Chinese painting image super-resolution reconstruction method and device based on joint control

Abstract

The application provides a Chinese painting image super-resolution reconstruction method and device based on joint control, and the method comprises the following steps: acquiring a low-resolution Chinese painting image to be reconstructed; inputting the low-resolution Chinese painting image into a joint control latent diffusion model, so that the joint control latent diffusion model can perform super-resolution reconstruction based on the low-resolution Chinese painting image, and output a reconstructed high-resolution Chinese painting image. The application can effectively extract and reconstruct the deep semantic information of the Chinese painting, effectively overcome the structure distortion and hallucination under high magnification, and realize the super-resolution reconstruction and lossless repair of the damaged Chinese painting image.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method and apparatus for super-resolution reconstruction of traditional Chinese painting images based on joint control. Background Technology

[0002] Traditional Chinese painting is an artistic treasure nurtured by five thousand years of Chinese civilization, carrying extremely rich characteristics of its era, philosophical thought, and regional culture, possessing immeasurable historical and academic value. However, in the current work of cultural relic protection, the authentic preservation and revitalization of traditional Chinese paintings face enormous challenges. Due to the extreme fragility of the material carrier of traditional Chinese paintings, they are easily eroded by the natural environment over the long course of history, resulting in yellowing, mold, and brittleness. Coupled with human damage during the transmission process, many precious authentic works have suffered irreversible physical degradation and information loss. To solve this problem, it is urgent to rely on cutting-edge digital methods to perform non-destructive restoration and digital reconstruction of damaged traditional Chinese paintings. Image super-resolution reconstruction technology, as a key underlying algorithm of computer vision, can effectively mine and extract deep features from severely degraded low-resolution images. Through nonlinear mapping to a higher-dimensional space, this technology can not only achieve physical magnification of image resolution but also accurately reconstruct the unique brushstrokes, textures, and details of traditional Chinese paintings, providing scientific and reliable technical support for the high-definition digital restoration and permanent storage of traditional Chinese painting artifacts.

[0003] Currently, research on super-resolution digital restoration of traditional Chinese paintings is still in its initial and exploratory stage. The following core dilemmas need to be overcome in algorithm design and practical implementation: (1) The barrier to building high-quality domain-specific datasets. Most of the existing public digital resources of Chinese paintings have low resolution and serious uncontrollable degradation. The lack of high-quality, lossless reference images constitutes the primary physical obstacle to training highly available super-resolution networks. (2) Differences in the distribution of underlying data and distortion in degradation modeling. Current mainstream super-resolution algorithms rely heavily on natural scene images as training benchmarks. However, the brushstrokes, ink washes, and texture of Xuan paper contained in Chinese paintings constitute a unique distribution of artistic features, which cannot be effectively provided by natural images. In addition, conventional super-resolution preprocessing often uses a single linear downsampling. This idealized degradation setting cannot realistically simulate the extremely complex high-order composite degradation process that Chinese paintings have experienced over hundreds of years of inheritance, such as material aging, wear and tear, and multiple digital compressions. (3) The lack of domain-specific generation priors and the generalization dilemma. Traditional Chinese painting boasts a rich variety of themes and techniques, resulting in an extremely complex high-dimensional feature space. Currently available open-source general-purpose super-resolution models lack targeted learning mechanisms specific to the artistic characteristics of traditional Chinese painting, making them unable to effectively extract and reconstruct its deep semantic information. When faced with diverse inputs from traditional Chinese paintings, these general-purpose models often exhibit poor generalization performance, easily leading to structural distortions, color banding, or severe style shifts. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method and apparatus for super-resolution reconstruction of traditional Chinese painting images based on joint control, so as to eliminate or improve one or more defects existing in the prior art.

[0005] One aspect of the present invention provides a method for super-resolution reconstruction of traditional Chinese painting images based on joint control, the method comprising the following steps: Obtain the low-resolution image of the traditional Chinese painting to be reconstructed; The low-resolution traditional Chinese painting image is input into a joint control latent diffusion model (VCD) to enable the VAE to perform super-resolution reconstruction based on the low-resolution image, outputting a reconstructed high-resolution image. The VAE includes a VAE encoder, a conditional guidance controller, a denoising network with a low-rank matrix introduced in the weight space, and a VAE decoder. The VAE encoder performs perceptual compression and noise addition on the low-resolution image, outputting a noisy low-dimensional latent feature map. The conditional guidance controller performs multi-scale feature extraction based on the spatial structure conditional feature map, outputting a multi-scale spatial control feature map and injecting it into the feature space of the denoising network. The spatial structure conditional feature map is obtained by extracting spatial structure conditional features from the low-resolution image. The denoising network iteratively generates prediction noise based on the noisy low-dimensional latent feature map, the multi-scale spatial control feature map, and the low-rank matrix, and performs reverse iterative denoising on the noisy low-dimensional latent feature map, outputting a denoised low-dimensional latent feature map. The VAE decoder reconstructs the image based on the denoised low-dimensional latent feature map, outputting the reconstructed high-resolution image.

