Image deblurring method based on adaptive amplitude filtering and gradient enhancement

Abstract

According to the image deblurring method based on adaptive amplitude filtering and gradient enhancement, a U-shaped network architecture is adopted, frequency domain information is optimized through an adaptive time-frequency fusion strategy, and artifacts are reduced. The method comprises a training stage and a testing stage, and comprises the following five steps: S1, in the training stage, obtaining an image deblurring data set and preprocessing an image; s2, constructing an image deblurring model based on adaptive amplitude filtering and gradient enhancement; s3, optimizing a loss function; s4, training the constructed image deblurring model of adaptive amplitude filtering and gradient enhancement; and S5, performing performance evaluation on the training model in the test stage. According to the method, testing is carried out on an unmanned aerial vehicle data set, the result shows that artifacts are effectively reduced in the aspects of PSNR, SSIM, calculation efficiency and the like, visual consistency is improved, and a new scheme is provided for high-quality blurred image restoration.

Description

Technical Field

[0001] The present invention relates to an image deblurring method based on adaptive amplitude filtering and gradient enhancement, which is applied to the enhancement task of blurred images in an unmanned aerial vehicle (UAV) vision system. The present invention provides an image deblurring method based on adaptive amplitude filtering and gradient enhancement. The method dynamically adjusts the amplitude frequency through an adaptive amplitude filtering module, and introduces gradient features combined with channel attention in the process of time-domain feature extraction to highlight the high-frequency information in the blurred image and suppress the blurring artifacts. Background Art

[0002] In recent years, UAV technology has been widely used and has played an important role in fields such as military reconnaissance, agricultural monitoring, disaster relief, environmental protection, and urban planning. Since UAVs are usually in a moving state and are affected by various external environmental factors, the captured images are often accompanied by different degrees of blurring phenomena such as motion blur, atmospheric disturbance blur, and focus blur. In recent years, deep learning methods have made breakthroughs in the UAV image deblurring task. Most deblurring methods mainly use multi-scale feature extraction, time-frequency domain and other feature enhancement means to deblur and improve the image quality. Conventional frequency-domain enhancement methods only enhance the frequency-domain components in a fixed range and cannot effectively extract high-frequency components, which risks amplifying the artifact features. During multi-scale feature extraction, the time-frequency domain conversion between different scales is prone to phase distortion due to scale mismatch, resulting in artifacts. Therefore, this patent establishes a multi-scale time-frequency domain conversion model and uses a method that combines time-domain and frequency-domain features and adaptively enhances the frequency-domain information to suppress blurring and artifacts and improve the UAV image deblurring performance. Summary of the Invention

[0003] The present invention proposes an image deblurring method based on adaptive amplitude filtering and gradient enhancement, which uses an adaptive amplitude filtering module to dynamically adjust the enhancement strategy of high and low amplitude frequencies, thereby avoiding the artifact problem of spectrum mixing in the multi-scale model training. At the same time, in the process of time-domain feature extraction, gradient features combined with the CA (channel attention) mechanism are introduced to improve the recovery ability of edges and high-frequency information.

[0004] Such as Figure 1As shown in the figure, the Gradient-Guided Frequency Adaptive Transformer Network (GFTA-Net) for adaptive amplitude filtering and gradient-enhanced image deblurring adopts a U-shaped encoder-decoder architecture, replacing the encoder-decoder with an improved GFBlock. It is divided into three layers, and each layer enhances the feature expression ability through downsampling and channel expansion. The first-layer encoding GFBlock1 takes the original image as input and keeps the size unchanged; the second-layer encoding GFBlock2 downsamples the spatial features to 0.5 times the original image size and expands the channels to 2 times the original; the third-layer encoding GFBlock3 further downsamples the original image by 0.25 times and expands the channels 4 times. The input image first extracts local information through a 3×3 convolution and mixes channel features through a 1×1 convolution. The encoding GFBlock extracts features layer by layer, and at the same time passes the low-level features to the decoding GFBlock through skip connections to retain multi-scale resolution information. The GFBlock combines cross-layer information for feature recovery. The GFBlock structure on the right shows the core computing unit of the model. Each encoder and decoder contains a Gradient-Channel Attention (GCA) and a Frequency-Spatial Self-Attention (FTSA) module. Among them, GCA adjusts the channel weights through gradient information, guiding the model to focus on edges and high-frequency details and reducing redundant information; FTSA effectively models high-frequency information by calculating self-attention in the frequency domain, preventing the loss of high-frequency details or the generation of blurring artifacts during the deblurring process. Overall, this model uses multi-scale feature fusion and frequency-domain attention mechanisms to strengthen the image deblurring ability and improve the image detail retention effect.

