Lace texture image diffusion generation method based on ADM and dynamic denoising filtering model

Abstract

The invention provides a lace texture image diffusion generation method based on an ADM and a dynamic denoising filtering model, the model combines a learnable Gaussian blur layer and a high-pass filtering technology, and the Gaussian blur layer aims at adaptively adjusting the blur degree of the image according to different input images and reducing high-frequency noise in the lace texture image. According to the high-pass filtering technology, key areas such as the contour and the edge of a main pattern in the image are enhanced, so that main texture features of the image are highlighted. The module is added into a down-sampling layer of an ADM network, the phenomenon that lace texture images generated by an original model are too smooth is improved, and the texture layering sense of the images is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field, and in particular to a lace texture image diffusion generation method based on ADM and a dynamic denoising filter model. Background Art

[0002] Currently, lace is widely used in all kinds of clothing, from being a decoration to becoming an indispensable element in clothing design. As a product of artistic color, people's appreciation of lace is very subjective. Therefore, how to design sufficiently diverse and high-quality lace products is a problem worth studying. At present, human designers spend a lot of time and energy on lace design, but there are still products that do not meet the aesthetic taste of customers. In order to meet people's demand for novel and beautiful patterns, it is necessary to design more diverse and high-quality lace products, which requires the use of new technologies and methods in the current field of computer vision.

[0003] The rapid development of the field of image generation has benefited from the progress of deep learning technology. At present, image synthesis has made great progress in recent years. New generation models are constantly being proposed, especially diffusion models, which greatly improve the quality and diversity of image generation, making this field have achieved remarkable achievements in both academic research and practical applications.

[0004] Compared with other models, the denoising diffusion model has been shown to generate high-quality and diverse images, and outperforms the current state-of-the-art generative models. However, the generated images still have problems such as insufficient texture features, overly smooth color representation, and insufficient texture hierarchy, which reduce the diversity and quality of the images. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide a lace texture image diffusion generation method based on ADM and dynamic denoising filter model, which combines a learnable Gaussian blur layer with high-pass filtering technology. The Gaussian blur layer is designed to adaptively adjust the blur intensity of the image according to the input image, thereby reducing the high-frequency noise in the lace texture image. The high-pass filtering technology highlights the main texture features of the image by enhancing the key areas such as the contour and edge of the main pattern in the image. Adding this module to the downsampling layer of the ADM network improves the over-smoothing phenomenon in the lace texture image generated by the original model and enhances the texture layering of the image.

[0006] To achieve the above object, the present invention adopts the following technical solution: a lace texture image diffusion generation method based on ADM and dynamic denoising filter model, comprising the following steps:

[0007] Step S1: Establishing a data set: by converting lace objects into lace texture images; firstly, performing image screening, cropping all lace texture images with blank edges, and retaining the patterned parts; secondly, filling all processed images with blanks to make the length and width of all lace images equal; finally, scaling all images to 512*512 pixels in equal proportions, and using them as training data sets for the lace generation model;

[0008] Step S2: Construct a fusion network based on dynamic denoising filtering, such as Figure 1 As shown in the figure, based on the Unet network, i.e., the encoder-decoder structure, the first four layers of the encoder are composed of a residual network ResBlock and a dynamic denoising filter module G(x), and the last three layers are composed of a self-attention module AttentionBlock and a residual network. Each layer of the encoder is composed of a self-attention module and a residual network; the encoder extracts the feature information of the input image to obtain a feature map; the dynamic denoising filter module is located in the encoder, denoising the extracted feature map and enhancing the high-frequency components in the image; the decoder corresponds to the feature map restoration process, and obtains the image information of the encoding process through jump connections to assist the process; the self-attention mechanism is located in the encoder and decoder respectively, obtaining the global dependency of the image and enhancing the high-frequency features; the network goal is to denoise by learning the forward process and generate a clean image in the reverse process;

[0009] Step S3: construct the loss calculation in the training process; the training process is to learn the noisy image and output the predicted clean image at the end to obtain the predicted noise amount; the goal of the network is to predict the noise value close to the noise value 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 loss is lower, the closer the image is, the better the learning effect is, and finally the image quality generated by the reverse process is better;

