A super-resolution method for high-resolution image generation

Abstract

The application relates to a super-resolution method for high-resolution image generation, comprising the following steps: constructing a paired training data set generated by uniform tiling cutting and degradation processing of a high-resolution image, wherein the data set comprises a low-quality image, a semantic text instruction and a high-quality image; based on a DiT architecture, training a conditional generation model with low-quality image latent features as a structural condition and a text instruction as a semantic condition; during preprocessing, uniformly tiling cutting is performed on a high-resolution input image to be enlarged, and each low-quality image tile is independently iteratively denoised and enlarged in the trained model; cross-block color difference of each high-quality image tile after enlargement is eliminated, and fusion splicing based on a weight mask is performed in an overlapping area to obtain a high-resolution output image. The application effectively solves the problems of detail style drift and inconsistent semantics in block processing in the existing diffusion model enlargement method, and realizes high-fidelity and high-consistency image super-resolution reconstruction.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically a super-resolution method for generating high-resolution images. Background Technology

[0002] Image upscaling (Super-Resolution, SR) aims to restore the detail and structural consistency of high-resolution images from lower-resolution inputs. This is a fundamental task in image processing and computer vision, with wide applications in high-fidelity upscaling of photographic, image enhancement, document and map reconstruction, and illustrative images such as animation, comics, and game illustrations. Current mainstream technologies include:

[0003] 1. Traditional interpolation methods: These methods originate from traditional signal and image processing, such as bilinear and bicubic interpolation. They use deterministic kernels for upsampling. Although they are simple, fast, and easy to deploy, they cannot recover high-frequency details, are prone to producing smoothing and jagged artifacts, and have poor style adaptability.

[0004] 2. Reconstruction methods based on convolutional neural networks (CNN): These methods mainly use pixel-level reconstruction loss (L1 / L2) and perceptual loss, such as SRCNN, EDSR, and RCAN. They minimize the prediction residual and optimize through pixel-level loss. They have good structural restoration, but often lack detail and texture, are prone to over-smoothing, and have limited adaptability to data with diverse styles.

[0005] 3. Perception-driven methods based on adversarial learning: These methods jointly optimize adversarial loss and perceptual loss, such as SRGAN and ESRGAN. They can generate richer textures and better visual perception, but they are prone to introducing artifacts and style drift, which limits stability and reproducibility. They may also cause contour distortion when processing images with sharp lines.

[0006] 4. Image Sampling Upscaling Methods Based on Diffusion Models: These methods utilize conditional diffusion or latent space diffusion as their core, achieving detail synthesis and style preservation through a stepwise generation process involving noise injection and denoising. They can be combined with multimodal conditions such as text, semantics, and guide maps. While this method synthesizes details through an iterative denoising process and possesses strong generation capabilities, the randomness of sampling can easily lead to inconsistencies in content details and style. Furthermore, its direct application to ultra-high-resolution images faces computational challenges, making it difficult to generate ultra-high-resolution images in a single step. To overcome the limitations of generation size, existing techniques employ block-based / sliding window / overlapping fusion strategies. However, these strategies amplify the inherent randomness of diffusion models, leading to new problems such as semantic incoherence between blocks and inconsistent color and brightness.

[0007] Therefore, existing technologies have limitations in high-fidelity and high-consistency image super-resolution tasks, and there is an urgent need for a technical solution that can balance detail generation capability with global consistency, faithfully reproduce the style of the original image, and process large-size images. Summary of the Invention

[0008] To address the aforementioned problems, this invention provides a super-resolution method for high-resolution image generation that effectively solves issues such as detail style drift and semantic inconsistency in block processing in existing diffusion model amplification methods, achieving high-fidelity and high-consistency image super-resolution reconstruction.

[0009] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a super-resolution method for generating high-resolution images, comprising the following steps:

[0010] S1. Constructing a pairwise training dataset: Collect high-resolution image source datasets and clean them to obtain high-resolution images. Cut the high-resolution images into tiles with a fixed size and overlap rate to obtain high-quality image tiles. Perform degradation processing on each high-quality image tile to generate a corresponding low-quality image tile, and assign semantic text instructions to the pairwise tiles to form a triplet training sample of low-quality image + semantic text instruction + high-quality image, which is the pairwise training dataset. The low-quality image is the initial image, and the high-quality image is the target image.