[0005] As Figure 2 shown in the process schematic diagram, the technical solution adopted by the present invention is divided into two stages and the following steps:

[0006] Training stage:

[0007] S1: Obtain an image deblurring dataset and preprocess the images during the training stage;

[0008] S2: Construct an image deblurring model based on adaptive amplitude filtering and gradient enhancement;

[0009] S3: Optimize the loss function;

[0010] S4: Train the constructed image deblurring model based on adaptive amplitude filtering and gradient enhancement;

[0011] Testing stage:

[0012] S5: Evaluate the performance of the trained model during the testing phase.

[0013] The specific steps of step S1 include the following steps:

[0014] S1.1: Obtain the UAV blurred image dataset, extract motion blur, blur artifacts, etc. from the dataset images to construct a dataset, and divide it into a training set, a validation set, and a test set according to the ratio of 8:1:1;

[0015] The specific steps of step S2 include the following steps:

[0016] S2.1: The deblurring network for initial adaptive amplitude filtering and gradient feature extraction adopts an upsampling and downsampling blurred image restoration network structure, including: a three-layer GFBlock encoder-decoder structure; the first-layer encoding GFBlock1 corresponds to the original scale of the blurred image, the second-layer encoding GFBlock2 corresponds to the scale of the blurred image after being downsampled once, the third-layer encoding GFBlock3 corresponds to the scale of the blurred image after being downsampled twice, and the first-layer decoding GFBlock1, the second-layer decoding GFBlock2, and the third-layer decoding GFBlock3 include: a gradient-guided attention module and a frequency-domain amplitude self-attention module;

[0017] S2.2: As Figure 3 shown, for the gradient-guided attention module GCA, the input image I extracts horizontal and vertical gradients through the Sobel operator , and enhances the edge features by calculating the gradient amplitude . The original feature map X is concatenated with the gradient map G to obtain . The Sobel operator acts as a high-pass filter to strengthen the image edges (high-frequency information) through convolution operations, alleviating the edge degradation problem in blurred images; the concatenation operation injects gradient information into the subsequent network, providing richer initial features for frequency-domain processing;

[0018] Apply channel attention (CA) to the concatenated features . The calculation process is as follows: First, compress the spatial dimension through global average pooling (GAP). The pooling formula is as follows:

[0019]

[0020] This step extracts channel-level statistics, reflecting the global importance of each channel. Subsequently, generate channel weights through a fully connected layer and an activation function:

[0021]

[0022] Where and is a learnable parameter, r is the compression ratio, δ is the ReLU activation, σ is the Sigmoid activation function, and finally the channel weights are adjusted by channel weighting. The weighting formula is as follows: , the gradient guidance module introduces gradient features into the CA mechanism through channel mapping of the fully connected layer to suppress redundant channels, enhance key channels, and improve the model's ability to express fuzzy sensitive features;

[0023] S2.3: As Figure 4 shown in the frequency domain amplitude self-attention module, the weighted input features are transformed into the frequency domain to extract amplitude information , and the energy distribution of different frequency components is reflected by the amplitude. Blurred images usually show low-frequency energy dominance and high-frequency energy attenuation. Directly processing the amplitude helps to restore high-frequency details.

[0024] Apply self-attention to the amplitude matrix M:

[0025] ,

[0026] By calculating the correlation between frequency components, the gradient features and channel features are combined with the global features, and the overall amplitude distribution of the image is adjusted according to the edge distribution, enhancing key frequencies and suppressing noise or artifact-related frequencies. For the enhanced amplitude perform inverse Fourier transform (IFFT) to obtain time-domain features , and at the same time, to avoid loss of phase information, the original phase is retained to ensure the structural consistency of the reconstructed image;

[0027] The total loss function formula for the step S3 is as follows:

[0028]

[0029] where λ1 is the weight of the pixel loss function, λ2 is the weight of the gradient consistency loss function, λ3 is the weight of the frequency domain loss function, and λ4 is the weight of the amplitude loss function;