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

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

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

[0013] Step S11: data screening; checking the collected images, removing the images containing text or other images that do not belong to lace content, and picking out the lace images with blank edges on both sides of the images, ensuring that all images are clean and have no blank spaces;

[0014] Step S12: image correction and scaling; the image selected in step S11 is corrected, and the blank content left on both sides of the image is deleted using professional image editing software to ensure that only the main body of the image is retained in all images, and the unprocessed image and the processed image are merged into one data set, and all images are scaled 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) Unet overall network:

[0018] It consists of two parts, namely the encoder and the decoder, each with seven layers, each layer transmits information through jump connections, and has 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 network extracts feature maps, the downsampling pooling layer reduces the image to half of the original, and the downsampling pooling layer magnifies the image by 2 times; the encoder is used to extract image features, and the decoder restores the image through the feature map obtained by the decoder; finally, the restored image result is output, and the output image is 6 channels; the self-attention mechanism is located in the last three layers of the encoder and decoder, and the attention calculation is performed on the extracted high-level semantic features, the global dependency is obtained, and the high-frequency features of the image are enhanced; the dynamic denoising filter module is the first four layers of the main encoder, connected after each residual network, to enhance the texture features;

[0019] (2) Dynamic denoising filter module:

[0020] It includes a learnable Gaussian blur layer, a convolution layer, and a Sigmoid activation function; the learnable Gaussian blur layer consists of two serialized 3x3 convolutions and a ReLU activation function. Each layer is a two-dimensional grouped convolution with a convolution kernel size of 3x3, a padding of 1, and the same number of input and output channels; the ReLU activation function is after each convolution layer to ensure nonlinear expression capabilities; the convolution layer is a 3x3 convolution; the convolution operation is a grouped convolution, and each channel is 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 nonlinear mapping on the output, compressing the output value to the [0,1] range, and generating filter weights.

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

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

[0023]

[0024] in α t =1-β t , x t represents the desired noise-added image, x t-1 Represents a noisy image; ∈ is the real noise, t is the training time step, β t is the variance schedule of any time step t; so the final noise addition formula is as follows:

[0025]

[0026] (2) Loss function:

[0027] The ultimate goal of the generative model is to enable the neural network to learn to generate clean images. That is, the reverse process requires the neural network to learn to denoise the noisy images. The formula of the reverse process is as follows:

[0028]

[0029] where μ θ and β θ They are all predicted values ​​of the neural network with parameter θ;

[0030] During the training phase, the variance of the forward process β t Can be learned by reparameterization or kept constant as a hyperparameter, usually with β fixed θ is a constant The p θ (x t-1| x t ) is compared with the forward posterior, and the training objective of the forward process is as follows:

[0031]

[0032] in is the variance of the posterior distribution, is the mean of the posterior distribution; however, μ is not directly θ and Comparison, because in the generation stage the goal is to restore the clean image from the noisy image, so the predicted noise ∈ t Better; therefore, the comparison of the loss function is the ratio of the loss between the real noise and the predicted noise, and the mean square error is calculated, and the loss function obtained is as follows:

[0033]

[0034] where ∈ θ is the noise predicted by the neural network; at the same time, the variance learning is added during the training process, and the loss term is added with a penalty term L vlb , and get the final loss function L hybrid ;

[0035]

[0036] L hybrid =L simple +λL vlb

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

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

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

[0040] Step S41: input the noisy data in S41 into the neural network training, perform a denoising process of feature extraction and restoration on the image, predict the noise and obtain the image after model denoising;

[0041] Step S42: Compare the predicted noise with 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: Gradient back propagation, using Adam optimizer to optimize network parameters with a learning rate of 1e-5.

