High-noise side-scan sonar image super-resolution processing method and device based on diffusion model

Abstract

The invention relates to a high-noise side-scan sonar image super-resolution processing technology and device based on a diffusion model, and the method comprises the steps: S1, constructing a sonar image super-resolution processing model based on the diffusion model, and the sonar image super-resolution processing model comprises an encoder and a decoder; the method comprises the following steps: carrying out unified size adjustment, normalization processing and specific channel extraction on a sonar side-scan image to generate training input of a codec; the method comprises the following steps: defining and optimizing a loss function of a codec; s2, generating a degraded image through preprocessing and downsampling of a sonar side-scan image, placing the degraded image into a trained encoder to generate potential representation and condition information, further training a UNet model in combination with a diffusion model, and improving noise prediction precision through a gradient descent method; s3, inputting the degraded image xlow and the water column region information c into an encoder tau to generate condition information d, and performing reverse diffusion operation by using a UNet network in combination with a randomly generated initial noise state zT; according to the method, the resolution and the quality of the sonar image are effectively improved by simulating the process of gradually recovering from a seriously degraded image to a high-quality image, and the method has the advantages that the method is high in resolution and quality, high in robustness and high in robustness, and the method is suitable for being applied to the field of sonar imaging and can be widely applied to the field of sonar imaging. The problem of poor image quality caused by high noise and low resolution of sonar images is solved, and the accuracy and effectiveness of subsequent image analysis are remarkably improved.

Description

Technical Field

[0001] The invention belongs to the field of image processing and underwater detection, and relates to a high-noise side-scan sonar image super-resolution processing method and device based on a diffusion model (Diffusion Model). Background Art

[0002] Sonar is an acoustic imaging technology used for terrain surveying and target detection in underwater environments. Since light waves decay rapidly in water, sound waves are the preferred means of detecting underwater targets, and can propagate over long distances in seawater with less loss. Sonar systems generate images of targets by emitting sound waves and receiving echoes after these sound waves interact with the bottom of the water or other underwater objects. However, due to the complex environment of seawater, the limitations on the number and quality of sensors, and the changes in the attitude and speed of the equipment itself, sonar images are often low in resolution and full of high noise, resulting in poor image quality, which seriously affects subsequent image processing and analysis. The acquisition of sonar images is limited by a variety of environmental factors. The data is not only precious but may also contain sensitive information, so its collection and sharing face severe challenges. In addition, high-noise side-scan sonar images are particularly difficult to process, which is particularly problematic under conditions of data scarcity.

[0003] To solve these problems, the "High Noise Side Scan Sonar Image Super-resolution Method Based on Diffusion Model" provides an effective solution. Diffusion model is an advanced generative model that can gradually recover clear images from high-noise data. This model effectively improves the resolution and quality of sonar images by simulating the process of gradually recovering from severely degraded images to high-quality images. By applying the diffusion model, the usability of sonar images can be significantly improved, thereby optimizing the accuracy of target detection and classification. Summary of the invention

[0004] The present invention aims to overcome the shortcomings of the existing high-noise side-scan sonar image processing technology, which leads to poor image quality due to low resolution, high noise and the like, and provides a sonar image super-resolution processing method and device based on a diffusion model.

[0005] By utilizing advanced diffusion model technology, the present invention can recover high-resolution, low-noise images from high-noise sonar data, significantly improving the quality of sonar images and the accuracy of subsequent image analysis.

[0006] In order to solve the technical problem of small sample size caused by the difficulty of sonar image acquisition, the technical solution adopted by the present invention is: a sonar image super-resolution processing method based on diffusion model. The subject of the super-resolution algorithm of the present invention is a latent diffusion model (Latent Diffusion Model). The final input of the algorithm of the present invention is a sonar scanning image x and its corresponding one-dimensional curve c of the water column area. The algorithm includes 4 main components:

[0007] Based on the features of the sonar side scan image, the input image is processed to interface with the diffusion model.

[0008] The entire image is compressed into a black and white intensity channel. The horizontal coordinate is converted into its directly corresponding echo time t, with the horizontal center corresponding to t=0, and then added to the input image as an additional channel. The continuous pure black area (also known as the water column area) close to the position of t=0 is extracted separately, and an additional channel with 1 outside the water column area and 0 inside the water column area is generated and added to the input image.