[0030] To ensure that the model output is as close as possible to the real clear image at the pixel level, the loss function can better retain edge information and is more robust to outliers. By the loss function can enhance the overall brightness, contrast, and detail recovery ability of the image; The loss function formula is as follows:

[0031]

[0032] Among them is the model output image, is the ground truth image;

[0033] The auxiliary edge loss function is based on the Sobel filter. The Sobel filter includes horizontal and vertical Sobel convolution kernels, and the convolution kernels are decomposed into the product form of an averaging convolution kernel and a differential convolution kernel. The averaging convolution kernel smooths the image, while the differential convolution kernel detects the edges of the image.

[0034] To restore a clear image, the distance between the output result image and the ground truth image in the Sobel edge space is directly minimized. Therefore, the auxiliary edge loss function is expressed by the following formula:

[0035]

[0036] Where and represent the gradients in the X and Y directions respectively, is the model output image, is the ground truth image;

[0037] Ensure that the model is optimized not only in the time domain but also matches the frequency distribution of the ground truth image in the frequency domain, reducing blurring artifacts, Emphasize the consistency of the entire spectrum to ensure that the model does not lose information at specific frequencies, Only constrain the amplitude of the spectrum to reduce the frequency shift caused by blurring,

[0038] The frequency domain loss function is as follows:

[0039]

[0040] The amplitude loss function is as follows:

[0041]

[0042] Step S4 specifically includes the following steps:

[0043] S4.1: Set the training parameters and use the Adam algorithm for optimization training. The initial preset parameters for training are: the batch size batch is 16, the momentum parameter momentum is 0.9, the initial learning rate is 0.001, and the number of iterations epoch is 10000;

[0044] S4.2: Put the training set and validation set images in the dataset into the image deblurring model with adaptive amplitude filtering and gradient enhancement for training;

[0045] S4.3: Train the model according to the set parameters. By observing the changing trend of the loss function, adjust the learning rate and the number of iterations of model training until the change of the loss function tends to be stable, and obtain the final trained model;

[0046] In step S5, according to the trained image deblurring model based on adaptive amplitude filtering and gradient enhancement, refer to Figure 6 , the left figure is the original blurred image, and the small figures from the upper left to the lower right are the clear images, the results of DeblurGAN, DeblurGANv2, MAXIM, MIMO-UNet, Restormer, NAFNet, and GFTA-Net respectively. The texture restoration ability and the ability to remove image blur artifacts of the model are evaluated through qualitative comparison of visual effects.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] Enhance the feature retention ability of blurred images: enhance key frequency information through adaptive amplitude filtering, reduce artifacts caused by phase distortion, and improve the detail restoration ability: supplement high-frequency information through the gradient feature extraction module to enhance the texture and edges of the image; at the same time, adopt a time-frequency domain collaborative optimization strategy to reduce artifacts caused by scale mismatch during the deblurring process. The present invention provides an image deblurring method based on adaptive amplitude filtering and gradient enhancement, effectively improving the deblurring effect of images and providing a new solution for the UAV image deblurring task. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a schematic flow chart in the present invention;

[0050] Figure 2 is a schematic diagram of the adaptive amplitude filtering and gradient enhancement model in the present invention;

[0051] Figure 3 is a schematic diagram of the gradient-guided attention module in the present invention;

[0052] Figure 4 is a schematic diagram of the frequency-domain amplitude self-attention module in the present invention;

[0053] Figure 5 is a gradient comparison diagram before and after adding different modules in the embodiment of the present invention;

[0054] Figure 6 is an example diagram of the test data of the UAV dataset UAV-Rain1k in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0055] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings of the present invention.

[0056] Embodiment:

[0057] As Figure 2 shown in the process schematic diagram, the technical solution adopted by the present invention is divided into two stages and the following steps:

[0058] Training stage:

[0059] S1: Obtain an image deblurring dataset in the training stage and preprocess the images;

[0060] S2: Construct an image deblurring model based on adaptive amplitude filtering and gradient enhancement;

[0061] S3: Optimize the loss function;

[0062] S4: Train the constructed image deblurring model with adaptive amplitude filtering and gradient enhancement;

[0063] Testing stage:

[0064] S5: Evaluate the performance of the trained model in the testing stage.