[0043] In a preferred embodiment: the determination steps in step S5 are as follows:

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

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

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

[0047] Compared with the prior art, the present invention has the following beneficial effects: the present invention proposes a learnable dynamic denoising filter module, which combines the two technologies of Gaussian blur and high-pass filtering. The Gaussian blur part can be regarded as preprocessing the input image to make it smoother and reduce high-frequency noise; the high-pass filter part strengthens the edge and detail features of the image by calculating the difference between the input image and the image after Gaussian blur, which not only improves the performance of the model in feature extraction. By adding this module to the diffusion model, the final generated image not only has clear edges and rich details, but also reduces the problem of the image being too smooth or the pattern being too dense. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0050] Figure 3 It is a flow chart of model generation in a preferred embodiment of the present invention. DETAILED DESCRIPTION

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

[0052] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0053] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application; as used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or their combinations.

[0054] A diffusion generation method for lace texture images based on ADM and dynamic denoising filter model, reference Figure 1-3 , including the following steps:

[0055] Step S1: Establishing a data set: by converting lace objects into lace texture images. First, perform image screening, crop all lace texture images with blank edges, and retain the patterned parts; second, fill all processed images with blanks to make the length and width of all lace images equal; finally, scale all images to 512*512 pixels in equal proportions, and use them as the training data set for the lace generation model.

[0056] Step S2: Construct a fusion network based on dynamic denoising filtering, such as Figure 1 As shown in the figure, based on the Unet network (encoder-decoder structure), the first four layers of the encoder are composed of a residual network (ResBlock) and a dynamic denoising filter module (G(x)), and the last three layers are composed of a self-attention module (AttentionBlock) and a residual network. Each layer of the encoder is composed of a self-attention module and a residual network. The encoder extracts the feature information of the input image to obtain a feature map; the dynamic denoising filter module is located in the encoder, denoising the extracted feature map and enhancing the high-frequency components in the image; the decoder corresponds to the feature map restoration process, and obtains the image information of the encoding process through jump connections to assist the process. The self-attention mechanism is located in the encoder and decoder respectively to obtain the global dependency of the image and enhance the high-frequency features. The network goal is to denoise by learning the forward process and generate a clean image in the reverse process.

[0057] Step S3: Construct the loss calculation in the training process. The training process is to learn the noisy image and output the predicted clean image at the end to obtain the predicted noise amount. The goal of the network is to predict the noise value close to the noise value of the input image. 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 loss is lower, the closer the image is, the better the learning effect is, and the better the image quality generated by the reverse process is.

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

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

[0060] The process of establishing a data set in step S1 includes the following steps:

[0061] Step S11: Data screening. Check the collected images, remove the images containing text or other non-lace content, and pick out the lace images with blank edges on both sides of the image to ensure that all images are clean and have no blanks.

[0062] Step S12: Image correction and scaling. Correct the image selected in step S11, use professional photo editing software to delete the blank content on both sides of the image, ensure that only the main body of the image is retained in all images, merge the unprocessed image and the processed image into one data set, and scale all images to unify the length and width of each image.

[0063] Step S13: Image filling and scaling. In order to reduce the amount of computation and meet the input size required by the backbone convolutional neural network, while keeping the aspect ratio of the pattern unchanged, the images processed in S12 are filled with blanks to make the length and width of each image equal, and then the images are scaled proportionally to 512*512 pixels to ensure that the patterns of all images are centered and the same size.

[0064] The network structure of step S2 includes:

[0065] (1) Unet overall network:

[0066] It consists of two parts, namely the encoder and the decoder, each with seven layers. Each layer transmits information through jump connections and has 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 network extracts feature maps, the downsampling pooling layer reduces the image to half of the original, and the downsampling pooling layer magnifies the image by 2 times. The encoder is used to extract image features, and the decoder restores the image through the feature map obtained by the decoder. Finally, the restored image result is output, and the output image is 6 channels. The self-attention mechanism is located in the last three layers of the encoder and decoder, respectively, and performs attention calculation on the extracted high-level semantic features, obtains global dependencies, and enhances the high-frequency features of the image. The dynamic denoising filter module is the first four layers of the main encoder, connected after each residual network, to enhance texture features.