[0006] In some embodiments of the present invention, the denoising network employs an improved U-Net network, which includes a U-Net encoder, an attention module incorporating a low-rank matrix, and a U-Net decoder. The denoising network iteratively generates prediction noise based on a noisy low-dimensional latent feature map, the multi-scale spatial control feature map, and the low-rank matrix, and performs reverse iterative denoising on the noisy low-dimensional latent feature map to output a denoised low-dimensional latent feature map, including: In each iteration, the U-Net encoder performs feature extraction based on the downsampled image output from the previous scale level. The extracted feature maps at each scale are then summed pixel-by-pixel with the corresponding scale spatial control feature maps of the multi-scale spatial control feature map to obtain the summed results at each scale. Convolution operations and downsampling are then performed based on the summed results at each scale to output the downsampled image. The first scale level performs feature extraction based on the denoised result output from the previous iteration. In the initial iteration, the first scale level performs feature extraction based on the noisy low-dimensional latent feature map. The spatial resolution of the multi-scale spatial control feature map and the multi-scale feature map extracted by the multi-scale levels of the U-Net encoder are matched one-to-one. The attention module performs mapping calculations based on the downsampled map output from the last scale level of the U-Net encoder, the low-rank matrix, and the original weight matrix, and outputs an attention feature map that is adaptively fine-tuned for artistic style. The U-Net decoder performs convolution and upsampling operations on the concatenation results output from the previous scale level at each scale level to obtain an upsampled image. Based on the upsampled image and the sum of the corresponding scale, it performs symmetrical concatenation of features at the same scale and outputs the concatenation result. The first scale level is upsampled based on the attention feature map, and the concatenation result output from the last scale level is prediction noise. The denoising network performs reverse denoising on the noisy low-dimensional latent feature map based on the predicted noise, outputs the denoising result as the input for the next iteration, until the last iteration outputs the denoised low-dimensional latent feature map.

[0007] In some embodiments of the present invention, the extraction of spatial structure condition features from the low-resolution traditional Chinese painting image includes: The high-frequency edges of the low-resolution traditional Chinese painting image are extracted using an edge detection algorithm to obtain a high-frequency edge conditional feature map; The low-frequency colors of the low-resolution traditional Chinese painting image are extracted using K-means color quantization to obtain a low-frequency color conditional feature map. The high-frequency edge condition feature map and the low-frequency color condition feature map are spliced ​​and fused in the channel dimension to obtain the spatial structure condition feature map.

[0008] In some embodiments of the present invention, the method further includes: The denoising network iteratively generates prediction noise based on a noisy low-dimensional latent feature map, the multi-scale spatial control feature map, the low-rank matrix, and text semantic features. It then performs reverse iterative denoising on the noisy low-dimensional latent feature map and outputs a denoised low-dimensional latent feature map. The text semantic features are obtained by extracting the semantic features of text prompt words using a text encoder.

[0009] In some embodiments of the present invention, the denoising network employs an improved U-Net network, which includes a U-Net encoder, an attention module incorporating a low-rank matrix, and a U-Net decoder. The denoising network iteratively generates predictive noise based on a noisy low-dimensional latent feature map, the multi-scale spatial control feature map, the low-rank matrix, and textual semantic features, and performs reverse iterative denoising on the noisy low-dimensional latent feature map to output a denoised low-dimensional latent feature map, including: In each iteration, the U-Net encoder performs feature extraction based on the downsampled image output from the previous scale level. The extracted feature maps at each scale are then summed pixel-by-pixel with the corresponding scale spatial control feature maps of the multi-scale spatial control feature map to obtain the summed results at each scale. Convolution operations and downsampling are then performed based on the summed results at each scale to output the downsampled image. The first scale level performs feature extraction based on the denoised result output from the previous iteration. In the initial iteration, the first scale level performs feature extraction based on the noisy low-dimensional latent feature map. The spatial resolution of the multi-scale spatial control feature map and the multi-scale feature map extracted by the multi-scale levels of the U-Net encoder are matched one-to-one. The attention module employs a cross-attention mechanism, performing mapping calculations and feature fusion based on the downsampled image output from the last scale level of the U-Net encoder, text semantic features, the low-rank matrix, and the original weight matrix. This allows for the addition and non-linear mapping of the feature fusion result with the downsampled image output from the last scale level, resulting in an attention feature map that is adaptively fine-tuned for artistic style. The U-Net decoder performs convolution and upsampling operations on the concatenation results output from the previous scale level at each scale level to obtain an upsampled image. Based on the upsampled image and the sum of the corresponding scale, it performs symmetrical concatenation of features at the same scale and outputs the concatenation result. The first scale level is upsampled based on the attention feature map, and the concatenation result output from the last scale level is prediction noise. The denoising network performs reverse denoising on the noisy low-dimensional latent feature map based on the predicted noise, outputs the denoising result as the input for the next iteration, until the last iteration outputs the denoised low-dimensional latent feature map.

[0010] In some embodiments of the present invention, the joint control potential diffusion model is obtained in advance through the following steps: A pre-defined joint control latent diffusion model is trained based on multiple low-resolution traditional Chinese painting images and corresponding multiple high-resolution original traditional Chinese painting images to obtain the trained joint control latent diffusion model. During the training process, the parameters of the VAE encoder, the part of the denoising network except for the low-rank matrix, and the VAE decoder are frozen, and the parameters of the conditional guidance controller and the low-rank matrix are updated. The multiple low-resolution traditional Chinese painting images are obtained by using a multidimensional high-order degradation algorithm to degrade the multiple high-resolution original traditional Chinese painting images.

[0011] In some embodiments of the present invention, the degradation processing of the multiple high-definition original Chinese paintings using a multidimensional higher-order degradation algorithm includes: The first stage of degradation is performed based on the multiple high-definition original Chinese paintings, and multiple first stage degradation sequences are output. The first stage degradation includes spatial convolution blurring based on a hybrid operator pool composed of isotropic Gaussian kernels and antitropic Gaussian kernels with a preset ratio, bicubic downsampling, additive white Gaussian noise injection, and JPEG compression. Semantic edge modulation is performed based on the multiple first-stage degradation sequences, and multiple modulation results are output. The semantic edge modulation includes line feature mask extraction based on the EDTER semantic edge extraction operator and local strong degradation of spatial variants applied to the core stroke region. Based on the multiple modulation results, a second-stage degradation is performed, and multiple second-stage degradation sequences are output. The second-stage degradation includes random Gaussian blurring, bicubic downsampling, additive white Gaussian noise injection, and JPEG compression in sequence. Frequency domain truncation is performed based on the multiple second-stage degradation sequences to output multiple low-resolution traditional Chinese painting images. The frequency domain truncation includes using a two-dimensional Sinc filter to truncate the end of the second-stage degradation sequence in the frequency domain.

