at the channel level. The self-learning parameter W is used to compensate the phase weight at different scales. A conservative strategy is adopted for W to retain all phase information at the beginning, and the loss function is used to gradually eliminate the fuzzy phase information. is the phase data extracted by the adaptive kernel,

[0029] The amplitude and phase after feature extraction are re-joined into new frequency domain data. The amplitude and phase splicing conversion formula is as follows:

[0030]

[0031] Where i is the imaginary unit, which represents the amplitude f in the frequency domain. r (t) and phase f i (t) is spliced, and the enhanced frequency domain amplitude and phase information are re-integrated, and the frequency domain information is converted into a time domain image through inverse fast Fourier transform. In this way, the image restoration process can fully retain high-frequency details while avoiding the distortion of low-frequency information. The formula is as follows:

[0032]

[0033]

[0034] in, represents the inverse fast Fourier transform, Represents the expansion operation of the channel, and finally obtains the output after phase amplitude enhancement ;

[0035] S2.3: If Figure 5The time-frequency domain channel enhancement module mentioned above enhances the image channel feature enhancement branch based on the original channel attention: the design core of the time-frequency domain channel enhancement module is to enhance the local feature interaction by using the 1D convolution kernel of adaptive size and combine the global pooling feature. Convolutional mapping enhances the interaction of global context features, realizes multi-scale feature extraction and dynamic weight generation technology, balances the expression weight of global context information and local detail information, and gives a given input feature. , where H and W represent the spatial dimensions, C represents the number of channels, and the initial convolution layer is and The depth-separable convolution obtains the channel feature X,

[0036] To further enhance the ability to express global and local information, global pooling is used to extract global features. 1D convolution is combined to model local channel dependencies, and 1x1 convolution is used to reduce the amount of computation. Global average pooling (GAP) is used on feature X to extract the global context features of each channel. The calculation formula for GAP is as follows:

[0037] in, It is the channel vector generated by the feature map after GAP, which contains the global spatial information of the input image. Then, 1D convolution is used to capture the local channel dependency. First, the kernel size of the 1D convolution is calculated:

[0038] , ,

[0039] Where b=1,γ=2 are preset hyperparameters. The adaptively adjusted 1D convolution kernel is combined with Global flat features, perform 1D convolution operation, the convolution kernel is ,get .

[0040] The formula for initially generating the weight relationship between channels by combining the local information of 1D convolution with the global information of global pooling is as follows:

[0041] in, and are the weight and bias parameters of the convolutional layer, It is the intermediate feature representation after the convolution operation;

[0042] Hyperbolic tangent activation function (Tanh), its formula is as follows:

[0043]

[0044] The introduction of the Tanh function significantly expands the negative value range of the attention weight, allowing the module to flexibly handle negative weights, thereby avoiding excessive absorption of potentially harmful signals. The resulting attention weight is It is used to dynamically adjust the input features in channels and space. The formula for the adjusted feature representation is as follows:

[0045] Among them, ⊗ represents the pixel-by-pixel channel weighted operation, and the adjusted features Improved multi-scale information flow and dynamic balance between global and local information;

[0046] Step S3 introduces a frequency-domain-based amplitude-phase joint loss function (TFLS) to improve the restoration effect of blurred images by comprehensively utilizing time domain and frequency domain information. The amplitude and phase information in the frequency domain capture the intensity changes and structural characteristics of the image respectively. Combined with the MFCMN model, the joint loss function optimizes image detail recovery through a combination of amplitude loss and phase loss. The specific formula is as follows:

[0047] L mag Amplitude difference loss function (L Loss):

[0048]

[0049] Phase consistency loss function (L phase Loss):

[0050]

[0051] L1 reconstruction loss function (L1 Loss):

[0052]

[0053] in, represents the pixel coordinates of the image, The predicted image pixel values ​​generated by the network, is the pixel value of the real image,

[0054] The final loss function is as follows:

[0055]