[0009] 1. Encoder-decoder to generate latent representation.

[0010] Encoder E(x)=z, decoder D(z)=x, where the letter x represents the sonar scanning image and the letter z represents the latent image.

[0011] 2. Neural Network UNet for Diffusion Process:

[0012] U(z i ,d)=z i-1

[0013] where z i It constitutes a diffusion process. From z0 to the final zT, the image quality deteriorates successively. z0 is the final output, z T is pure noise. The letter d represents the additional condition (Conditioning) that controls the diffusion process.

[0014] 3. Encoder for generating the additional condition d that controls the diffusion process:

[0015] τ(x,c)=d

[0016] The conceptual process of the super-resolution diffusion model is to add noise to the image first, and then decode the noise into a high-definition image. When adding noise, first input the high-definition image from the input end, then generate z0 latent image through the encoder E(x)=z, and then continue to add noise according to the law of the diffusion process to obtain the output z T The process from z0 to z T, the image quality gradually decreases and finally becomes pure noise. When decoding noise, the pure noise is reversely diffused using a neural network and additional control conditions to generate a clearer image z′0, and then the decoder D(z)=x is used to decode z′0 to generate a high-resolution image result.

[0017] A first aspect of the present invention relates to a method for super-resolution processing of high-noise side-scan sonar images based on a diffusion model, comprising:

[0018] S1. Construct a sonar image super-resolution processing model based on a diffusion model, wherein the sonar image super-resolution processing model includes an encoder E(x)=z and a decoder D(z)=x structure; generate the training input of the encoder and decoder by uniformly resizing, normalizing, and extracting specific channels of the sonar side scan images; define and optimize the loss function of the encoder and decoder, so that the model can effectively compress and restore the image, thereby improving the processing accuracy while retaining the key information of the image;

[0019] S2, generate degraded images by preprocessing and downsampling the sonar side scan images, put them into the trained encoder to generate potential representation and conditional information, further train the UNet model in combination with the diffusion model, and improve the noise prediction accuracy through the gradient descent method;

[0020] S3, by transforming the degraded image x low The water column area information c is input into the encoder τ to generate the conditional information d, and combined with the randomly generated initial noise state z T , using the UNet network for reverse diffusion operation; after multiple iterations of noise elimination, a clear potential representation is obtained, and the final super-resolution image is generated through the trained decoder D and the noise prediction network ∈

[0021] Wherein, step S1 specifically includes:

[0022] S11, collecting sonar side scan images and preprocessing them;

[0023] S111, adjusting all images to a uniform size to meet network input requirements;

[0024] S112, normalize the image pixel values ​​to [0,1] to improve the stability of model training;

[0025] S113, compressing the entire image into a black and white intensity channel;

[0026] S114, converting the horizontal coordinate into its corresponding echo time t, wherein the exact center in the horizontal direction corresponds to t=0, and adding it to the input image as an additional channel;

[0027] S115, extract the continuous pure black area (i.e., the water column area) close to the position t=0 separately, generate an additional channel with 1 outside the water column area and 0 inside the water column area, and add it to the input image;

[0028] S12. Define loss function L1=||xD(E(x))||2;

[0029] S121. Use the gradient descent algorithm to minimize the loss function L1. Through iterative optimization, adjust the parameters of the encoder and decoder so that the latent representation z compressed by the encoder E(x) can be restored as closely as possible to a high-quality image similar to the original image x through the decoder D(z).

[0030] Wherein, step S2 specifically includes:

[0031] S21, the image is resized to a uniform size and normalized, and then a degraded image x is generated by downsampling and adding noise {low} ; Then, use the trained encoder E(x) = z to encode x {low} Process it and get the potential representation z0;

[0032] S211, adjusting all images to a uniform size to meet network input requirements;

[0033] S212, normalize the image pixel values ​​to [0,1] to improve the stability of model training;

[0034] S213, downsampling the sonar side scan image to reduce the resolution, and adding noise to degrade it to generate a degraded image x low , corresponds to the input image in super-resolution applications;

[0035] S214, use the encoder E(x)=z trained in step S1 to process the degraded image x low , get the encoded potential representation z0;

[0036] S22, generate intermediate state z for UNet training i ;