[0012] In some embodiments of the present invention, the step of training a preset joint control latent diffusion model based on multiple low-resolution traditional Chinese painting images and corresponding multiple high-resolution original traditional Chinese painting images by optimizing the global loss function of joint control, thereby obtaining the trained joint control latent diffusion model, wherein the global loss function is: In the formula, This represents the true latent feature map of the lossless, high-resolution original Chinese painting in the latent feature space. This represents the time step in the noise reduction process. Indicates text prompt words, This represents a high-frequency edge condition feature map. This represents a low-frequency color condition feature map. This represents real random Gaussian noise added to the low-dimensional latent feature map during the diffusion stage. Indicates the predicted noise value. These represent the parameters of the denoising network. Represents a multivariate standard normal distribution. Represents the identity matrix. Indicates at time step The noisy latent feature tensor at time t, i.e., the first... The denoising result of the next iteration. This represents multi-scale spatial control characteristics. This represents the original weight matrix of the denoising network in the weight space. and Represents a low-rank matrix. This represents the mean squared error loss function.

[0013] Another aspect of the present invention provides a super-resolution reconstruction device for traditional Chinese painting images based on joint control. The device includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the aforementioned method.

[0014] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned method.

[0015] Another aspect of the present invention provides a computer program product including computer instructions that, when executed by a processor, implement the steps of the aforementioned method.

[0016] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0017] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. The components in the drawings are not drawn to scale but are merely illustrative of the principles of the invention. For ease of illustration and description of certain parts of the invention, corresponding portions in the drawings may be enlarged, i.e., may appear larger relative to other components in an exemplary device actually manufactured according to the invention. In the drawings: Figure 1 This is a flowchart illustrating a method for super-resolution reconstruction of traditional Chinese painting images based on joint control, according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the specific process of a method for super-resolution reconstruction of traditional Chinese painting images based on joint control in one embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of a conditional guidance controller in one embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a low-rank adaptive module in one embodiment of the present invention; Figure 5 This is a flowchart illustrating a multidimensional higher-order degradation algorithm in one embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0020] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0021] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0022] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0023] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0024] Figure 1 This is a flowchart of a method for super-resolution reconstruction of traditional Chinese painting images based on joint control, as described in one embodiment of the present invention. Figure 1 and Figure 2 As shown, the method includes the following steps: Step S110: Obtain the low-resolution Chinese painting image to be reconstructed.

[0025] Specifically, the low-resolution traditional Chinese painting image to be reconstructed is the traditional Chinese painting image that has been damaged in the real environment.

[0026] Step S120: The low-resolution traditional Chinese painting image is input into the Joint Control Latent Diffusion Model (JCDLM) so that the JCDLM can perform super-resolution reconstruction based on the low-resolution traditional Chinese painting image and output a reconstructed high-resolution traditional Chinese painting image. The JCDLM includes a VAE encoder, a conditional guidance controller, a denoising network that introduces a low-rank matrix into the weight space, and a VAE decoder. The VAE encoder performs perceptual compression and noise addition on the low-resolution traditional Chinese painting image, outputting a noisy low-dimensional latent feature map. The conditional guidance controller performs multi-scale feature extraction based on the spatial structure conditional feature map, outputting a multi-scale spatial control feature map and injecting it into the feature space of the denoising network. The spatial structure conditional feature map is obtained by extracting spatial structure conditional features from the low-resolution traditional Chinese painting image. The denoising network iteratively generates prediction noise based on the noisy low-dimensional latent feature map, the multi-scale spatial control feature map, and the low-rank matrix, and performs reverse iterative denoising on the noisy low-dimensional latent feature map, outputting a denoised low-dimensional latent feature map. The VAE decoder reconstructs the image based on the denoised low-dimensional latent feature map, outputting the reconstructed high-resolution traditional Chinese painting image.

[0027] Specifically, the perceptual compression function of the VAE encoder is to compress or map the low-resolution traditional Chinese painting image from its computationally intensive, high-dimensional pixel space to a low-dimensional latent feature space, thereby significantly reducing computational complexity. The VAE decoder's function is to reverse-map and decode the denoised low-dimensional latent feature map back to the high-dimensional pixel space, outputting the final reconstructed high-resolution traditional Chinese painting image.

[0028] In some embodiments, the denoising network employs an improved U-Net network, which includes a U-Net encoder, an attention module incorporating a low-rank matrix, and a U-Net decoder. The denoising network iteratively generates predictive noise based on a noisy low-dimensional latent feature map, the multi-scale spatial control feature map, and the low-rank matrix, and performs reverse iterative denoising on the noisy low-dimensional latent feature map to output a denoised low-dimensional latent feature map, including: In each iteration, the U-Net encoder performs feature extraction based on the downsampled image output from the previous scale level. The extracted feature maps at each scale are then summed pixel-by-pixel with the corresponding scale spatial control feature maps of the multi-scale spatial control feature map to obtain the summed results at each scale. Convolution operations and downsampling are then performed based on the summed results at each scale to output the downsampled image. The first scale level performs feature extraction based on the denoised result output from the previous iteration. In the initial iteration, the first scale level performs feature extraction based on the noisy low-dimensional latent feature map. The spatial resolution of the multi-scale spatial control feature map and the multi-scale feature map extracted by the multi-scale levels of the U-Net encoder are matched one-to-one. The attention module performs mapping calculations based on the downsampled map output from the last scale level of the U-Net encoder, the low-rank matrix, and the original weight matrix, and outputs an attention feature map that is adaptively fine-tuned for artistic style. The U-Net decoder performs convolution and upsampling operations on the concatenation results output from the previous scale level at each scale level to obtain an upsampled image. Based on the upsampled image and the sum of the corresponding scale, it performs symmetrical concatenation of features at the same scale and outputs the concatenation result. The first scale level is upsampled based on the attention feature map, and the concatenation result output from the last scale level is prediction noise. The denoising network performs reverse denoising on the noisy low-dimensional latent feature map based on the predicted noise, outputs the denoising result as the input for the next iteration, until the last iteration outputs the denoised low-dimensional latent feature map.