[0056] in, Represents amplitude loss, which is used to quantify the difference in frequency domain amplitude; Represents phase loss, which is used to measure the consistency of frequency domain phase. Phase loss function The learning process of Δθ_pred is indirectly constrained by optimizing the loss. It represents the predicted image generated by deblurring, Represents the real image, amplitude loss is used to enhance high-frequency features to ensure that the details and contrast of the image are fully restored; phase loss retains structural information through structural consistency, the phase of the clear image in the training set is compared with the phase of the output image to obtain the amplitude loss error; the amplitude of the clear image in the training set is compared with the amplitude of the output image to obtain the phase loss error; the pixels of the clear image in the training set are compared with the pixels of the output image to obtain the pixel L1 loss error; finally, based on the amplitude loss error, phase loss error, and pixel L1 loss error, the loss is calculated by joint weighting.

[0057] The step S4 specifically includes the following steps:

[0058] S4.1: Set the training parameters and use the Adam algorithm for optimization training. The default training parameters are a batch size of 16, a momentum parameter of 0.9, an initial learning rate of 0.001, and a cosine annealing schedule.

[0059] S4.2: Put the training set and validation set images in the dataset into the image deblurring model based on amplitude phase and time domain channel fusion for training;

[0060] S4.3: Train the model according to the set parameters, repeat the above training stages, adjust the learning rate and number of iterations of model training, until the loss function changes tend to be stable, and obtain the final training model;

[0061] The step S5 is based on the trained image deblurring model based on the fusion of amplitude phase and time domain channels, such as Figure 6 The figure below shows a visual comparison of the proposed method with currently used image deblurring models. The left image shows the original blurred image, and the smaller images (from top left to bottom right) show the results of the clear image, DeblurGAN, DeblurGANv2, MAXIM, MIMO-UNet, Restormer, NAFNet, and MFCMN, respectively. Under the same experimental conditions (using the GoPro test subset), MFCMN's restoration results demonstrate superior visual performance in the following aspects.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] The present invention proposes an image deblurring method based on amplitude, phase and time domain channel fusion, wherein the amplitude, phase and time domain channel fusion network includes a frequency domain phase amplitude enhancement module. From the perspective of the image frequency domain, the high and low frequency information of the features is utilized by simultaneously acquiring information in the spatial domain and frequency domain. Furthermore, the amplitude, phase and time domain channel fusion network reduces the blurring effect of low-resolution images, effectively restoring low-resolution blurred images in equipment such as drones to clearer images. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a schematic diagram of the network model based on amplitude phase and time domain channel fusion of the present invention;

[0065] Figure 2 It is a schematic diagram of the process of the present invention;

[0066] Figure 3 Schematic diagram of the encoder-decoder module structure of the present invention;

[0067] Figure 4 Schematic diagram of the phase and amplitude enhancement module within the codec of the present invention;

[0068] Figure 5 Schematic diagram of the time-frequency domain channel enhancement module within the codec of the present invention;

[0069] Figure 6 This is an example diagram of GoPro test data in an embodiment of the present invention. DETAILED DESCRIPTION

[0070] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings of the present invention. To make the above-mentioned objects, features and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0071] like Figure 2 As shown, the specific implementation of the present invention is divided into two stages and the following steps:

[0072] Training phase:

[0073] S1: In the training phase, the image deblurring dataset is obtained and the images are preprocessed;

[0074] S2: Construct an image deblurring model based on the fusion of amplitude, phase and time domain channels;

[0075] S3: optimize the loss function;

[0076] S4: Training and constructing image deblurring model based on amplitude phase and time domain channel fusion;

[0077] Testing phase:

[0078] S5: The performance of the trained model is evaluated in the testing phase.

[0079] The step S1 specifically includes the following steps:

[0080] S1.1: Prepare a standard image deblurring dataset. The selected datasets are the GoPro dataset, the HIDE dataset, and a hybrid dataset constructed by combining the GoPro and UAV-Rain1k datasets. While the GoPro dataset primarily targets motion blur from handheld devices, the high-frequency dynamic blur patterns it contains (such as rapid translation and rotation) are highly similar to the motion blur observed during high-speed drone flight. By blending GoPro data with the UAV-Rain1k dataset, the model can learn more general dynamic blur characteristics, enhancing its adaptability to drone scenarios. Preprocess the dataset by randomly chopping the experimental data into 256×256 pixels and performing random horizontal flipping before feeding it into the model training. This creates the training set for the image deblurring dataset.