[0065] The specific steps of step S1 include the following steps:

[0066] S1.1: Select a certain number of blurred images and UAV blurred images from the GoPro dataset, HIDE dataset, and UAV-Rain1k dataset. Before inputting the experimental data into the model for training, randomly cut the data into 256×256 size and perform random horizontal up and down flipping, and divide it into a training set, a validation set, and a test set according to the ratio of 8:1:1;

[0067] The specific steps of step S2 include the following steps:

[0068] S2.1: As Figure 2 shown: The network adopts a U-shaped architecture, extracts initial features through a 3×3 convolutional layer and a 1×1 convolutional layer, and keeps the input size as C×H×W. In the encoding process, the input image enters the first encoding GFBlock1, and the dimension of the feature map becomes , and then is further encoded to . In the decoding process, the input image enters the first decoding GFBlock3, and the dimension of the feature map is upsampled from to , and then is further decoded and upsampled to During the encoding and decoding process of each layer, the gradient-guided attention module is used to adaptively adjust the importance of different channels, thereby enhancing the retention of key information. The frequency-domain amplitude self-attention module FTSA effectively models high-frequency information by calculating self-attention in the frequency domain, preventing the loss of high-frequency details or the generation of blurring artifacts during the deblurring process. Overall, the model utilizes multi-scale feature fusion and frequency-domain attention mechanism to strengthen the image deblurring ability and improve the detail retention effect.

[0069] S2.2: As Figure 3 shown: For the gradient-guided attention module, the input image I extracts horizontal and vertical gradients through the Sobel operator , and enhances the edge features by calculating the gradient amplitude . The original feature map X is concatenated with the gradient map G to obtain . The Sobel operator, as a high-pass filter, strengthens the image edges (high-frequency information) through convolution operations, alleviating the problem of edge degradation in blurred images; the concatenation operation injects gradient information into the subsequent network, providing richer initial features for frequency-domain processing; applying channel attention (CA) to the concatenated features has the following calculation process: First, compress the spatial dimension through global average pooling (GAP),

[0070] Pooling formula:

[0071] This step extracts channel-level statistics, reflecting the global importance of each channel; subsequently, generate channel weights through a fully connected layer and an activation function: , where and are learnable parameters, r is the compression ratio, δ is the ReLU activation, σ is the Sigmoid activation function, and finally adjust the channel weights through channel weighting. The weighting formula is as follows: ; The gradient-guided module introduces gradient features into the CA mechanism through channel mapping of the fully connected layer to suppress redundant channels and enhance key channels, improving the model's ability to express blur-sensitive features;

[0072] S2.3: As Figure 4 shown: For the frequency-domain amplitude self-attention module, convert to the frequency domain and extract the amplitude information . The energy distribution of different frequency components is reflected by the amplitude. Blurred images usually show low-frequency energy dominance and high-frequency energy attenuation. Directly processing the amplitude helps to restore high-frequency details; applying self-attention to the amplitude matrix MM:

[0073] ,

[0074] By calculating the correlation between frequency components, the gradient features and channel features are combined with the global features, and the overall amplitude distribution of the image is adjusted according to the edge distribution to enhance the key frequencies and suppress the noise or artifact-related frequencies; for the enhanced amplitude perform the inverse Fourier transform (IFFT) to obtain the time-domain features , and at the same time, to avoid loss of phase information, the original phase is retained to ensure the structural consistency of the reconstructed image;

[0075] The total loss function formula adopted in the step S3 is as follows:

[0076]

[0077] where λ1 is the weight of the pixel loss function, λ2 is the weight of the gradient consistency loss function, λ3 is the weight of the frequency-domain loss function, λ4 is the weight of the amplitude loss function;

[0078] To ensure that the model output is as close as possible to the real clear image at the pixel level, the loss function can better retain edge information and is more robust to outliers. Through the loss function, the overall brightness, contrast and detail recovery ability of the image can be enhanced; The loss function formula is as follows:

[0079]

[0080] where is the model output image, is the real image;

[0081] The auxiliary edge loss function is based on the Sobel filter; the Sobel filter includes horizontal and vertical Sobel convolution kernels, and the convolution kernels are decomposed into the product form of an averaging convolution kernel and a differential convolution kernel, where the averaging convolution kernel smooths the image and the differential convolution kernel detects the edges of the image;