[0029] Specifically, the improved U-Net network is a complete denoising backbone network. Its role is to receive feature injections from the conditionally guided controller in the latent feature space, progressively predicting and eliminating noise in low-dimensional latent features. In the reverse denoising process, the conditionally guided controller controls the feature map in a multi-scale space. In the U-Net encoder, at each scale level of the U-Net encoder, the injected features are fused using residual feature fusion. Spatial control feature map at various scales and the corresponding scale feature map extracted through feature extraction Perform pixel-by-pixel summation: By combining downsampling to gradually reduce the image size, the spatial structure features are precisely aligned and integrated into the generation path of the backbone network. This achieves direct intervention from the external physical space to the internal generation path, guiding the backbone network to perform high-quality pixel rendering under strict skeleton constraints. This achieves physical structure constraints in the feature space, "hard-coding" the outline skeleton of the traditional Chinese painting into the flowing data. Unlike the weak constraints of existing diffusion models that rely on cross-attention mechanisms to inject textual prompts, this direct and powerful feature addition can achieve hard physical skeleton locking. Furthermore, this innovative design of the invention achieves complete decoupling of image generation capabilities and spatial control capabilities: U-Net is only responsible for providing general generation priors and high-quality pixel rendering, while the conditional guidance controller precisely locks the physical skeleton and macroscopic color layout of the traditional Chinese painting through hard injection.

[0030] The conditional guided controller is a lightweight convolutional neural network independent of the backbone network (i.e., the denoising network) of the joint control latent diffusion model. This network consists of multiple scale feature extraction modules comprised of a series of cascaded residual blocks and downsampling layers. Its input is an externally extracted spatial structure conditional feature map, obtained by concatenating and fusing high-frequency edge features extracted via EDTER with low-frequency color features obtained through K-means processing along the channel dimension. After multi-scale feature extraction via residual blocks and downsampling layers, the controller outputs a multi-scale spatial control feature set that matches the spatial resolution of the multiple scale levels of the U-Net encoder. Because the U-Net encoder uses a pyramid structure and contains multiple feature scale levels with different spatial resolutions (in this specific example, there are four scale levels), the conditional guidance controller also outputs four feature tensors of different scales to achieve accurate spatial alignment between the external spatial features and the features generated internally by the encoder. Figure 3 As shown, this multi-scale spatial control feature set is designed to contain four feature tensors of different scales to ensure that its spatial resolution corresponds one-to-one with the spatial resolution of the four feature scale levels in the U-Net encoder.

[0031] The improved U-Net network's attention module is formed by side-channeling a trainable low-rank matrix into the linear projection layer of the original U-Net network's attention module, based on the original weight matrix, thus achieving efficient parameter fine-tuning. In a specific example, two trainable low-rank matrices A and B are sequentially connected in the side-channel of the linear projection layer. These two matrices form a low-rank adaptive module (LoRA), which is then combined with the original weight matrix from the other path of the linear projection layer. Parallel connection, such as Figure 4 As shown. The weights are updated after LoRA injection as follows: ,in, , , The eigendimensionality and rank of the original weight matrix are... much smaller Its function is to accurately inject the unique textures of traditional Chinese painting, such as brushstrokes, ink washes, and the microscopic texture of Xuan paper, into large models with extremely small parameters and very low computational cost, thereby achieving efficient style domain transfer.

[0032] In the specific mapping calculation process, the downsampled map output of the last scale level of the U-Net encoder is regarded as a matrix. The original U-Net network uses the pre-trained original weight matrix. and Multiplication, while this invention uses a calculation formula. right Perform mapping and output attention feature map. This invention meticulously integrates the unique ink wash and Xuan paper texture of traditional Chinese painting into deep semantics. Compared to the conventional approach of adjusting all parameters of a large model in existing technologies, this specific design not only significantly reduces the amount of parameter updates and memory usage during model training, and exerts a natural regularization effect, but also fundamentally avoids catastrophic forgetting when the model learns the microscopic texture of traditional Chinese ink wash and Xuan paper. Thus, while preserving the basic semantic generation capabilities of the pre-trained model, it achieves high-fidelity mapping of the style of traditional Chinese painting. This invention also achieves decoupling of the feature space and weight space of the improved U-Net network. Through joint control of the feature space and weight space, it effectively overcomes structural breakage under high-magnification magnification.

[0033] The U-Net decoder uses upsampling and symmetrical stitching of features at the same scale to accurately recover lost spatial details at each layer as the image is enlarged back from a smaller size, thus ensuring that the generated image does not distort in terms of detail structure. After multi-step iterative deep fusion of spatial structure constraints and Chinese painting style priors by the improved U-Net network and denoising, clean high-frequency latent features are obtained, which are then used to obtain a denoised low-dimensional latent feature map.

[0034] In some embodiments, the extraction of spatial structure condition features from the low-resolution traditional Chinese painting image includes: The high-frequency edges of the low-resolution traditional Chinese painting image are extracted using an edge detection algorithm to obtain a high-frequency edge conditional feature map; The low-frequency colors of the low-resolution traditional Chinese painting image are extracted using K-means color quantization to obtain a low-frequency color conditional feature map. The high-frequency edge condition feature map and the low-frequency color condition feature map are spliced ​​and fused in the channel dimension to obtain the spatial structure condition feature map.

[0035] Specifically, the EDTER edge detection algorithm is used to extract high-frequency edges from low-resolution traditional Chinese painting images, resulting in a high-frequency edge conditional feature map. This map serves as the structural condition for constraining brushstroke features such as line drawing, aiming to extract and lock the macroscopic physical skeleton of the painting. It provides the network with hard spatial constraints on brushstroke features like line drawing and object outlines, effectively preventing structural distortions or meaningless high-frequency illusions from arising during the joint control latent diffusion model during the reconstruction and generation of the painting. K-means color quantization is then used to extract low-frequency color conditions from the low-resolution painting images, resulting in a low-frequency color conditional feature map. This map extracts low-frequency color distribution information from the painting, constraining large areas of flat washes and ink wash areas. This prevents severe color gamut shifts or color banding at high magnification, ensuring high consistency between the ink wash layers and the original painting. Finally, the structural conditions and low-frequency color conditions are concatenated and fused along the channel dimension to obtain spatial structural conditional features.