[0081] The step S2 specifically includes the following steps:

[0082] S2.1: Introducing a U-shaped network model, adding a channel fusion module and a phase-amplitude enhancement module after the backbone network of the original network model to enhance the model's multi-scale feature extraction capabilities;

[0083] S2.2: Reference Figure 3 The first layer of the time-frequency encoder / decoder (TFBLOCK) consists of two MFFN modules, a TF-SCA module, an LN, and a convolutional layer. The module contains 6 convolutional layers and 2 residual blocks. The overall three-layer deep encoding and decoding architecture is adopted. A convolutional layer with a stride of 2 is inserted between every two encoding layers to downsample the feature map. The decoder uses Pixel-Shuffle (PS) for upsampling at the corresponding position. It also contains 6 convolutional layers and 2 residual blocks. In addition, between the symmetric layers of the time-frequency encoder and decoder, additional skip connections are introduced to fuse feature information, thereby enhancing feature expression capabilities.

[0084] S2.3: Reference Figure 4 The structure and operation flow of the frequency domain phase amplitude enhancement module are as follows: First, the input image is decomposed in the frequency domain at multiple scales using the fast Fourier transform (FFT) to extract the amplitude and phase information of different frequency bands, which can be expressed as follows: In the formula Indicates setting a patch window on the channel to reduce the amount of data processed each time; F represents the fast Fourier transform (FFT), which enhances the extracted frequency domain amplitude and phase information separately, especially in the high-frequency part, by enhancing the amplitude information. The formula is as follows:

[0085]

[0086]

[0087] At the same time, an adaptive phase compensation strategy is introduced to enhance effective edge and texture information during image upsampling and downsampling. The phase in the frequency domain represents the phase offset or delay of the signal at each frequency. The phase can be calculated using the inverse tangent function. The phase formula is as follows:

[0088]

[0089]

[0090] Among them, among them, represents the real part of the complex number X(f), Represents the imaginary part of the complex number X(f). The amplitude data is then enhanced by softmax. , atan2 is the inverse tangent function, which returns the angle of the tangent value of a given point. In order to further enhance the flow of information, the module also introduces an adaptive phase compensation strategy to enhance the information of different frequency bands at the channel level. The self-learning parameter W is used to compensate the phase weight at different scales. A conservative strategy is adopted for W to retain all phase information at the beginning, and the loss function is used to gradually eliminate the fuzzy phase information. is the phase data extracted by the adaptive kernel,

[0091] The amplitude and phase after feature extraction are re-joined into new frequency domain data. The amplitude and phase splicing conversion formula is as follows:

[0092]

[0093] Where i is the imaginary unit, which represents the amplitude f in the frequency domain. r (t) and phase f i (t) is spliced, and the enhanced frequency domain amplitude and phase information are re-integrated, and the frequency domain information is converted into a time domain image through inverse fast Fourier transform. In this way, the image restoration process can fully retain high-frequency details while avoiding the distortion of low-frequency information. The formula is as follows:

[0094]

[0095]

[0096] in, represents the inverse fast Fourier transform, Represents the expansion operation of the channel, and finally obtains the output after phase amplitude enhancement ;

[0097] S2.4: Reference Figure 5 , the structure and operation flow of the time-frequency domain channel enhancement module are as follows: given the input feature , where H and W represent the spatial dimensions, C represents the number of channels, and the initial convolution layer is and The channel feature X is obtained by depthwise separable convolution. To further enhance the ability to express global and local information, global features are extracted through global pooling. 1D convolution is combined to model local channel dependencies, and 1x1 convolution is used to reduce the dimension and computational complexity. Global average pooling (GAP) is used to extract the global context feature of each channel from the feature X. The calculation formula of GAP is as follows:

[0098]

[0099] in, It is the channel vector generated by the feature map after GAP, which contains the global spatial information of the input image. Then, 1D convolution is used to capture the local channel dependency. First, the kernel size of the 1D convolution is calculated:

[0100] , ,

[0101] Where b=1,γ=2 are preset hyperparameters. The adaptively adjusted 1D convolution kernel is combined with Global flat features, perform 1D convolution operation, the convolution kernel is ,get .

