A lace texture image diffusion generation method based on ADM and dynamic denoising filter model

Abstract

This invention provides a method for generating lace texture images based on an ADM (Adaptive Demographic Model) and dynamic denoising filtering model. This model combines a learnable Gaussian blur layer with a high-pass filtering technique. The Gaussian blur layer adaptively adjusts the blur intensity based on the input image, reducing high-frequency noise in the lace texture image. The high-pass filtering technique enhances key areas such as the outline and edges of the main pattern in the image, thereby highlighting the main texture features. Adding this module to the downsampling layer of the ADM network improves the overly smooth appearance of the lace texture images generated by the original model and enhances the image's texture depth.

Description

Technical Field

[0001] This invention relates to the field of technology, and in particular to a method for generating lace texture images based on ADM and a dynamic denoising filtering model. Background Technology

[0002] Currently, lace is widely used in various types of clothing, evolving from a decorative element to an indispensable part of fashion design. As an artistic product, the appreciation of lace is highly subjective. Therefore, designing a diverse range of high-quality lace products is a worthwhile research topic. While designers invest significant time and effort in lace design, some products still fail to meet customer aesthetic preferences. To satisfy the demand for novel and aesthetically pleasing patterns, it is necessary to design more diverse and high-quality lace products, which requires leveraging new technologies and methods in the field of computer vision.

[0003] The rapid development of the field of image generation has benefited from the advancement of deep learning technology. Currently, image synthesis has made great progress in recent years, with new generation models constantly being proposed, especially diffusion models, which have greatly improved the quality and diversity of image generation, resulting in significant achievements in both academic research and practical applications.

[0004] Compared to other models, the denoising diffusion model has been proven to generate high-quality and diverse images, outperforming current state-of-the-art generative models. However, its generated images still suffer from insufficient texture features, overly smooth color representation, and inadequate texture hierarchy, leading to reduced image diversity and quality. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a lace texture image diffusion generation method based on ADM and a dynamic denoising filtering model. This model combines a learnable Gaussian blur layer with a high-pass filtering technique. The Gaussian blur layer aims to adaptively adjust the blurring intensity of the image according to different input images, reducing high-frequency noise in the lace texture image. The high-pass filtering technique enhances key areas such as the outline and edges of the main pattern in the image, thereby highlighting the main texture features of the image. Adding this module to the downsampling layer of the ADM network improves the overly smooth phenomenon in the lace texture image generated by the original model and enhances the texture hierarchy of the image.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for generating lace texture images based on ADM and a dynamic denoising filtering model, comprising the following steps:

[0007] Step S1: Create a dataset: Convert real lace objects into lace texture images; first, filter the images by cropping all lace texture images with blank edges, keeping only the patterned parts; second, fill all the processed images with blanks to make all lace images equal in length and width; finally, scale all images proportionally to 512*512 pixels to use as the training dataset for the lace generation model.

[0008] Step S2: Construct a fusion network based on dynamic denoising filtering, such as... Figure 1 As shown, based on the Unet network, i.e., the encoder-decoder structure, the first four layers of the encoder consist of a residual network ResBlock and a dynamic denoising filter module G(x), while the last three layers consist of a self-attention module AttentionBlock and a residual network. Each layer of the encoder consists of a self-attention module and a residual network. The encoder extracts feature information from the input image to obtain a feature map. The dynamic denoising filter module is located in the encoder and denoises the extracted feature map and enhances the high-frequency components in the image. The decoder corresponds to the feature map reconstruction process and obtains image information from the encoding process through skip connections to assist this process. The self-attention mechanism is located in both the encoder and decoder to obtain the global dependencies of the image and enhance high-frequency features. The goal of this network is to denoise through the forward process and generate a clean image through the reverse process.

[0009] Step S3: Construct the loss calculation during training; the training process involves learning from the noisy image and finally outputting the predicted clean image to obtain the predicted noise level; the network aims to predict noise values ​​close to the noise values ​​of the input image, and its loss function is the conventional L2 loss. The input image and the image after network denoising are compared. If the predicted noise is closer to the real noise, the lower the loss, the closer the images are, the better the learning effect, and the better the quality of the image generated by the reverse process.