[0036] In some embodiments, the method further includes: The denoising network iteratively generates prediction noise based on a noisy low-dimensional latent feature map, the multi-scale spatial control feature map, the low-rank matrix, and text semantic features. It then performs reverse iterative denoising on the noisy low-dimensional latent feature map and outputs a denoised low-dimensional latent feature map. The text semantic features are obtained by extracting the semantic features of text prompt words using a text encoder.

[0037] Specifically, since the latent diffusion model requires textual instructions to activate its generative capabilities, the text encoder is responsible for receiving external textual prompts, extracting their semantic features, and mapping them from natural language to high-dimensional text embedding vectors, which serve as the basic generative prior conditions. The textual prompts are descriptive instructions used to guide the generation direction of the large model, and the text embedding vectors contain semantic priors describing the style of traditional Chinese painting. In the super-resolution task of this invention, the text is typically a pre-defined, general term describing the overall style of the image, such as: "A high-quality traditional Chinese painting, ink wash, highly detailed, masterpiece." Because the specific lines and color structure of the image are already rigidly constrained by the conditional guidance controller, the textual prompts are merely a global semantic prior auxiliary input. In actual implementation, the textual prompts can be set to a fixed general text or an empty string.

[0038] In some embodiments, the denoising network employs an improved U-Net network, which includes a U-Net encoder, an attention module incorporating a low-rank matrix, and a U-Net decoder. The denoising network iteratively generates predictive noise based on a noisy low-dimensional latent feature map, the multi-scale spatial control feature map, the low-rank matrix, and textual semantic features, and performs reverse iterative denoising on the noisy low-dimensional latent feature map to output a denoised low-dimensional latent feature map, including: In each iteration, the U-Net encoder performs feature extraction based on the downsampled image output from the previous scale level. The extracted feature maps at each scale are then summed pixel-by-pixel with the corresponding scale spatial control feature maps of the multi-scale spatial control feature map to obtain the summed results at each scale. Convolution operations and downsampling are then performed based on the summed results at each scale to output the downsampled image. The first scale level performs feature extraction based on the denoised result output from the previous iteration. In the initial iteration, the first scale level performs feature extraction based on the noisy low-dimensional latent feature map. The spatial resolution of the multi-scale spatial control feature map and the multi-scale feature map extracted by the multi-scale levels of the U-Net encoder are matched one-to-one. The attention module employs a cross-attention mechanism, performing mapping calculations and feature fusion based on the downsampled image output from the last scale level of the U-Net encoder, text semantic features, the low-rank matrix, and the original weight matrix. This allows for the addition and non-linear mapping of the feature fusion result with the downsampled image output from the last scale level, resulting in an attention feature map that is adaptively fine-tuned for artistic style. The U-Net decoder performs convolution and upsampling operations on the concatenation results output from the previous scale level at each scale level to obtain an upsampled image. Based on the upsampled image and the sum of the corresponding scale, it performs symmetrical concatenation of features at the same scale and outputs the concatenation result. The first scale level is upsampled based on the attention feature map, and the concatenation result output from the last scale level is prediction noise. The denoising network performs reverse denoising on the noisy low-dimensional latent feature map based on the predicted noise, outputs the denoising result as the input for the next iteration, until the last iteration outputs the denoised low-dimensional latent feature map.

[0039] In this embodiment, the attention module introduces a low-rank matrix in the same way as in the previous embodiment. During the specific mapping calculation process, the downsampled image output from the last scale level of the U-Net encoder is first... and text semantic features The mapping is represented by a query vector (Query, Q), a key vector (Key, K), and a value vector (Value, V), and the calculation formula is as follows: In the formula, , and This is the original weight matrix. , , , , and This is a trainable low-rank matrix specific to the style of traditional Chinese painting. Then, by calculating the similarity between the query vector (from the image) and the key vector (from the text), the text information (value vector) that each region in the image should absorb is determined. The calculation formula is as follows: In the formula, The result is the feature fusion output, where D is a scaling factor, typically the size of the feature dimension. The feature fusion result is then fed back into the downsampled image of the last scale level output of the U-Net encoder via a residual connection. Then, the features are passed through a feedforward neural network for non-linear mapping to obtain the attention feature map of the current module.

[0040] During the training phase, initial values ​​need to be assigned to the low-rank matrix mentioned above in this embodiment. To ensure that the low-rank adaptive module injected through the bypass does not destroy the original generative capability of the pre-trained model in the initial state, this invention adopts an asymmetric initialization strategy, that is, using a random Gaussian distribution initialization method to initialize the dimensionality-reduced low-rank matrix. , , Perform random assignment to force initialization of the upgraded low-rank matrix. , , For a matrix consisting entirely of zeros, training is performed on a low-rank matrix. , , and , , Update the matrix. Initially all zeros, in the first step of training, the product... It must be an all-zero matrix, resulting in the weights after LoRA injection in the initial state. This means that when the model starts training, it is equivalent to the original pre-trained latent diffusion model, allowing the training process to start smoothly.

[0041] In some embodiments, the joint control potential diffusion model is pre-trained through the following steps: A pre-defined joint control latent diffusion model is trained based on multiple low-resolution traditional Chinese painting images and corresponding multiple high-resolution original traditional Chinese painting images to obtain the trained joint control latent diffusion model. During the training process, the parameters of the VAE encoder, the part of the denoising network except for the low-rank matrix, and the VAE decoder are frozen, and the parameters of the conditional guidance controller and the low-rank matrix are updated. The multiple low-resolution traditional Chinese painting images are obtained by using a multidimensional high-order degradation algorithm to degrade the multiple high-resolution original traditional Chinese painting images.