[0037] S221, generate noise z T ;

[0038] S222, initial noise state z is mixed by weight T and the encoded latent representation z0 to generate the intermediate state z i , where z0 is the potential representation obtained by the encoder E(x), α i is the noise weight of the i-th step calculated according to the diffusion process;

[0039] S23, the intermediate state zi Input UNet for processing:

[0040] S231, generating additional conditions d (Conditioning) required in the diffusion process, the specific steps include: low and the one-dimensional curve c of the water column area are input into the encoder τ, and τ(x {low} ,c)=d, where the one-dimensional curve c of the water column area corresponds to the water column area mask in the input channel, and the encoder τ adopts the internal structure of the convolutional neural network (CNN) and is trained together with UNet;

[0041] S232, use the UNet model to input the intermediate state z i Process it with condition d to generate the updated state z i-1 , satisfying U(z i ,d,i)=z {i-1} ;

[0042] S233, during the training process, the noise prediction function ∈(z i ,d,i) predict the noise and control the progress of the reverse diffusion process through the following formula Where a and b are manually adjusted constants that control the reverse diffusion process;

[0043] S24, use the gradient descent method to minimize the loss function L2 = ||∈(z i ,τ(x {low} ,c),i)-z T ||2, to improve the accuracy of noise prediction in the reverse diffusion process and enhance the performance of the UNet model.

[0044] Wherein, step S3 specifically includes:

[0045] S31, randomly generate initial noise state z T ;

[0046] S32: Degrade the image x low and the one-dimensional curve c of the water column area are input into the encoder τ to generate the conditional information conditioning τ(x {low} ,c)=d;

[0047] S33, the initial noise state z T And conditional information d is input into the UNet network to obtain the predicted noise ∈(z T ,d,T);

[0048] S34, according to the predicted noise ∈(z T ,d,T) performs reverse diffusion operation, which specifically includes eliminating a certain proportion of prediction noise, calculating and obtaining the reverse diffusion result Where a and b are constants that control the diffusion process;

[0049] S35, repeating steps S33 and S34, and gradually iterating and performing reverse diffusion until the final clear code z0 is obtained;

[0050] S36, using decoder D to decode the clear code z0 to generate a super-resolution image

[0051] Wherein, in step S11, the input channels of the sonar image include:

[0052] Merge the three RGB channels into one brightness channel. The brightness information represents the echo intensity in the sonar image.

[0053] An additional echo time channel t is added to provide physical hints to help the network better distinguish the water column area and the seabed target area as well as the side scan direction;

[0054] The mask of the water column area, where the pixel value of the water column area is 0 and the pixel value outside is 1, indicating the specific location of the water column area.

[0055] Wherein, in step S221, the noise adding module gradually adds noise according to the law of the diffusion process until a pure noise image is generated.

[0056] The control condition generator in step S32 is used to adjust the denoising parameters in the reverse diffusion process to optimize the image restoration effect.

[0057] In step S34, the neural network UNet uses the control condition generator τ(x {low} ,c)=dThe control condition d generated by d is used for image denoising.

[0058] Wherein, the decoder in step S35 is used to restore the compressed latent image into a clear high-resolution image.

[0059] A second aspect of the present invention relates to a high-noise side-scan sonar image super-resolution processing device based on a diffusion model, comprising:

[0060] An encoder to convert the noisy side-scan sonar image into a latent image;

[0061] A noise addition module that gradually adds noise to the latent image;

[0062] A neural network, UNet, to process noisy images and perform inverse diffusion;

[0063] A control condition generator, used to generate control conditions for regulating the reverse diffusion process;

[0064] A decoder to restore the denoised latent image to a high-resolution image.

[0065] The innovation of the present invention is that the channels of the input image x of the super-resolution algorithm of the present invention and the traditional super-resolution algorithm are different, namely the following channels:

[0066] 1. Merge the three RGB channels into one brightness channel. Brightness information has a physical meaning in sonar images and indicates the intensity of the echo.

[0067] 2. An additional echo time channel t is added to provide the network with physically meaningful prompts, so that the network can better distinguish the water column area and the seabed target area, and better identify the current side scan direction, and then make targeted processing.