[0010] Step S4: Train the fusion network constructed in step S2 using the lace texture image dataset established in step S1, using the loss calculation function in step S3 during the training process;

[0011] Step S5: After training is complete, use the trained model to input a random noise image to generate a clean lace texture image.

[0012] In a preferred embodiment: the process of establishing the dataset in step S1 includes the following steps:

[0013] Step S11: Data filtering; inspect the collected images, remove images containing text or other content that is not related to lace, and pick out lace images with blank edges on both sides to ensure that all images are clean and without blank spaces;

[0014] Step S12: Image correction and scaling; Correct the images selected in step S11 by using professional image editing software to delete the blank content on both sides of the image, ensuring that all images retain only the main body of the image, merge the unprocessed images and the processed images into a dataset, and scale all images to unify the length and width of each image.

[0015] Step S13: Image Filling and Scaling; Fill the images processed in step S12 with blanks to make the length and width of each image equal, and then scale the images proportionally to 512*512 pixels to ensure that the patterns of all images are centered and the same size.

[0016] In a preferred embodiment: the network structure of step S2 includes:

[0017] (1) Overall Unet network:

[0018] It consists of two parts: an encoder and a decoder, each with seven layers. Each layer transmits information through skip connections, forming a symmetrical structure. Each layer of the encoder consists of two residual networks followed by a downsampling pooling layer. Each layer of the decoder consists of an upsampling pooling layer followed by two residual networks. The residual networks extract feature maps, the downsampling pooling layers reduce the image to half its original size, and the downsampling pooling layers perform a 2x upsampling operation on the image. The encoder extracts image features, and the decoder uses the feature maps obtained from the decoder to reconstruct the image. Finally, the reconstructed image is output, and the output image has 6 channels. The self-attention mechanism is located in the last three layers of both the encoder and decoder, performing attention calculations on the extracted high-level semantic features to obtain global dependencies and enhance the high-frequency features of the image. The dynamic denoising and filtering module is the first four layers of the main encoder, connected after each residual network, to enhance texture features.

[0019] (2) Dynamic noise reduction and filtering module:

[0020] It includes a learnable Gaussian blur layer, convolutional layers, and a Sigmoid activation function. The learnable Gaussian blur layer consists of two sequential 3x3 convolutions and a ReLU activation function. Each layer is a two-dimensional grouped convolution with a 3x3 kernel size, padding of 1, and the same number of input and output channels. The ReLU activation function is applied after each convolutional layer to ensure non-linear expressiveness. Each convolutional layer is a 3x3 convolution. The convolution operation is a grouped convolution, with each channel processed independently. The blurred image is generated by performing two convolutions on the input image and applying the ReLU activation function. The Sigmoid activation function is used to perform a non-linear mapping on the output, compressing the output values ​​to the [0,1] range and generating filter weights.

[0021] In a preferred embodiment: the loss calculation process in step S3 comprises two parts:

[0022] (1) The formula for calculating forward noise is as follows:

[0023]

[0024] in α t =1-β t x t This represents the desired image with added noise, x. t-1 Represents an image with added noise; ∈ represents the real image with added noise, t is the training time step, and β t Given the variance scheduling for any time step t, the final noise addition formula is as follows:

[0025]

[0026] (2) Loss function:

[0027] The ultimate goal of generative models is to enable neural networks to learn to generate clean images. This means the reverse process requires the neural network to learn to denoise images with added noise. The formula for the reverse process is as follows:

[0028]

[0029] Where μ θ and β θ These are all neural network predictions with parameter θ;

[0030] During the training phase, the variance β of the forward process t It can be learned through reparameterization or kept constant as a hyperparameter, typically with β fixed. θ constant p θ (x t-1| x t By comparing the forward and posterior perspectives, the training objective for the forward process is obtained as follows:

[0031]

[0032] in Let V be the variance of the posterior distribution. μ is the mean of the posterior distribution; however, during the training phase, μ is not directly used. θ and In comparison, since the goal in the generation stage is to restore a clean image from a noisy image, the predicted noise ∈ t Better; therefore, the loss function is compared as the ratio of the loss between the actual noise and the predicted noise, and the mean square error is calculated. The resulting loss function is as follows:

[0033]

[0034] Where ∈ θ The noise is used for neural network prediction; simultaneously, variance is learned during training, and a penalty term L is added to the loss term. vlb The final loss function L is obtained. hybrid ;

[0035]

[0036] L hybrid =L simple +λL vlb

[0037] Where v is the vector output by the model, Σ θ( x t ,t ) This represents the new variance.