[0042] During training, freezing the original weight matrices of the backbone or denoising network of the pre-trained joint control latent diffusion model preserves the superior macroscopic composition and basic semantic generation capabilities learned by the large base model on massive images, avoiding or even completely preventing catastrophic forgetting during fine-tuning for specific artistic styles. Backpropagation is used to update and optimize the parameters of the conditional controller and the low-rank matrix (or LoRA formed by the low-rank matrix). Parameters involved in model training, such as learning rate, batch size, and number of training steps, can be configured according to actual needs.

[0043] In some embodiments, the degradation processing of the multiple high-resolution original Chinese paintings using a multidimensional higher-order degradation algorithm includes: The first stage of degradation is performed based on the multiple high-definition original Chinese paintings, and multiple first stage degradation sequences are output. The first stage degradation includes spatial convolution blurring based on a hybrid operator pool composed of isotropic Gaussian kernels and antitropic Gaussian kernels with a preset ratio, bicubic downsampling, additive white Gaussian noise injection, and JPEG compression. Semantic edge modulation is performed based on the multiple first-stage degradation sequences, and multiple modulation results are output. The semantic edge modulation includes line feature mask extraction based on the EDTER semantic edge extraction operator and local strong degradation of spatial variants applied to the core stroke region. Based on the multiple modulation results, a second-stage degradation is performed, and multiple second-stage degradation sequences are output. The second-stage degradation includes random Gaussian blurring, bicubic downsampling, additive white Gaussian noise injection, and JPEG compression in sequence. Frequency domain truncation is performed based on the multiple second-stage degradation sequences to output multiple low-resolution traditional Chinese painting images. The frequency domain truncation includes using a two-dimensional Sinc filter to truncate the end of the second-stage degradation sequence in the frequency domain.

[0044] like Figure 5As shown, for each high-definition original Chinese painting image, in the first stage of degradation, the high-definition original Chinese painting image is subjected to spatial convolution blurring, initial bicubic downsampling, initial additive white Gaussian noise injection, and initial JPEG compression using a hybrid operator pool composed of a certain proportion of isotropic Gaussian kernels and heterotropic Gaussian kernels. Preferably, the hybrid operator pool is composed of 70% isotropic Gaussian kernels and 30% heterotropic Gaussian kernels. Among them, the 70% isotropic Gaussian kernels are used to simulate conventional optical defocusing, and the 30% heterotropic Gaussian kernels are specifically used to simulate the asymmetric diffusion of ink following the Xuan paper fibers and the texture features of the brush side stroke. Bicubic downsampling and JPEG compression are used to realistically simulate the frequency domain quantization distortion and high-frequency detail loss of ancient paintings in the process of digital archiving and storage. The injection of additive white Gaussian noise reproduces the thermal noise of the physical acquisition equipment.

[0045] During semantic edge modulation, the EDTER (Edge Detection Transformer, an edge detection method based on the Transformer architecture) semantic edge extraction operator is used to extract feature masks of lines, and local strong degradation of spatial variants is applied to the core stroke areas. By accurately locating core strokes such as line drawings through the mask, and applying stronger blurring and information erasure to these areas, the effect is to forcibly cut off the dependence of the joint control latent diffusion model on local shallow pixels, forcing the model to use global prior knowledge for deep semantic completion during training, thereby realistically replicating the evolution trajectory of structural damage in ancient paintings.

[0046] In the second stage of degradation, random Gaussian blurring, bicubic downsampling, additive white Gaussian noise injection, and JPEG compression are performed sequentially. This can replicate the deep nonlinear composite degradation caused by multiple forwardings, adaptive scaling, and secondary recoding of Chinese painting images on online social platforms. This further widens the gap between the synthetic data and the distribution of real damaged ancient paintings, providing a highly challenging blind super-resolution data base for the model.

[0047] In the frequency domain truncation process, a two-dimensional Sinc filter is introduced to truncate the end of the degraded sequence. Utilizing the low-pass characteristic of the filter, ringing artifacts are generated at high-frequency edges of the image, realistically restoring the digital pollution caused by arbitrary scaling of the image. Traditional Chinese paintings often feature continuous brushstrokes with water-like ripples, which, combined with the ringing artifacts generated by Sinc filtering, create visual deception. This frequency domain truncation operation forces the joint control latent diffusion model to perform deep semantic analysis, enabling it to accurately distinguish between natural ink diffusion and artificially truncated oscillations, ensuring the purity and crispness of the brushstroke edges in the final super-resolution reconstruction. Therefore, by using a multi-dimensional high-order degradation algorithm to process lossless high-resolution original images of traditional Chinese paintings, the physical and digital composite damage of traditional Chinese paintings can be realistically reproduced, thus overcoming the barrier to constructing high-quality traditional Chinese painting datasets for training highly available super-resolution networks.

[0048] In some embodiments, based on multiple low-resolution traditional Chinese painting images and corresponding multiple high-resolution original traditional Chinese painting images, a preset joint control latent diffusion model is trained by optimizing the global loss function of the joint control, resulting in a trained joint control latent diffusion model. The global loss function is: In the formula, This represents the true latent feature map of the lossless, high-resolution original Chinese painting in the latent feature space. This represents the time step in the noise reduction process. Indicates text prompt words, This represents a high-frequency edge condition feature map. This represents a low-frequency color condition feature map. This represents real random Gaussian noise added to the low-dimensional latent feature map during the diffusion stage. Indicates the predicted noise value. These represent the parameters of the denoising network. Represents a multivariate standard normal distribution. Represents the identity matrix. Indicates at time step The noisy latent feature tensor at time t, i.e., the first... The denoising result of the next iteration. This represents multi-scale spatial control characteristics. This represents the original weight matrix of the denoising network in the weight space. and Represents a low-rank matrix. Let represent the mean squared error loss function. By optimizing the mean squared error loss function between the predicted noise and the added real noise, the entire joint control latent diffusion model can be robustly driven to approach the optimal solution at extremely low cost, achieving a unity between the low-level pixel-level restoration of damaged traditional Chinese paintings and high-dimensional artistic aesthetics.