[0011] S2: Conditional generation model training: Based on the DiT architecture, obtain triplet training samples, use the latent features of low-quality image tiles as structural conditions, use the semantic text instructions as semantic conditions, train a conditional generation model, the training objective of the model is to learn the mapping relationship from noisy state to the latent representation of the target high-quality image, continuously update the model parameters, and obtain the trained magnified model.

[0012] S3: Preprocessing of high-resolution input image to be magnified: Receive the high-resolution input image to be magnified, calculate the target size according to the target magnification factor and perform alignment adjustment, and then perform tile cutting on the adjusted image according to the fixed size and overlap rate parameters consistent with step S1 to obtain a set of low-quality image tile datasets to be processed, and record the cutting coordinates of each low-quality image tile.

[0013] S4: Block-based iterative denoising and amplification: Input the low-quality image tiles obtained in step S3 into the amplification model trained in step S2, and perform iterative denoising under structural conditions and preset or generated semantic instructions to obtain the corresponding high-quality image tiles after amplification.

[0014] S5: Color Correction and Fusion Stitching: Discrete wavelet transform is performed on all high-quality image tiles obtained in step S4 to extract low-frequency components and perform block color correction using the wavelet method; subsequently, based on the recorded cutting coordinates and the weight mask of each tile, the corrected tiles are weighted, fused, and stitched to generate the final high-resolution output image.

[0015] Preferably, in step S1, the degradation processing includes downsampling, blurring, color shifting, adding noise, and simulating JPEG compression artifacts; the semantic text instructions are used to describe the degradation type and / or the desired restoration effect.

[0016] Preferably, step S2 specifically includes:

[0017] S21: Perform latent encoding on the high-quality image tiles and low-quality image tiles in the triplet training samples respectively to obtain the target latent representation and the initial latent representation as structural conditions;

[0018] S22: Inject random noise into the latent representation of the target and construct a training target;

[0019] S23: At each time step, the current noisy target latent representation and the structured condition latent representation are concatenated for feature concatenation and used as the first layer input of the DiT network; the semantic text instruction is encoded and then integrated into the middle layer of the DiT network through an attention mechanism;

[0020] S24: The model predicts the noise residual based on the first layer input, time step information and fused semantic conditions, and optimizes it by minimizing the loss function between the predicted residual and the true residual to obtain an amplification model for block image amplification.

[0021] Preferably, in step S3, the alignment adjustment is to round up the calculated target size to an integer multiple of the fixed size of the tile to ensure the integrity of the cut.

[0022] Preferably, in step S5, the low-frequency component color correction in the discrete wavelet transform specifically involves using the low-frequency component of the high-resolution input image to be magnified in the corresponding tile region as a global reference benchmark.

[0023] Preferably, in step S5, the weight mask is a mask that decreases linearly from the center of the tile to the edge, and the attenuation region corresponds to the overlapping region between tiles.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] This invention significantly improves the style reproduction and feature consistency of magnified images through unified tiled training and degradation processing, conditional generation models, and wavelet domain color correction and weighted fusion in post-processing. It greatly enhances the stability of different blocks in block magnification, ensuring the integrity of the stitched image while adding a large amount of detail. It effectively suppresses semantic drift, color unevenness, and boundary artifacts caused by block processing, and significantly reduces detail decay (such as twisted line ends and kinking of multiple lines when they intersect). It broadens the application boundaries of high-fidelity image magnification, as it can adapt to input images of different sizes and freely select the magnification factor through block size conversion. Attached Figure Description

[0026] Figure 1 This is a schematic diagram comparing the effects of the method provided by the present invention in the embodiment (a in the figure is the high-resolution input image to be enlarged, and b is the high-resolution output image after enlargement).

[0027] Figure 2 This is a schematic diagram of the process of this invention. Detailed Implementation

[0028] The following will combine Figure 1-2 The present invention will be described in detail below. The illustrative embodiments and descriptions herein are used to explain the invention, but are not intended to limit the invention.