[0038] In a preferred embodiment, the training process in step S4 is as follows:

[0039] Step S41: Before each round of training, input a batch of data into the current diffusion model network, randomly select a time step t from [1,2000], and add noise that conforms to a normal distribution to the batch of images;

[0040] Step S41: Input the noisy data from S41 into the neural network for training, perform a denoising process of feature extraction and restoration on the image, predict the noise, and obtain the denoised image after modeling.

[0041] Step S42: Compare the predicted noise and the real noise, that is, compare the L2 loss of the real image and the predicted image, and calculate the loss of the current batch of data. The calculation method is shown in step S3.

[0042] Step S43: Backpropagation of gradients, using the Adam optimizer to optimize network parameters at a learning rate of 1e-5.

[0043] In a preferred embodiment, the determination step in step S5 is as follows:

[0044] Step S51: Randomly sample a noisy image and input the number of steps to restore the image;

[0045] Step S52: Input the data into the neural network and perform noise prediction in the reverse process to obtain the denoised image; repeat the above steps until the number of steps is used up.

[0046] Step S53: Finally, output a denoised 512*512 pixel color lace texture image.

[0047] Compared with existing technologies, this invention has the following advantages: This invention proposes a learnable dynamic denoising filtering module that combines Gaussian blur and high-pass filtering techniques. The Gaussian blur can be seen as a preprocessing step for the input image, making it smoother and reducing high-frequency noise; the high-pass filtering enhances the edge and detail features of the image by calculating the difference between the input image and the Gaussian-blurred image, thus improving the model's feature extraction performance. Adding this module to the diffusion model results in an image with clear edges and rich details, while also reducing the problems of overly smooth or overly dense patterns in the image. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the model architecture of a preferred embodiment of the present invention;

[0049] Figure 2 This is a flowchart of the model operation of a preferred embodiment of the present invention;

[0050] Figure 3 This is a flowchart of the model generation process according to a preferred embodiment of the present invention. Detailed Implementation

[0051] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0052] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0053] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations according to this application; as used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise; furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0054] A method for generating lace texture images based on ADM and a dynamic denoising filtering model, referenced Figure 1-3 This includes the following steps:

[0055] Step S1: Create a dataset: Convert real lace objects into lace texture images. First, filter the images by cropping all lace texture images with blank edges, retaining only the patterned parts; second, fill all the processed images with blanks to make all lace images equal in length and width; finally, scale all images proportionally to 512*512 pixels to use as the training dataset for the lace generation model.

[0056] Step S2: Construct a fusion network based on dynamic denoising filtering, such as... Figure 1 As shown, based on the Unet network (encoder-decoder structure), the first four layers of the encoder consist of a residual network (ResBlock) and a dynamic denoising filter module (G(x)), while the last three layers consist of a self-attention module (AttentionBlock) and a residual network. Each layer of the encoder consists of a self-attention module and a residual network. The encoder extracts feature information from the input image to obtain a feature map. The dynamic denoising filter module, located in the encoder, denoises the extracted feature map and enhances the high-frequency components in the image. The decoder corresponds to the feature map reconstruction process, obtaining image information from the encoding process through skip connections to assist this process. The self-attention mechanism is located in both the encoder and decoder, acquiring global dependencies in the image and enhancing high-frequency features. The network aims to denoise through the forward process and generate a clean image through the reverse process.

[0057] Step S3: Calculate the loss during training. The training process involves learning from the noisy image and finally outputting a predicted clean image, thus obtaining the predicted noise level. The network aims to predict noise values ​​close to the noise values ​​of the input image. Its loss function is a standard L2 loss. The input image and the network-denoised image are compared. The closer the predicted noise is to the actual noise, the lower the loss. The closer the images are, the better the learning effect, and the better the quality of the image generated by the reverse process.