[0082] To restore the clear image, the distance between the output result image and the ground truth image in the Sobel edge space is directly minimized; therefore, the auxiliary edge loss function is expressed by the following formula:

[0083]

[0084] where and represent the gradients in the X direction and Y direction respectively, is the model output image, is the real image;

[0085] Ensure that the model is optimized not only in the time domain but also in the frequency domain to match the frequency distribution of real images, reducing blurring artifacts. Emphasize the consistency of the entire spectrum to ensure that the model does not lose information at specific frequencies. Only constrain the amplitude of the spectrum to reduce the frequency shift caused by blurring; the frequency-domain loss function is as follows:

[0086]

[0087] The amplitude loss function is as follows:

[0088]

[0089] The specific steps of step S4 include the following steps:

[0090] S4.1: Set training parameters, use the Adam algorithm for optimization training, set the batch size batch of training to 16, the momentum parameter momentum to 0.9, the initial learning rate to 0.001, and the number of iterations epoch to 10000;

[0091] S4.2: Put the training set and validation set images in the dataset into the image deblurring model of adaptive amplitude filtering and gradient enhancement for training;

[0092] S4.3: Train the model according to the set parameters, adjust the learning rate and the number of iterations of model training by observing the change trend of the loss function until the change of the loss function tends to be stable, and obtain the final trained model;

[0093] Such as Figure 5 Figure 6 As shown, step S5 evaluates the deblurred image using the Peak Signal-to-Noise Ratio (PSNR) and the Structural Similarity (SSIM), and also adds evaluation metrics such as the number of model parameters and the deblurring execution time to evaluate the model.

[0094] S5.1: Refer to Figure 5, from left to right are the original gradient map, the model gradient map enhanced by GCA alone, and the model gradient map enhanced by both GCA and FTSA modules. As shown in the figure, the GCA module uses the channel attention mechanism to adaptively weight the extracted gradient information to enhance key edge features. Compared with the original gradient map, this method can better highlight the main edge structure and reduce unnecessary background noise. Finally, on the basis of GCA, FTSA is introduced. The amplitude information of the image is extracted through FFT frequency domain transformation, and the self-attention mechanism is used to enhance key frequency components. This joint method can further optimize the gradient features, improve the expression ability of edge details, reduce high-frequency noise at the same time, and improve the clarity of the image. The results show that this method can effectively enhance the detail and edge information, making the structural features of the image more obvious and sharp;

[0095] S5.2: Refer to Figure 6 , the left figure is the original blurred image, and the small figures from the upper left to the lower right are the clear images, the results of DeblurGAN, DeblurGANv2, MAXIM, MIMO-UNet, Restormer, NAFNet, and GFTA-Net respectively. Through the comparison of the results of MIMO-UNet with the adaptive amplitude filtering and gradient feature extraction network (GFTA-Net), it can be shown in the qualitative comparison of visual effects that the present invention can effectively improve the deblurring ability of the convolutional neural network, effectively improve the texture clarity and filter out the blurred artifacts.

Claims

1. An image deblurring method using adaptive amplitude filtering and gradient enhancement, 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 adaptive amplitude filtering and gradient enhancement; S3: optimize the loss function; S4: Image deblurring model constructed by training adaptive amplitude filtering and gradient enhancement; Testing phase: S5: The performance of the trained model is evaluated in the testing phase.

2. The image deblurring method of adaptive amplitude filtering and gradient enhancement according to claim 1, characterized in that: The step S1 specifically includes the following steps: S1.1: Obtain a dataset of blurred drone images, extract motion blur, blur artifacts and other images from the dataset images to construct a dataset, and divide it into training set, validation set and test set according to the ratio of 8:1:

1.