[0049] In summary, the super-resolution reconstruction method for traditional Chinese painting images based on joint control provided in this invention proposes a joint control super-resolution framework that integrates a conditional guided controller and a low-rank matrix. This framework, relying on a latent diffusion model, innovatively achieves decoupled control of traditional Chinese painting features: on the one hand, it introduces a multi-scale spatial control feature map output by the conditional guided controller into the feature space, injecting edge and color quantization features to achieve macroscopic physical skeleton constraints; on the other hand, it introduces a LoRA formed by a low-rank matrix into the weight space, fine-tuning the injected ink wash and Xuan paper texture priors with extremely low computational overhead. Through the synergy of these two approaches, the deep semantic information of traditional Chinese paintings can be effectively extracted and reconstructed, and the structural distortion and generation illusion under high-magnification super-resolution can be effectively overcome, achieving super-resolution reconstruction and lossless restoration of damaged traditional Chinese painting images.

[0050] This method also breaks through the limitations of general natural images, selectively collecting and constructing over ten thousand high-resolution traditional Chinese painting datasets. A multi-dimensional high-order degradation algorithm is proposed to realistically replicate the damage in all the high-resolution traditional Chinese painting datasets, thus constructing a high-quality dataset specifically for training the aforementioned framework. Simultaneously, the degradation algorithm abandons the assumption of a single linear degradation, innovatively proposing a two-stage high-order degradation model that integrates EDTER semantic edge modulation. This model deeply couples factors such as optical blurring, noise, compression, and frequency domain truncation, realistically replicating the complex physical and digital damage preserved in the history of traditional Chinese paintings, providing a high-fidelity data foundation for model training.

[0051] Corresponding to the above method, this embodiment of the invention also provides a super-resolution reconstruction device for traditional Chinese painting images based on joint control. The device includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the aforementioned method.

[0052] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned method. The computer-readable storage medium may be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium known in the art.

[0053] This invention also provides a computer program product, including computer instructions that, when executed by a processor, implement the steps of the aforementioned method.

[0054] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0055] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0056] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for super-resolution reconstruction of traditional Chinese painting images based on joint control, characterized in that, The method includes: Obtain the low-resolution image of the traditional Chinese painting to be reconstructed; The low-resolution traditional Chinese painting image is input into a joint control latent diffusion model (VCD) to enable the VAE to perform super-resolution reconstruction based on the low-resolution image, outputting a reconstructed high-resolution image. The VAE includes a VAE encoder, a conditional guidance controller, a denoising network with a low-rank matrix introduced in the weight space, and a VAE decoder. The VAE encoder performs perceptual compression and noise addition on the low-resolution image, outputting a noisy low-dimensional latent feature map. The conditional guidance controller performs multi-scale feature extraction based on the spatial structure conditional feature map, outputting a multi-scale spatial control feature map and injecting it into the feature space of the denoising network. The spatial structure conditional feature map is obtained by extracting spatial structure conditional features from the low-resolution image. The denoising network iteratively generates prediction noise based on the noisy low-dimensional latent feature map, the multi-scale spatial control feature map, and the low-rank matrix, and performs reverse iterative denoising on the noisy low-dimensional latent feature map, outputting a denoised low-dimensional latent feature map. The VAE decoder reconstructs the image based on the denoised low-dimensional latent feature map, outputting the reconstructed high-resolution image.

2. The method according to claim 1, characterized in that, The denoising network employs an improved U-Net network, which includes a U-Net encoder, an attention module incorporating a low-rank matrix, and a U-Net decoder. The denoising network iteratively generates prediction noise based on a noisy low-dimensional latent feature map, the multi-scale spatial control feature map, and the low-rank matrix, and performs reverse iterative denoising on the noisy low-dimensional latent feature map, outputting a denoised low-dimensional latent feature map, including: In each iteration, the U-Net encoder performs feature extraction based on the downsampled image output from the previous scale level. The extracted feature maps at each scale are then summed pixel-by-pixel with the corresponding scale spatial control feature maps of the multi-scale spatial control feature map to obtain the summed results at each scale. Convolution operations and downsampling are then performed based on the summed results at each scale to output the downsampled image. The first scale level performs feature extraction based on the denoised result output from the previous iteration. In the initial iteration, the first scale level performs feature extraction based on the noisy low-dimensional latent feature map. The spatial resolution of the multi-scale spatial control feature map and the multi-scale feature map extracted by the multi-scale levels of the U-Net encoder are matched one-to-one. The attention module performs mapping calculations based on the downsampled map output from the last scale level of the U-Net encoder, the low-rank matrix, and the original weight matrix, and outputs an attention feature map that is adaptively fine-tuned for artistic style. The U-Net decoder performs convolution and upsampling operations on the concatenation results output from the previous scale level at each scale level to obtain an upsampled image. Based on the upsampled image and the sum of the corresponding scale, it performs symmetrical concatenation of features at the same scale and outputs the concatenation result. The first scale level is upsampled based on the attention feature map, and the concatenation result output from the last scale level is prediction noise. The denoising network performs reverse denoising on the noisy low-dimensional latent feature map based on the predicted noise, outputs the denoising result as the input for the next iteration, until the last iteration outputs the denoised low-dimensional latent feature map.

3. The method according to claim 1, characterized in that, The extraction of spatial structure condition features from the low-resolution traditional Chinese painting image includes: The high-frequency edges of the low-resolution traditional Chinese painting image are extracted using an edge detection algorithm to obtain a high-frequency edge conditional feature map; The low-frequency colors of the low-resolution traditional Chinese painting image are extracted using K-means color quantization to obtain a low-frequency color conditional feature map. The high-frequency edge condition feature map and the low-frequency color condition feature map are spliced ​​and fused in the channel dimension to obtain the spatial structure condition feature map.