[0102] The formula for initially generating the weight relationship between channels by combining the local information of 1D convolution with the global information of global pooling is as follows:

[0103]

[0104] in, and are the weight and bias parameters of the convolutional layer, It is the intermediate feature representation after the convolution operation;

[0105] Hyperbolic tangent activation function (Tanh), its formula is as follows:

[0106]

[0107] The introduction of the Tanh function significantly expands the negative value range of the attention weight, allowing the module to flexibly handle negative weights, thereby avoiding excessive absorption of potentially harmful signals. The resulting attention weight is It is used to dynamically adjust the input features in channels and space. The formula for the adjusted feature representation is as follows:

[0108]

[0109] Among them, ⊗ represents the pixel-by-pixel channel weighted operation, and the adjusted features Improved multi-scale information flow and dynamic balance between global and local information;

[0110] The step S3 specifically includes the following steps:

[0111] S3.1: Calculate the difference between the network output image and the clear image in the training set to get the loss and return the gradient to update the network parameters. Repeat this process until the number of training times reaches the preset number. and The reconstruction results and real clear data output by the network are respectively. The loss functions of network training are the combination of the following three loss functions:

[0112] Lmag amplitude difference loss function (L Loss):

[0113]

[0114] Phase consistency loss function (Lphase Loss):

[0115]

[0116] L1 reconstruction loss function (L1 Loss):

[0117]

[0118] in, represents the pixel coordinates of the image, The predicted image pixel values ​​generated by the network, is the pixel value of the real image,

[0119] The final loss function is as follows:

[0120]

[0121] in, 、 and are weight hyperparameters, set as =0.7, =0.1 and =0.2, by adjusting these three weights, the joint loss function can achieve a good balance between restoring the pixel intensity and frequency characteristics of the image.

[0122] The step S4 specifically includes the following steps:

[0123] S4.1: Use the Adam optimizer (β1 = 0.9, β2 = 0.9, weight decay 0) to train the model. Set the batch size to 16, the total number of iterations to 200K, the initial learning rate to 1e−4 and gradually reduce it to 1e−6, and use a cosine annealing schedule.

[0124] S4.2: Inputting the blurred image into an initial amplitude-phase and time-domain channel fusion network to obtain an output image;

[0125] S4.3: Based on the loss error between the output image and the clear image, calculate the loss and perform gradient inversion, and update the parameters of the initial frequency-domain phase amplitude and time-domain channel fusion network;

[0126] S4.4: Repeat the above training steps until the number of training times reaches a preset number, and obtain the target amplitude phase and time domain channel fusion network.

[0127] The step S5 specifically includes the following steps:

[0128] S5.1: Reference Figure 6 The left picture is the original blurred image, and the small pictures from the upper left to the lower right are the clear images respectively. The results of DeblurGAN, DeblurGANv2, MAXIM, MIMO-UNet, Restormer, NAFNet, and MFCMN. Under the same experimental conditions (GoPro test subset is used uniformly), the restoration results of MFCMN show better visual effects in the following aspects. By comparing with typical methods such as MIMO-UNet, it can be seen that the present invention effectively alleviates the common edge artifact effects and texture loss problems in low-resolution blurred images through frequency domain phase compensation and channel enhancement mechanism.

[0129] The main feature of the present invention lies in the design of the frequency domain phase amplitude enhancement module MFFN, which includes a frequency domain phase amplitude enhancement module based on the amplitude phase and time domain channel fusion network. From the perspective of the image frequency domain, the high and low frequency information of the feature is utilized by simultaneously acquiring information in the spatial domain and frequency domain. The frequency domain phase amplitude enhancement module successfully integrates the frequency domain information of the feature into the convolutional neural network, increases the model's acquisition and processing of frequency domain information, and improves the performance of the model.