[0068] 3. Mask of the water column area. Set the pixels inside the water column area to 0 and the pixels outside to 1 to indicate the specific location of the water column area.

[0069] The present invention is also different from the traditional algorithm in Conditioning: an additional one-dimensional curve c of the water column region is added to the input of the encoder \tau that generates the additional condition d. The mask of the water column region in the input channel and the one-dimensional curve c of the water column region correspond to each other.

[0070] The invention has the advantages that by optimizing the channel setting of the input image and improving the condition generation process, the accuracy and adaptability of sonar image super-resolution processing are improved, so that the model can more effectively process high-noise sonar images.

[0071] 1. Introduction of brightness channel: This invention combines the traditional RGB channels into one brightness channel, so that the algorithm can more accurately capture the echo intensity information in the sonar image. This brightness information has physical meaning in the sonar image, which improves the effect of super-resolution processing.

[0072] 2. Addition of echo time channel: By adding the echo time channel, the present invention provides the network with prompts related to the physical characteristics of the sonar image, so that the model can better distinguish between the water column area and the seabed target area, and accurately identify the side scan direction, thereby achieving targeted processing.

[0073] 3. Combination of water column area mask and one-dimensional curve: The present invention introduces the water column area mask and the corresponding one-dimensional curve in the condition generation process, so that the model can more accurately locate and process the water column area when generating conditions, further improving the accuracy and adaptability of processing complex sonar images. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1This is a flow chart of defining the loss function of the present invention.

[0075] Figure 2 This is a training flowchart of UNet of the present invention.

[0076] Figure 3 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0077] In order to better understand the above technical solution, the technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0078] Example 1

[0079] like Figure 3 As shown, a high-noise side-scan sonar image super-resolution method based on a diffusion model is implemented, and the implementation process specifically includes the following steps:

[0080] S1. Construct a sonar image super-resolution processing model based on a diffusion model, wherein the sonar image super-resolution processing model includes an encoder E(x)=z and a decoder D(z)=x structure. The training input of the encoder and decoder is generated by uniformly resizing, normalizing, and extracting specific channels of the sonar side scan images. By defining and optimizing the loss function of the encoder and decoder, the model can effectively compress and restore the image, thereby improving the processing accuracy while retaining the key information of the image. This step mainly trains the encoder E(x)=z and decoder D(z)=x of the model.

[0081] S11. Collect sonar side scan image data to ensure that the data contains appropriate information about underwater terrain and targets; preprocess the images, including resizing and pixel normalization, to ensure that they meet the standards and requirements of network input. Compress the entire image into a black and white intensity channel. Convert the horizontal coordinate to its directly corresponding echo time t, with the horizontal center corresponding to t=0, and then add it to the input image as an additional channel. Separate the continuous pure black area (also known as the water column area) close to t=0, generate an additional channel with 1 outside the water column area and 0 inside the water column area, and add it to the input image.

[0082] S12. Define the loss function:

[0083] L1=||xD(E(x))||2 (1)

[0084] The loss function is optimized using a gradient descent algorithm to update the model parameters so that it can better reconstruct the original image.

[0085] S2. Degraded images are generated by preprocessing and downsampling the sonar side scan images, and put into the trained encoder to generate potential representation and conditional information. The UNet model is further trained in combination with the diffusion model, and the noise prediction accuracy is improved by the gradient descent method.

[0086] S21, the image is adjusted to a uniform size and normalized. The sonar side scan image is downsampled to reduce the resolution and noise is added to degrade it. The degraded image x {low} The input image can be used for the resolution method. Then, use the trained encoder E(x)=z to {low} Process it and get the potential representation z0.

[0087] S22, get z i , generating the input for UNet training.

[0088]

[0089] S23, additional conditions d for the diffusion process;

[0090] x {low} And the one-dimensional curve c of the water column area is put into the encoder τ, and we get:

[0091] τ(x {low} ,c)=d (3)

[0092] The one-dimensional curve c of the water column area corresponds to the mask of the water column area in the input channel; τ adopts the internal structure of CNN and is trained together with UNet.

[0093] With the additional conditioning (d) of UNet and the diffusion process, z i Put it into UNet as input and get:

[0094] U(z i ,d,i)=z {i-1} (4)

[0095] In practice, the traditional “learning noise” design is adopted, and the neural network actually trained is the noise prediction function ∈(z i ,d,i), is also a UNet structure, satisfying:

[0096]

[0097] Where a and b are manually adjusted constants that control the progress of the reverse diffusion process.