[0067] (2) Dynamic denoising filter module:

[0068] Includes learnable Gaussian blur layer, convolution layer and Sigmoid activation function. The learnable Gaussian blur layer consists of two serialized 3x3 convolutions and ReLU activation function. Each layer is a 2D grouped convolution (each channel is processed independently), the convolution kernel size is 3x3, the padding is 1, and the number of input and output channels is the same. The ReLU activation function is after each convolution layer to ensure nonlinear expression ability. The convolution layer is a 3x3 convolution. The convolution operation is a grouped convolution, and each channel is processed independently. The blurred image is generated by convolving the input image twice and using the ReLU activation function. The Sigmoid activation function is used to perform nonlinear mapping on the output, compressing the output value to the [0,1] range, and generating filter weights.

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

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

[0071]

[0072] in α t =1-β t , x t represents the desired noise-added image, x t-1 Represents a noisy image. ∈ is the real noise, t is the training time step, β t is the variance schedule of any time step t. So the final noise addition formula is as follows:

[0073]

[0074] (2) Loss function:

[0075] The ultimate goal of the generative model is to enable the neural network to learn to generate clean images. That is, the reverse process requires the neural network to learn to denoise the noisy images. The formula of the reverse process is as follows:

[0076]

[0077] where μ θ and β θ They are all predicted values ​​of the neural network with parameter θ.

[0078] During the training phase, the variance of the forward process β t Can be learned by reparameterization or kept constant as a hyperparameter, usually with β fixed θ is a constant The p θ (x t-1| x t ) is compared with the forward posterior, and the training objective of the forward process is as follows:

[0079]

[0080] in is the variance of the posterior distribution, is the mean of the posterior distribution. However, in the training phase, μ is not directly θ and Comparison, because in the generation stage the goal is to restore the clean image from the noisy image, so the predicted noise ∈ t Therefore, the loss function is the ratio of the loss between the real noise and the predicted noise, and the mean square error is calculated, and the loss function is as follows:

[0081]

[0082] where ∈ θ is the noise predicted by the neural network. At the same time, the variance learning is added during the training process, and the loss term is added with a penalty term L vlb , and get the final loss function Lhybrid .

[0083]

[0084] L hybrid =L simple +λL vlb

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

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

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

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

[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: Gradient back propagation, using Adam optimizer to optimize network parameters with a learning rate of 1e-5.

[0091] The determination steps in step S5 are as follows:

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

[0093] Step S52: Input to the neural network, perform noise prediction in the reverse process, and obtain the denoised image. Repeat the above steps until the number of steps is exhausted.

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

Claims

1. A lace texture image diffusion generation method based on ADM and dynamic denoising filter model, characterized by: The following steps are involved: Step S1: Establishing a data set: by converting lace objects into lace texture images; firstly, performing image screening, cropping all lace texture images with blank edges, and retaining the patterned parts; secondly, filling all processed images with blanks to make the length and width of all lace images equal; finally, scaling all images to 512*512 pixels in equal proportions, and using them as training data sets for the lace generation model; Step S2: Construct a fusion network based on dynamic denoising filtering, as shown in Figure 1. Based on the Unet network, that is, the encoder-decoder structure, the first four layers of the encoder are composed of a residual network ResBlock and a dynamic denoising filtering module G(x), and the last three layers are composed of a self-attention module AttentionBlock and a residual network. Each layer of the encoder is composed of a self-attention module and a residual network; the encoder extracts feature information of 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 the high-frequency components in the image; the decoder corresponds to the feature map restoration process, and obtains the image information of the encoding process through jump connections to assist the process; the self-attention mechanism is located in the encoder and the decoder respectively, to obtain the global dependency of the image and enhance the high-frequency features; the network goal is to denoise by learning the forward process and generate a clean image in the reverse process; Step S3: construct the loss calculation in the training process; the training process is to learn the noisy image and output the predicted clean image at the end to obtain the predicted noise amount; the goal of the network is to predict the noise value close to the noise value 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 loss is lower, the closer the image is, the better the learning effect is, and finally the image quality generated by the reverse process is better; Step S4: using the lace texture image dataset established in step S1 to train the fusion network constructed in step S2, and using the loss calculation function in step S3 during the training process; Step S5: After the training is completed, the trained model is used to input a random noise image to generate a clean lace texture image.