[0130] The above is only a preferred specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. An image deblurring method based on an amplitude-phase and time-domain channel fusion network, characterized in that: It is divided into two phases and the following steps: Training phase: S1: In the training phase, the image deblurring dataset is obtained and the images are preprocessed; S2: Construct an image deblurring model based on the amplitude-phase and time-domain channel fusion network; S3: optimize the loss function; S4: Train and build an image deblurring model based on the amplitude phase and time domain channel fusion network; Testing phase: S5: The performance of the trained model is evaluated in the testing phase.

2. The image deblurring method based on the amplitude phase and time domain channel fusion network according to claim 1, characterized in that: The step S1 specifically includes the following steps: S1.1: Obtain an image deblurring dataset, the dataset being a given channel Width High is 3 256 256 dataset D.

3. The image deblurring method based on the amplitude-phase and time-domain channel fusion network according to claim 1, characterized in that: The step S2 specifically includes the following steps: S2.1: Based on the blurred images and clear images in the training set, an initial amplitude-phase and time-domain channel fusion network is trained to obtain a target amplitude-phase and time-domain channel fusion network; S2.2: Inputting the blurred image to be measured into the target-based amplitude-phase and time-domain channel fusion network to perform image deblurring processing to obtain a clear image of the target; S2.3: The initial amplitude-phase and time-domain channel fusion network adopts a U-shaped network structure for blurred image restoration, including three-layer scaling and an improved time-frequency coding and decoding architecture. The codec and scaling layers include: a primary codec and scaling, a secondary codec and scaling, and a tertiary codec and scaling. S2.4: The single-layer time-frequency encoder / decoder consists of two phase-amplitude enhancement modules, a time-frequency channel enhancement module, and a LN convolutional layer. The module contains six convolutional layers and two residual blocks. The overall architecture uses a three-layer deep U-shaped network encoder / decoder. A convolutional layer with a stride of 2 is inserted between every two encoding layers to downsample the feature maps. The decoder uses pixel-shuffle upsampling at the corresponding position. It also contains six convolutional layers and two residual blocks. In addition, skip connections are introduced between the symmetric layers of the time-frequency encoder and decoder to fuse feature information, thereby enhancing feature expression capabilities. S2.5: The phase-amplitude enhancement module adds a branch for extracting the phase-amplitude features of the image in the frequency domain. The branch processing process of the phase-amplitude enhancement is as follows: First, the fast Fourier transform (FFT) is used to perform multi-scale frequency domain decomposition on the input image to extract the amplitude and phase information of different frequency bands, which can be expressed as follows: In the formula Indicates setting a patch window on the channel to reduce the amount of data processed each time. F represents the fast Fourier transform (FFT), which enhances the extracted frequency domain amplitude and phase information separately, especially in the high-frequency part, by enhancing the amplitude information. The formula is as follows: ； ； At the same time, the fuzzy phase is decomposed into the real phase The compensation of Δθ is implicitly learned through the loss function, which enhances the effective edge and texture information during the up- and down-sampling of the image, thereby reducing the artifacts caused by the frequency domain phase mismatch. The phase in the frequency domain represents the phase offset or delay of the signal at each frequency. The phase can be calculated using the inverse tangent function. The phase formula is as follows: ； ； Among them, among them, represents the real part of the complex number X(f), Represents the imaginary part of the complex number X(f); the amplitude data is then enhanced by softmax to obtain , atan2 is the inverse tangent function, which returns the angle of the tangent value of a given point. In order to further enhance the flow of information, the module also introduces an adaptive phase compensation strategy to enhance the information of different frequency bands at the channel level. The self-learning parameter W is used to compensate the phase weights at different scales. A conservative strategy is adopted for W to retain all phase information at the beginning, and the loss function is used to gradually eliminate the phase information of the artifact. is the phase data extracted by the adaptive kernel, The amplitude and phase after feature extraction are re-joined into new frequency domain data. The amplitude and phase splicing conversion formula is as follows: Where i is the imaginary unit, which represents the amplitude f in the frequency domain. r (t) and phase f i (t) is spliced, and the enhanced frequency domain amplitude and phase information are re-integrated, and the frequency domain information is converted into a time domain image through inverse fast Fourier transform. In this way, the image restoration process can fully retain high-frequency details while avoiding the distortion of low-frequency information. The formula is as follows: ； ； in, represents the inverse fast Fourier transform, Represents the expansion operation of the channel, and finally obtains the output after phase amplitude enhancement ; S2.6: The time-frequency domain channel enhancement module enhances the image channel feature enhancement branch based on the original channel attention: The core of the design of the time-frequency domain channel enhancement module is to use multi-scale feature extraction and dynamic weight generation technology to balance the expression weights of global context information and local detail information. Given the input feature , where H and W represent the spatial dimensions, C represents the number of channels, and the initial convolution layer is and The channel feature X is obtained by depthwise separable convolution. To further enhance the ability to express global and local information, global features are extracted through global pooling, local channel dependencies are modeled by combining 1D convolution, and 1x1 convolution is used to reduce the computational complexity. Global average pooling (GAP) is used to extract the global context feature of each channel from the feature X. The calculation formula of GAP is as follows: ； in, It is the channel vector generated by the feature map after GAP, which contains the global spatial information of the input image. Then, 1D convolution is used to capture the local channel dependency. First, the kernel size of the 1D convolution is calculated: ， ， ； Among them, b=1, γ=2 are preset hyperparameters; the adaptively adjusted 1D convolution kernel is combined with Global flat features, perform 1D convolution operation, the convolution kernel is ,get ; The formula for initially generating the weight relationship between channels by combining the local information of 1D convolution with the global information of global pooling is as follows: ； in, and are the weight and bias parameters of the convolutional layer, It is the intermediate feature representation after the convolution operation; Hyperbolic tangent activation function (Tanh), its formula is as follows: ； The introduction of the Tanh function significantly expands the negative value range of the attention weight, allowing the module to flexibly handle negative weights, thereby avoiding excessive absorption of potentially harmful signals. The resulting attention weight is It is used to dynamically adjust the input features in channels and space. The formula for the adjusted feature representation is as follows: ； Among them, ⊗ represents the pixel-by-pixel channel weighted operation, and the adjusted features Improves multi-scale information flow and dynamically balances global and local information.