[0029] This invention aims to overcome the problems of semantic inconsistency, color drift, and detail distortion in existing diffusion model block-based upscaling methods, and provides a high-fidelity super-resolution solution. The core idea is to construct a pairwise dataset with unified tiled structures, train a conditional generative model, and introduce frequency domain color correction and intelligent fusion strategies after inference to ensure high consistency of the block processing results globally.

[0030] A super-resolution method for high-resolution image generation includes the following steps, such as... Figure 2 As shown:

[0031] S1. Constructing a pairwise training dataset: Collect high-resolution image source datasets and clean them to obtain high-resolution images. Cut the high-resolution images into tiles according to a fixed size (e.g., megapixels) and overlap rate to obtain high-quality image tiles. Perform degradation processing on each high-quality image tile to generate a corresponding low-quality image tile, and assign semantic text instructions to the pairwise tiles. Semantic text instructions can be obtained through manual annotation, template generation based on degradation parameters, or automatic generation using a multimodal large language model, or by directly using fixed instructions such as denoising and amplification, forming a triplet training sample of low-quality image + semantic text instruction + high-quality image, i.e., a pairwise training dataset, where the low-quality image is the initial image and the high-quality image is the target image.

[0032] The degradation processing includes, but is not limited to, combinations of downsampling, blurring (Gaussian blur, motion blur), color shift, sensor noise (adding Gaussian additive noise), and simulating JPEG compression artifacts. The type, order, and parameters of each degradation operation can be randomly configured to simulate diverse degradation conditions in real-world scenarios. Semantic text instructions are used to describe the degradation type and / or the desired restoration effect.

[0033] S2: Conditional Generation Model Training: Based on the DiT architecture, triplet training samples are obtained. The latent features of low-quality image tiles are used as structural conditions, and semantic text instructions are used as semantic conditions to train a conditional generation model. The conditional generation model is trained based on a diffusion model or flow matching model framework, with its backbone network being a Transformer architecture, U-Net architecture, or other suitable deep neural network architecture. The training objective of the conditional generation model is to learn the mapping relationship from noisy states to the latent representation of the target high-quality image. The model parameters are continuously updated to obtain a trained magnified model. The specific steps are as follows:

[0034] S21: Using a pre-trained variational autoencoder (VAE), the encoder performs latent encoding on high-quality and low-quality image tiles in the triplet training samples to obtain the target latent representation and the initial latent representation as structural conditions. The pre-trained VAE keeps its parameters frozen during the training and inference phases.

[0035] S22: Inject random noise into the latent representation of the target and construct a training target, wherein the noise injection and the construction of the training target are based on one or a combination of the following frameworks:

[0036] The Diffusion Probability Model (DDPM) framework, where the training objective is to predict the noise added to the model;

[0037] The Flow Matching framework, where the training objective is to predict the velocity field between the data and the noise;

[0038] Or other equivalent generative model training frameworks;

[0039] S23: At each time step, incorporate structural conditions into the generative model. The incorporation methods include, but are not limited to:

[0040] The channel dimension of the noisy target latent representation is concatenated and used as the input to the first layer of the DiT network, injected into the first layer of the DiT network through an additional control branch, or injected into the first layer of the DiT network through feature addition / modulation; semantic text instructions are encoded by a text encoder and then incorporated into the middle layer of the DiT network through a cross-attention mechanism or other conditional modulation methods.

[0041] S24: Based on the first layer input, time step information and fused semantic conditions, the model predicts the target noise residual corresponding to the training framework, and optimizes it by minimizing the loss function between the predicted residual and the real residual, thus obtaining an amplification model for block image amplification.

[0042] S3: Preprocessing of the high-resolution input image to be magnified: Receive the high-resolution input image to be magnified, calculate the target size according to the target magnification factor and perform alignment adjustment. This alignment adjustment is to round up the calculated target size to an integer multiple of the fixed tile size to ensure the integrity of the cutting. Then, according to the fixed size and overlap rate parameters consistent with step S1, the adjusted image is tiled to obtain a set of low-quality image tile datasets to be processed, and the cropping coordinates of each low-quality image tile are recorded for reconstruction. Then, latent space initialization (i.e., latent encoding) is performed, the low-quality image tiles are encoded and injected with controlled intensity noise. The initialization method includes: the controlled intensity noise is sampled from the middle time step, or the pure noise latent representation is sampled from the standard normal distribution.