4. The method according to claim 1, characterized in that, The method further includes: The denoising network iteratively generates prediction noise based on a noisy low-dimensional latent feature map, the multi-scale spatial control feature map, the low-rank matrix, and text semantic features. It then performs reverse iterative denoising on the noisy low-dimensional latent feature map and outputs a denoised low-dimensional latent feature map. The text semantic features are obtained by extracting the semantic features of text prompt words using a text encoder.

5. The method according to claim 4, characterized in that, The denoising network employs an improved U-Net network, which includes a U-Net encoder, an attention module incorporating a low-rank matrix, and a U-Net decoder. The denoising network iteratively generates predictive noise based on a noisy low-dimensional latent feature map, the multi-scale spatial control feature map, the low-rank matrix, and textual semantic features. It then performs reverse iterative denoising on the noisy low-dimensional latent feature map, outputting a denoised low-dimensional latent feature map, including: In each iteration, the U-Net encoder performs feature extraction based on the downsampled image output from the previous scale level. The extracted feature maps at each scale are then summed pixel-by-pixel with the corresponding scale spatial control feature maps of the multi-scale spatial control feature map to obtain the summed results at each scale. Convolution operations and downsampling are then performed based on the summed results at each scale to output the downsampled image. The first scale level performs feature extraction based on the denoised result output from the previous iteration. In the initial iteration, the first scale level performs feature extraction based on the noisy low-dimensional latent feature map. The spatial resolution of the multi-scale spatial control feature map and the multi-scale feature map extracted by the multi-scale levels of the U-Net encoder are matched one-to-one. The attention module employs a cross-attention mechanism, performing mapping calculations and feature fusion based on the downsampled image output from the last scale level of the U-Net encoder, text semantic features, the low-rank matrix, and the original weight matrix. This allows for the addition and non-linear mapping of the feature fusion result with the downsampled image output from the last scale level, resulting in an attention feature map that is adaptively fine-tuned for artistic style. The U-Net decoder performs convolution and upsampling operations on the concatenation results output from the previous scale level at each scale level to obtain an upsampled image. Based on the upsampled image and the sum of the corresponding scale, it performs symmetrical concatenation of features at the same scale and outputs the concatenation result. The first scale level is upsampled based on the attention feature map, and the concatenation result output from the last scale level is prediction noise. The denoising network performs reverse denoising on the noisy low-dimensional latent feature map based on the predicted noise, outputs the denoising result as the input for the next iteration, until the last iteration outputs the denoised low-dimensional latent feature map.

6. The method according to claim 1 or 4, characterized in that, The joint control potential diffusion model is pre-trained through the following steps: A pre-defined joint control latent diffusion model is trained based on multiple low-resolution traditional Chinese painting images and corresponding multiple high-resolution original traditional Chinese painting images to obtain the trained joint control latent diffusion model. During the training process, the parameters of the VAE encoder, the part of the denoising network except for the low-rank matrix, and the VAE decoder are frozen, and the parameters of the conditional guidance controller and the low-rank matrix are updated. The multiple low-resolution traditional Chinese painting images are obtained by using a multidimensional high-order degradation algorithm to degrade the multiple high-resolution original traditional Chinese painting images.

7. The method according to claim 6, characterized in that, The degradation processing of the multiple high-resolution original Chinese paintings using a multidimensional higher-order degradation algorithm includes: The first stage of degradation is performed based on the multiple high-definition original Chinese paintings, and multiple first stage degradation sequences are output. The first stage degradation includes spatial convolution blurring based on a hybrid operator pool composed of isotropic Gaussian kernels and antitropic Gaussian kernels with a preset ratio, bicubic downsampling, additive white Gaussian noise injection, and JPEG compression. Semantic edge modulation is performed based on the multiple first-stage degradation sequences, and multiple modulation results are output. The semantic edge modulation includes line feature mask extraction based on the EDTER semantic edge extraction operator and local strong degradation of spatial variants applied to the core stroke region. Based on the multiple modulation results, a second-stage degradation is performed, and multiple second-stage degradation sequences are output. The second-stage degradation includes random Gaussian blurring, bicubic downsampling, additive white Gaussian noise injection, and JPEG compression in sequence. Frequency domain truncation is performed based on the multiple second-stage degradation sequences to output multiple low-resolution traditional Chinese painting images. The frequency domain truncation includes using a two-dimensional Sinc filter to truncate the end of the second-stage degradation sequence in the frequency domain.

8. The method according to claim 6, characterized in that, The method involves training a pre-defined joint control latent diffusion model based on multiple low-resolution images of traditional Chinese paintings and corresponding multiple high-resolution original images of traditional Chinese paintings, by optimizing the global loss function of the joint control, to obtain the trained joint control latent diffusion model. The global loss function is: In the formula, This represents the true latent feature map of the lossless, high-resolution original Chinese painting in the latent feature space. This represents the time step in the noise reduction process. Indicates text prompt words, This represents a high-frequency edge condition feature map. This represents a low-frequency color condition feature map. This represents real random Gaussian noise added to the low-dimensional latent feature map during the diffusion stage. Indicates the predicted noise value. These represent the parameters of the denoising network. Represents a multivariate standard normal distribution. Represents the identity matrix. Indicates at time step The noisy latent feature tensor at time t, i.e., the first... The denoising result of the next iteration. This represents multi-scale spatial control characteristics. This represents the original weight matrix of the denoising network in the weight space. and Represents a low-rank matrix. This represents the mean squared error loss function.

9. A device for super-resolution reconstruction of traditional Chinese painting images based on joint control, comprising a processor, a memory, and computer instructions stored in the memory, characterized in that, The processor is configured to execute the computer instructions, and when the computer instructions are executed, the device implements the steps of the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method as described in any one of claims 1 to 8.

Images