3. The image deblurring method of adaptive amplitude filtering and gradient enhancement according to claim 1, characterized in that: The step S2 specifically includes the following steps: S2.1: A blurred image restoration network structure using up and down sampling, including: a three-layer encoding and decoding GFBlock structure; the first layer of encoding and decoding GFBlock1 corresponds to the original scale of the blurred image, the second layer of encoding and decoding GFBlock2 corresponds to the scale of the blurred image after downsampling once, and the third layer of encoding and decoding GFBlock3 corresponds to the scale of the blurred image after downsampling twice. The first layer of encoding and decoding GFBlock1, the second layer of encoding and decoding GFBlock2, and the third layer of encoding and decoding GFBlock3 include: a gradient-guided attention module, a Gaussian linear error (GELU), and a frequency domain amplitude self-attention module; S2.2: In encoding GFBlock and decoding GFBlock, the gradient-guided attention module GCA is used to focus on factors such as edge artifacts. The gradient-guided attention module extracts horizontal and vertical gradients from the input image I through the Sobel operator. , by calculating the gradient amplitude Enhance edge features; the original feature map X is concatenated with the gradient map G to obtain ;The Sobel operator acts as a high-pass filter. It strengthens the image edge (high-frequency information) through convolution operation, alleviates the edge degradation problem in blurred images, and the splicing operation injects gradient information into the subsequent network, providing richer initial features for frequency domain processing. After splicing, the features Apply channel attention (CA), the calculation process is as follows: First, compress the spatial dimension through global average pooling (GAP), the pooling formula is as follows: ； This step extracts channel-level statistics to reflect the global importance of each channel, and then generates channel weights through a fully connected layer and an activation function: ； in and is a learnable parameter, r is the compression ratio, δ is the ReLU activation, and σ is the Sigmoid activation function; finally, the channel weight is adjusted by channel weighting, and the weighting formula is as follows: ,The gradient guided module introduces the gradient features into the CA mechanism through the ,channel mapping of the fully connected layer to suppress the redundant channels, ,enhance the key channels, and improve the model’s ability to ,express fuzzy sensitive features; S2.3: When encoding GFBlock and decoding GFBlock, the frequency domain amplitude self-attention module FTSA is used to reduce noise and blur components. The frequency domain amplitude self-attention module Convert to frequency domain and extract amplitude information The amplitude reflects the energy distribution of different frequency components. Blurred images usually show that low-frequency energy is dominant and high-frequency energy is attenuated. Directly processing the amplitude helps to restore high-frequency details. Apply self-attention to the magnitude matrix MM: ， ； By calculating the correlation between frequency components, the gradient features and channel features are combined with the global features, the overall amplitude distribution of the image is adjusted according to the edge distribution, the key frequencies are enhanced to suppress the noise or artifact related frequencies; the enhanced amplitude Perform inverse Fourier transform (IFFT) to obtain time domain features ,At the same time, in order to avoid the loss of phase information, the original phase is retained to ensure the structural consistency of the reconstructed image.

4. The image deblurring method of adaptive amplitude filtering and gradient enhancement according to claim 1, characterized in that: The step S3 adopts the joint use of the pixel loss function L1, the multi-scale frequency reconstruction loss function FFT-L1, the amplitude frequency domain loss function FFT-Lmag and the gradient consistency loss function Lgrad. The four loss functions integrate the perceptual quality and frequency domain characteristics of the image. The joint loss function is expressed as: ； Among them, λ1 is The pixel loss function weight, λ2 is Gradient consistency loss function weight, λ3 is The frequency domain loss function weight, λ4 is Magnitude loss function weights.

5. The image deblurring method of adaptive amplitude filtering and gradient enhancement according to claim 1, characterized in that: The step S4 specifically comprises the following steps: S4.1: Set the training parameters and use the Adam algorithm for optimization training. The initial preset parameters for training are: batch size is 16, momentum parameter is 0.9, initial learning rate is 0.001, and number of iterations is 10000. S4.2: Put the training set and validation set images in the dataset into the adaptive amplitude filtering and gradient enhancement image deblurring model for training; S4.3: Train the model according to the set parameters, and adjust the learning rate and number of iterations of the model training by observing the changing trend of the loss function until the loss function changes tend to be stable, and obtain the final training model.

6. The image deblurring method of adaptive amplitude filtering and gradient enhancement according to claim 1, characterized in that: The step S5 is based on the trained image deblurring model based on adaptive amplitude filtering and gradient enhancement, referring to Figure 6, the left picture is the original blurred image, the small pictures from the upper left to the lower right are the clear images, DeblurGAN, DeblurGANv2, MAXIM, MIMO-UNet, Restormer, NAFNet, GFTA-Net results, and the texture restoration ability and image blur artifact removal ability of the model are evaluated through qualitative comparison of visual effects.

Images