[0043] S4: Block-based iterative denoising and amplification: After initialization, the core generation process is started. The low-quality image tiles obtained in step S3 are input into the amplification model trained in step S2. Under the structural conditions and preset or generated semantic instructions, iterative denoising is performed in combination with the sampler corresponding to the training framework. The sampler includes, but is not limited to, DDPM / DDIM sampling, ODE solvers (such as Euler method, Heun method) or other applicable sampling strategies. This process is repeated, and the denoised latent representation is restored to the pixel space through the VAE decoder to obtain the amplified high-quality image tiles.

[0044] S5: Color Correction and Fusion Stitching: Using the low-frequency components of the high-resolution input image to be magnified in the corresponding tile region as a global reference benchmark, discrete wavelet transform is performed on all high-quality image tiles obtained in step S4 to extract the low-frequency components and perform block color correction using the wavelet method; subsequently, based on the recorded cutting coordinates and the weight mask of each tile, the corrected tiles are weighted, fused, and stitched to generate the final high-resolution output image, where the weight mask is a mask that decays linearly from the center of the tile to the edge, and the decay region corresponds to the overlapping region between tiles.

[0045] In one embodiment, the super-resolution method for high-resolution image generation provided by the present invention mainly includes the following stages:

[0046] The first stage involves the construction and normalization of paired training datasets.

[0047] First, collect a large number of high-resolution image source datasets (e.g., comics, illustrations with over 8 million pixels) to build a training basis. Then, perform the following steps:

[0048] Unified tile cutting: Using the "overlapping uniform stepped tile cutting" algorithm, cut all high-resolution image source data into tiles of a fixed size (e.g., 1024×1024 pixels), given the original image size W×H and tile size w×h; if W < w or H < h, no cutting is performed. This algorithm ensures that the image is fully covered, and there is a controllable overlapping area between tiles (e.g., 10% overlap rate). This guarantees that the subsequent model can learn boundary context information, and finally obtain a tile dataset in millions of pixels.

[0049] Paired training data generation and degradation modeling: Apply composite degradation processing to each high-quality image tile (HR Tile) to generate the corresponding low-quality image tile (LR Tile). The degradation sequence includes: bicubic downsampling to the target low resolution (e.g., 256×256 pixels, corresponding to a 4-fold magnification), adding Gaussian noise with a standard deviation sigma randomly sampled from [0.02, 0.08], and simulating artifacts generated by JPEG compression quality Q randomly sampled from [25, 75]. At the same time, generate a semantic text instruction (e.g., "remove jpeg artifacts and noise") describing the degradation type or expected restoration effect for each (low-quality tile, high-quality tile) pair, thus forming a structured triple training sample (I_small, C_T, I_target). Before inputting into the model, upsample the low-quality image tile to the same size as the high-quality image tile to obtain consistent spatial dimensions for subsequent latent encoding.

[0050] Second stage: Training of the conditional generation model

[0051] The goal of this stage is to model the image conversion problem as a condition-driven residual prediction task, using the dit model with a multi-modal feature interface as the base to train an available magnification model.

[0052] Latent Encoding and Condition Extraction: A pre-trained Variational Autoencoder (VAE) encodes the target image I_target from the training triples into the latent space, obtaining the target latent representation z_0(I_target). Similarly, the VAE encodes the initial image I_small into the latent space, obtaining the initial latent representation z_0(I_small). The initial image I_small is then upsampled to the same spatial dimension as I_target using bicubic interpolation, and then encoded into the latent space again using the VAE, resulting in the initial latent representation z_0(I_small), making it have the same latent space dimension as z_0(I_target). The initial latent representation is used as the structural anchor, i.e., the structural condition C_I = z_0(I_small) is defined. Text instructions (and the optional initial image I_small) are input into an existing text (visual-language model) encoder to extract deep semantic embeddings, defined as semantic conditions C_T.

[0053] Noisy Modeling and Target Construction: Starting from time step t in the range [0, 1], noise epsilon is sampled from the standard normal distribution. Noise is injected into the target latent representation z_0(I_target) at time step t to obtain the noisy latent representation z_t at the current time.