[0058] Step S4: Train the fusion network constructed in step S2 using the lace texture image dataset established in step S1, using the loss calculation function in step S3 during the training process.

[0059] Step S5: After training is complete, use the trained model to input a random noise image to generate a clean lace texture image.

[0060] The process of establishing the dataset in step S1 includes the following steps:

[0061] Step S11: Data Filtering. Inspect the collected images, remove images containing text or other content not related to lace, and pick out images with blank edges on both sides of the lace image to ensure that all images are clean and without blank spaces.

[0062] Step S12: Image Correction and Scaling. Correct the images selected in Step S11 by using professional image editing software to delete the blank content on both sides of the images, ensuring that only the main body of the image is retained. Merge the unprocessed and processed images into a dataset, and scale all images to unify the length and width of each image.

[0063] Step S13: Image Filling and Scaling. To reduce computational load and meet the input size requirements of the backbone convolutional neural network, while maintaining the aspect ratio of the pattern, the images processed in S12 are filled with blanks to make each image equal in length and width. Then, the images are scaled proportionally to 512*512 pixels to ensure that the pattern in all images is centered and the same size.

[0064] The network structure in step S2 includes:

[0065] (1) Overall Unet network:

[0066] The system consists of two parts: an encoder and a decoder, each with seven layers. Information is passed between layers via skip connections, forming a symmetrical structure. Each layer of the encoder consists of two residual networks followed by a downsampling pooling layer. Each layer of the decoder consists of an upsampling pooling layer followed by two residual networks. The residual networks extract feature maps, while the downsampling pooling layers reduce the image to half its original size and then enlarge it by a factor of two. The encoder extracts image features, and the decoder uses the feature maps obtained from the decoder to reconstruct the image. The final output is the reconstructed image, which has six channels. A self-attention mechanism is located in the last three layers of both the encoder and decoder. It performs attention calculations on the extracted high-level semantic features, obtaining global dependencies and enhancing the high-frequency features of the image. The dynamic denoising and filtering module consists of the first four layers of the main encoder, connected after each residual network, enhancing texture features.

[0067] (2) Dynamic noise reduction and filtering module:

[0068] It includes a learnable Gaussian blur layer, convolutional layers, and a Sigmoid activation function. The learnable Gaussian blur layer consists of two sequential 3x3 convolutions and a ReLU activation function. Each layer is a two-dimensional grouped convolution (each channel is processed independently), with a 3x3 kernel size and padding of 1. The number of input and output channels is the same. The ReLU activation function is applied after each convolutional layer to ensure non-linear expressiveness. Each convolutional layer is a 3x3 convolution. The convolution operation is grouped convolution, with each channel processed independently. By performing two convolutions on the input image and applying the ReLU activation function, a blurred image is generated. The Sigmoid activation function is used to non-linearly map the output, compressing the output values ​​to the [0,1] range to generate filter weights.

[0069] The loss calculation process in step S3 consists of two parts:

[0070] (1) The formula for calculating forward noise is as follows:

[0071]

[0072] in α t =1-β t x t This represents the desired image with added noise, x. t-1 This represents an image with added noise. ∈ represents the ground truth with added noise, t is the training time step, and β... t Let the variance be regulated for any time step t. Therefore, the final noise addition formula is as follows:

[0073]

[0074] (2) Loss function:

[0075] The ultimate goal of generative models is to enable neural networks to learn to generate clean images. This means the reverse process requires the neural network to learn to denoise images with added noise. The formula for the reverse process is as follows:

[0076]

[0077] Where μ θ and β θ These are all neural network predictions with parameter θ.

[0078] During the training phase, the variance β of the forward process t It can be learned through reparameterization or kept constant as a hyperparameter, typically with β fixed. θ constant p θ (x t-1| x t By comparing the forward and posterior perspectives, the training objective for the forward process is obtained as follows:

[0079]

[0080] in Let V be the variance of the posterior distribution. Let μ be the mean of the posterior distribution. However, during the training phase, μ is not directly used... θ and In comparison, since the goal in the generation stage is to restore a clean image from a noisy image, the predicted noise ∈ t Better. Therefore, the loss function is compared by the ratio of the loss between the actual noise and the predicted noise, and the mean square error is calculated. The resulting loss function is as follows:

[0081]

[0082] Where ∈ θ This represents the noise predicted by the neural network. Simultaneously, variance is learned during training, and a penalty term L is added to the loss term. vlb The final loss function L is obtained.hybrid .