[0098] L2=||∈(z i ,τ(x {low} ,c),i)-zT ||2 (6)

[0099] Define the loss function L2 and use gradient descent to optimize it to minimize the loss.

[0100] S3, by transforming the degraded image x low The water column area information c is input into the encoder τ to generate the conditional information d, and combined with the randomly generated initial noise state z T , using the UNet network for reverse diffusion operations. After multiple iterations of noise elimination, a clear potential representation is obtained, and the final super-resolution image is generated through the trained decoder D and the noise prediction network ∈

[0101] S31, randomly generate noise z T

[0102] S32, put x {low} And the one-dimensional curve c of the water column area is put into the encoder τ to get conditioning:

[0103] τ(x {low} ,c)=d (7)

[0104] S33, put z T Put it into UNet and get the predicted noise:

[0105] ∈(z T ,d,T) (8)

[0106] S34, eliminate a certain proportion of the prediction noise and obtain the reverse diffusion result:

[0107]

[0108] S35. Repeat S33 and S34 until z0 is obtained.

[0109] S36, decode z0 with decoder D to obtain:

[0110]

[0111] Example 2

[0112] This embodiment relates to a high-noise side-scan sonar image super-resolution processing device based on a diffusion model, which is used to implement the high-noise side-scan sonar image super-resolution method based on a diffusion model described in Example 1, including:

[0113] An encoder to convert the noisy side-scan sonar image into a latent image;

[0114] A noise addition module that gradually adds noise to the latent image;

[0115] A neural network, UNet, to process noisy images and perform inverse diffusion;

[0116] A control condition generator, used to generate control conditions for regulating the reverse diffusion process;

[0117] A decoder to restore the denoised latent image to a high-resolution image.

[0118] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A high-noise side-scan sonar image super-resolution processing method based on a diffusion model, characterized in that: The following steps are involved: S1. Construct a sonar image super-resolution processing model based on a diffusion model, wherein the sonar image super-resolution processing model includes an encoder E(x)=z and a decoder D(z)=x structure; generate the training input of the encoder and decoder by uniformly resizing, normalizing, and extracting specific channels of the sonar side scan images; define and optimize the loss function of the encoder and decoder, so that the model can effectively compress and restore the image, thereby improving the processing accuracy while retaining the key information of the image; S2, generate degraded images by preprocessing and downsampling the sonar side scan images, put them into the trained encoder to generate potential representation and conditional information, further train the UNet model in combination with the diffusion model, and improve the noise prediction accuracy through the gradient descent method; S3, by transforming the degraded image x low The water column area information c is input into the encoder τ to generate the conditional information d, and combined with the randomly generated initial noise state z T , using the UNet network for reverse diffusion operation; after multiple iterations of noise elimination, a clear potential representation is obtained, and the final super-resolution image is generated through the trained decoder D and the noise prediction network ∈ 2. The method for super-resolution processing of high-noise side-scan sonar images based on a diffusion model according to claim 1, characterized in that: Step S1 specifically includes: S11, collecting sonar side scan images and preprocessing them; S111, adjusting all images to a uniform size to meet network input requirements; S112, normalizing the image pixel values ​​to [0, 1] to improve the stability of model training; S113, compressing the entire image into a black and white intensity channel; S114, converting the horizontal coordinate into its corresponding echo time t, wherein the exact center in the horizontal direction corresponds to t=0, and adding it to the input image as an additional channel; S115, extract the continuous pure black area (i.e., the water column area) close to the position t=0 separately, generate an additional channel with 1 outside the water column area and 0 inside the water column area, and add it to the input image; S12. Define the loss function L1 = ||xD(E(x))||2 S121. Use the gradient descent algorithm to minimize the loss function L1. Through iterative optimization, adjust the parameters of the encoder and decoder so that the latent representation z compressed by the encoder E(x) can be restored as closely as possible to a high-quality image similar to the original image x through the decoder D(z).