4. The image deblurring method based on the amplitude phase and time domain channel fusion network according to claim 1, characterized in that: Step S3 introduces a frequency-domain-based amplitude-phase joint loss function (TFLS) to improve the restoration effect of blurred images by comprehensively utilizing time domain and frequency domain information. The amplitude and phase information in the frequency domain capture the intensity changes and structural characteristics of the image respectively. Combined with the MFCMN model, the frequency domain loss function optimizes image detail recovery through a combination of amplitude loss and phase loss. The phase loss function indirectly constrains the learning process of motion phase Δθ compensation by optimizing the loss, and adds amplitude loss to enhance high-frequency features to ensure that the details and contrast of the image are fully restored.

5. The image deblurring method based on the amplitude-phase and time-domain channel fusion network according to claim 1, characterized in that: The step S4 specifically includes the following steps: S4.1: Set the training parameters and use the Adam algorithm for optimization training. The default training parameters are a batch size of 16, a momentum parameter of 0.9, an initial learning rate of 0.001, and a cosine annealing schedule. S4.2: Inputting the blurred image into an initial amplitude-phase and time-domain channel fusion network to obtain an output image; S4.3: Based on the loss error between the output image and the clear image, calculate the loss and perform gradient inversion, and update the parameters of the initial frequency-domain phase amplitude and time-domain channel fusion network; S4.4: Repeat the above training steps until the number of training times reaches a preset number, and obtain the target amplitude phase and time domain channel fusion network.

6. The image deblurring method based on the amplitude phase and time domain channel fusion network according to claim 1, characterized in that: In step S5, based on the trained image deblurring model based on amplitude-phase and time-domain channel fusion, the proposed method was visually compared with currently common image deblurring models under the same experimental conditions (using the GoPro test subset) to qualitatively evaluate the deblurring effect of different network models. The deblurring effect of the model was evaluated by comparing texture details and the filtering effect of blur artifacts.

Images