[0083]

[0084] L hybrid =L simple +λL vlb

[0085] Where v is the vector output by the model, Σ θ (x t ,t) represents the new variance.

[0086] The training process in step S4 is as follows:

[0087] Step S41: Before each round of training begins, input a batch of data into the current diffusion model network, randomly select a time step t from [1,2000], and add noise that conforms to a normal distribution to the batch of images.

[0088] Step S41: Input the noisy data from S41 into the neural network for training, perform a denoising process of feature extraction and restoration on the image, predict the noise, and obtain the denoised image of the model.

[0089] Step S42: Compare the predicted noise and the real noise, that is, compare the L2 loss of the real image and the predicted image, and calculate the loss of the current batch of data. The calculation method is shown in step S3.

[0090] Step S43: Backpropagation of gradients, using the Adam optimizer to optimize network parameters at a learning rate of 1e-5.

[0091] The determination step in step S5 is as follows:

[0092] Step S51: Randomly sample a noisy image and input the number of steps to restore the image.

[0093] Step S52: Input the data into the neural network and perform noise prediction in the reverse process to obtain the denoised image. Repeat the above steps until all steps are used up.

[0094] Step S53: Finally, output a denoised 512*512 pixel color lace texture image.

Claims

1. A method for generating lace texture images based on ADM and a dynamic denoising filtering model, characterized in that: Includes the following steps: Step S1: Create a dataset: Convert real lace objects into lace texture images; first, filter the images by cropping all lace texture images with blank edges, keeping only the patterned parts; second, fill all the processed images with blanks to make all lace images equal in length and width; finally, scale all images proportionally to 512*512 pixels to use as the training dataset for the lace generation model. Step S2: Construct a fusion network based on dynamic denoising filtering, using the Unet network, i.e., an encoder-decoder structure. The first four layers of the encoder consist of a residual network ResBlock and a dynamic denoising filtering module G(x), while the last three layers consist of a self-attention module AttentionBlock and a residual network. The encoder extracts feature information from the input image to obtain a feature map. The dynamic denoising filtering module is located in the encoder, denoising the extracted feature map and enhancing high-frequency components in the image. The decoder corresponds to the feature map restoration process, obtaining image information from the encoding process through skip connections to assist this process. The self-attention mechanism is located in both the encoder and decoder, acquiring global dependencies in the image and enhancing high-frequency features. The network aims to denoise by learning the forward process and generate a clean image in the reverse process. Step S3: Construct the loss calculation during training; the training process involves learning from the noisy image and finally outputting the predicted clean image to obtain the predicted noise level; the network aims to predict noise values ​​close to the noise values ​​of the input image, and its loss function is L2 loss. The input image and the image after network denoising are compared. If the predicted noise is closer to the real noise, the lower the loss, the closer the images are, the better the learning effect, and the better the quality of the image generated by the reverse process. Step S4: Train the fusion network constructed in step S2 using the lace texture image dataset established in step S1, using the loss calculation function in step S3 during the training process; Step S5: After training is complete, use the trained model to input a random noise image to generate a clean lace texture image; The fusion network based on dynamic denoising filtering in step S2 includes: (1) Overall Unet network: It consists of two parts: an encoder and a decoder, each with seven layers. Each layer transmits information through skip connections, forming a symmetrical structure. Each layer of the encoder includes two residual networks followed by a downsampling pooling layer. Each layer of the decoder includes an upsampling pooling layer followed by two residual networks. The residual networks extract feature maps, the downsampling pooling layers reduce the image to half its original size, and the upsampling pooling layers magnify the image by a factor of 2. The encoder extracts image features, and the decoder uses the feature maps obtained from the encoder to reconstruct the image. Finally, the reconstructed image is output, which has 6 channels. A self-attention mechanism performs attention calculations on the extracted high-level semantic features to obtain global dependencies and enhance the high-frequency features of the image. The dynamic denoising and filtering module consists of the first four layers of the main encoder, connected after each residual network, to enhance texture features. (2) Dynamic noise reduction filtering module: It includes a learnable Gaussian blur layer, convolutional layers, and a Sigmoid activation function. The learnable Gaussian blur layer consists of two sequential 3x3 convolutions and a ReLU activation function. Each layer is a two-dimensional grouped convolution with a kernel size of 3x3 and padding of 1. The number of input and output channels is the same. The ReLU activation function is applied after each convolutional layer to ensure non-linear expressiveness. The convolutional layer is a 3x3 convolution. The convolution operation is a grouped convolution, with each channel processed independently. The blurred image is generated by performing two convolutions on the input image and using the ReLU activation function. The Sigmoid activation function is used to perform a non-linear mapping on the output, compressing the output value to the range of [0, 1] to generate filter weights.