3. The method for super-resolution processing of high-noise side-scan sonar images based on a diffusion model as claimed in claim 1, characterized in that: Step S2 specifically includes: S21, the image is resized to a uniform size and normalized, and then a degraded image x is generated by downsampling and adding noise {low} ; Then, use the trained encoder E(x) = z to encode x {low} Process it and get the potential representation z0; S211, adjusting all images to a uniform size to meet network input requirements; S212, normalizing the image pixel values ​​to [0, 1] to improve the stability of model training; S21 3. Downsample the sonar side scan image to reduce the resolution and add noise to degrade it to generate a degraded image x low , corresponds to the input image in super-resolution applications; S214, use the encoder E(x)=z trained in step S1 to process the degraded image x low , get the encoded potential representation z0; S22, generate intermediate state z for UNet training i ; S221, generate noise z T ; S222, initial noise state z is mixed by weight T and the encoded latent representation z0 to generate the intermediate state z i , where z0 is the potential representation obtained by the encoder E(x), α i is the noise weight of the i-th step calculated according to the diffusion process; S23, the intermediate state z i Input UNet for processing; S231, generating additional conditions d (Conditioning) required in the diffusion process, the specific steps include: low and the one-dimensional curve c of the water column area are input into the encoder τ, and τ(x {low} , c)=d, where the one-dimensional curve c of the water column area corresponds to the water column area mask in the input channel, and the encoder τ adopts the internal structure of the convolutional neural network (CNN) and is trained together with UNet; S232, use the UNet model to input the intermediate state z i Process it with condition d to generate the updated state z i-1 , satisfying U(z i , d, i) = z {i-1} ; S233, during the training process, the noise prediction function ∈(z i , d, i) predict the noise and control the progress of the reverse diffusion process through the following formula Where a and b are manually adjusted constants that control the reverse diffusion process; S24, use the gradient descent method to minimize the loss function L2 = ||∈(z i ,τ(x {low} , c), i)-z T ||2, to improve the accuracy of noise prediction in the reverse diffusion process and enhance the performance of the UNet model.

4. The method for super-resolution processing of high-noise side-scan sonar images based on a diffusion model according to claim 1, characterized in that: Step S3 specifically includes: S31, randomly generate initial noise state z T ; S32: Degrade the image x low and the one-dimensional curve c of the water column area are input into the encoder τ to generate the conditional information conditioning τ(x {low} , c) = d; S33, the initial noise state z T And conditional information d is input into the UNet network to obtain the predicted noise ∈(z T , d, T); S34, according to the predicted noise ∈(z T , d, T) to perform reverse diffusion operations, including eliminating a certain proportion of prediction noise, calculating and obtaining the reverse diffusion results Where a and b are constants that control the diffusion process; S35, repeating steps S33 and S34, and gradually iterating and performing reverse diffusion until the final clear code z0 is obtained; S36, using decoder D to decode the clear code z0 to generate a super-resolution image 5. The method according to claim 1, characterized in that Step S11, the input channels of the sonar image include: Merge the three RGB channels into one brightness channel. The brightness information represents the echo intensity in the sonar image. An additional echo time channel t is added to provide physical hints to help the network better distinguish the water column area and the seabed target area as well as the side scan direction; The mask of the water column area, where the pixel value of the water column area is 0 and the pixel value outside is 1, indicating the specific location of the water column area.

6. The method according to claim 3, characterized in that In step S221, the noise adding module gradually adds noise according to the law of the diffusion process until a pure noise image is generated.

7. The method according to claim 4, characterized in that The control condition generator in step S32 is used to adjust the denoising parameters in the reverse diffusion process to optimize the image restoration effect.

8. The method according to claim 4, characterized in that Step S34: The neural network UNet uses the control condition generator τ(x {low} , c) = d to generate the control condition d for image denoising.

9. The method according to claim 4, characterized in that In step S35, the decoder is used to restore the compressed latent image into a clear high-resolution image.

10. A high-noise side-scan sonar image super-resolution processing device based on a diffusion model, characterized in that: include: An encoder to convert the noisy side-scan sonar image into a latent image; A noise addition module that gradually adds noise to the latent image; A neural network, UNet, to process noisy images and perform inverse diffusion; A control condition generator, used to generate control conditions for regulating the reverse diffusion process; A decoder to restore the denoised latent image to a high-resolution image.

Images