2. According to claim 1, a lace texture image diffusion generation method based on ADM and dynamic denoising filter model is characterized by: The process of establishing a data set in step S1 includes the following steps: Step S11: data screening; checking the collected images, removing the images containing text or other images that do not belong to lace content, and picking out the lace images with blank edges on both sides of the images, ensuring that all images are clean and have no blank spaces; Step S12: image correction and scaling; the image selected in step S11 is corrected, and the blank content left on both sides of the image is deleted using professional image editing software to ensure that only the main body of the image is retained in all images, and the unprocessed image and the processed image are merged into one data set, and all images are scaled 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 image diffusion based on ADM and dynamic denoising filter model according to claim 1, characterized in that: The network structure of step S2 includes: (1) Unet overall network: It consists of two parts, namely the encoder and the decoder, each with seven layers, each layer transmits information through jump connections, and has 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 network extracts feature maps, the downsampling pooling layer reduces the image to half of the original, and the downsampling pooling layer magnifies the image by 2 times; the encoder is used to extract image features, and the decoder restores the image through the feature map obtained by the decoder; finally, the restored image result is output, and the output image is 6 channels; the self-attention mechanism is located in the last three layers of the encoder and decoder, and the attention calculation is performed on the extracted high-level semantic features, the global dependency is obtained, and the high-frequency features of the image are enhanced; the dynamic denoising filter module is the first four layers of the main encoder, connected after each residual network, to enhance the texture features; (2) Dynamic denoising filter module: It includes a learnable Gaussian blur layer, a convolution layer, and a Sigmoid activation function; the learnable Gaussian blur layer consists of two serialized 3x3 convolutions and a ReLU activation function. Each layer is a two-dimensional grouped convolution with a convolution kernel size of 3x3, a padding of 1, and the same number of input and output channels; the ReLU activation function is after each convolution layer to ensure nonlinear expression capabilities; the convolution layer is a 3x3 convolution; the convolution operation is a grouped convolution, and each channel is 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 nonlinear mapping on the output, compressing the output value to the [0,1] range, and generating filter weights.

4. The method for generating lace texture image diffusion based on ADM and dynamic denoising filter model according to claim 1, characterized in that: The loss calculation process of step S3 includes two parts: (1) The calculation formula for adding noise forward is as follows: in α t =1-β t , x t represents the desired noise-added image, x t-1 Represents a noisy image; ∈ is the real noise, t is the training time step, β t is the variance schedule of any time step t; so the final noise addition formula is as follows: (2) Loss function: The ultimate goal of the generative model is to enable the neural network to learn to generate clean images. That is, the reverse process requires the neural network to learn to denoise the noisy images. The formula of the reverse process is as follows: where μ θ and β θ They are all predicted values ​​of the neural network with parameter θ; During the training phase, the variance of the forward process β t Can be learned by reparameterization or kept constant as a hyperparameter, usually with β fixed θ is a constant The p θ( x t-1 | x t) Compared with the forward and posterior, the training objectives of the forward process are as follows: in is the variance of the posterior distribution, is the mean of the posterior distribution; however, μ is not directly θ and Comparison, because in the generation stage the goal is to restore the clean image from the noisy image, so the predicted noise ∈ t Better; therefore, the comparison of the loss function is the ratio of the loss between the real noise and the predicted noise, and the mean square error is calculated, and the loss function obtained is as follows: where ∈ θ is the noise predicted by the neural network; at the same time, the variance learning is added during the training process, and the loss term is added with a penalty term L vlb , and get the final loss function L hybrid ; THE hybrid =L simple +λL vlb Where v is the vector output by the model, Σ θ (x t ,t) is the new variance.

5. The method for generating lace texture image diffusion based on ADM and dynamic denoising filter model according to claim 1, characterized in that: The training process of step S4 is as follows: Step S41: before each round of training, a batch of data is input into the current diffusion model network, a time step t is randomly selected from [1,2000], and noise that conforms to the normal distribution is added to the batch of images; Step S41: input the noisy data in S41 into the neural network training, perform a denoising process of feature extraction and restoration on the image, predict the noise and obtain the image after model denoising; Step S42: Compare the predicted noise with 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 S43: Gradient back propagation, using Adam optimizer to optimize network parameters with a learning rate of 1e-5.

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

Images