2. The method for generating lace texture images based on ADM and a dynamic denoising filtering model according to claim 1, characterized in that: The process of establishing the dataset in step S1 includes the following steps: Step S11: Data filtering; inspect the collected images, remove images containing text or other content that is not related to lace, and pick out lace images with blank edges on both sides to ensure that all images are clean and without blank spaces; Step S12: Image correction and scaling; Correct the images selected in step S11 by using professional image editing software to delete the blank content on both sides of the image, ensuring that all images retain only the main body of the image, merge the unprocessed images and the processed images into a dataset, and scale all images to unify the length and width of each image. Step S13: Image Filling and Scaling; Fill the images processed in step S12 with blanks to make the length and width of each image equal, and then scale the images proportionally to 512*512 pixels to ensure that the patterns of all images are centered and the same size.

3. The method for generating lace texture images based on ADM and a dynamic denoising filtering model according to claim 1, characterized in that: The loss calculation process in step S3 consists of two parts: (1) The formula for calculating forward noise is as follows: in , , This represents the desired image with added noise. This represents an image with added noise. To add noise to the real training, where t is the training time step. Given the variance scheduling for any time step t, the final noise addition formula is as follows: (2) Loss function: The ultimate goal of generative models is to enable neural networks to learn to generate clean images. This means the reverse process requires the neural network to learn to denoise images with added noise. The formula for the reverse process is as follows: in and All parameters are The neural network prediction value; During the training phase, the variance of the forward process It can be learned through reparameterization or kept constant as a hyperparameter, usually fixed. constant ,Will Comparing with the forward and posterior perspectives, the training objective for the forward process is as follows: in Let V be the variance of the posterior distribution. It is the mean of the posterior distribution; however, it is not directly used during the training phase. and The comparison is made because the goal in the generation stage is to restore a clean image from a noisy image; therefore, the prediction of noise is crucial. Better; therefore, the loss function is compared as the ratio of the loss between the actual noise and the predicted noise, and the mean square error is calculated. The resulting loss function is as follows: in The noise is used to predict the neural network; simultaneously, variance is learned during training, and a penalty term is added to the loss term. The final loss function is obtained. ; Where v is the vector output by the model, This represents the new variance.

4. The method for generating lace texture images based on ADM and a dynamic denoising filtering model according to claim 1, characterized in that: The training process in step S4 is as follows: Step S41: Before each round of training begins, input a batch of data into the current diffusion model network, randomly select a time step t from [1,2000], and add noise that conforms to a normal distribution to the batch of images; Step S42: Input the noisy data from S41 into the neural network for training, perform a denoising process of feature extraction and restoration on the image, predict the noise, and obtain the denoised image after modeling; Step S43: Compare the predicted noise and the real noise, that is, compare the L2 loss of the real image and the predicted image, and calculate the loss of the current batch of data. The calculation method is shown in step S3. Step S44: Backpropagation of gradients, using the Adam optimizer to optimize network parameters at a learning rate of 1e-5.

5. The method for generating lace texture images based on ADM and a dynamic denoising filtering model according to claim 1, characterized in that: The determination step in step S5 is as follows: Step S51: Randomly sample a noisy image and input the number of steps to restore the image; Step S52: Input the data into the neural network and perform noise prediction in the reverse process to obtain the denoised image; repeat the above steps until the number of steps is used up. Step S53: Finally, output a denoised 512*512 pixel color lace texture image.

Images