[0054] z_t = (1 - t) * z_0(I_target) + t * epsilon;

[0055] Constructing the true residual of the training target: Calculate the true residual term R_target, which points from the latent representation of the target image to the noise. In the flow matching architecture, this residual term corresponds to the velocity vector:

[0056] R_target = epsilon - z_0(I_target).

[0057] Conditional Injection and Fusion: The structural condition C_I (i.e., z_0(I_small)) is concatenated with the current noisy latent representation z_t in the first layer of the DiT network using latent channel concatenation.

[0058] Input = Concat(z_t, C_I).

[0059] The semantic condition C_T is then injected into the intermediate layer of DiT. Through the joint self-attention mechanism, the latent image representation and the semantic condition C_T interact and fuse, thereby guiding the generation process to conform to the semantic requirements of the text instructions.

[0060] Loss calculation and optimization: Input the constructed input, time step t, and semantic condition C_T into the DiT model to obtain the predicted residual R_pred (only the part corresponding to z_t is taken), and calculate the mean squared error between the predicted residual R_pred and the true residual R_target as the loss function: L_total = ||R_pred - R_target||^2. Minimize this loss through backpropagation, update the model parameters, and obtain the amplified model.

[0061] Phase 3: Preprocessing and Tile Formation Before Reasoning

[0062] When a user inputs an image to be enlarged and specifies the magnification factor (e.g., 4x), the system first performs preprocessing, dividing the high-resolution input image into blocks and recording the cropping coordinates:

[0063] Size Calculation and Alignment: Based on the original image size (width W_orig, height H_orig) and the magnification factor, calculate the total number of target pixels P_target = W_orig × H_orig × scale², and calculate the new target size H_new = sqrt(P_target / A) and W_new = P_target / H_new while maintaining the aspect ratio A = W_orig / H_orig. To meet the alignment requirements of subsequent block segmentation and model processing (e.g., a multiple of 16), H_new and W_new are rounded and aligned. The original input image is then upsampled to the aligned target size (W_aligned × H_aligned) using bicubic interpolation, resulting in the upsampled image.

[0064] Consistent parameter tileization: Using the exact same target tile size (1024×1024 pixels) and overlap rate (10%) as in the first stage, the aligned image is cropped. If W < w or H < h, no cropping is performed, and the complete image is returned directly. The top-left corner coordinates (x_i, y_j) of each tile T_{i,j} in the original image are recorded to form a list of tiles to be processed. Finally, a set of tile datasets, all with a size of 1024x1024 (1MP), is obtained for subsequent model inference.

[0065] Fourth stage: Block-based iterative noise reduction and amplification

[0066] Using each low-quality image tile T_{i,j} obtained in the third stage as input, execute the following inference process independently:

[0067] Condition preparation: For each tile T_{i,j} to be processed, calculate the structural condition: C_I = VAE(T_{i,j}); the semantic condition C_T is calculated uniformly based on the global editing instructions provided by the user; and a noisy latent representation z_{t=1} is initialized from the standard normal distribution according to the target resolution.

[0068] Iterative generation: The dynamic displacement parameter *mu* is calculated based on the image resolution. A nonlinear transformation is performed on the uniform time-step sequence in the [0, 1] interval to generate a sampling time-step sequence t_N, t_{N-1}, ..., t_0. Multiple iterations are performed from t=1 to t=0 according to a predetermined sampler (such as DDIM or a flow-matching sampler). In each step, the currently generated latent representation *z_{t_i}* is concatenated with a fixed structural condition *C_I* (channel or sequence dimension), and input along with *C_T* into the trained DiT model. After processing via an attention mechanism, the predicted residual *R_i* is output, and the latent representation is updated according to the flow-matching formula.

[0069] z_{t_{i-1}} = z_{t_i} - (t_i - t_{i-1}) * R_i,

[0070] When iterating to t_0 approximately equal to 0, the final latent representation z_out is obtained, which is then restored to a high-quality image tile T'_{i,j} in pixel space by the VAE decoder.

[0071] Phase 5: Image Proofreading and Stitching

[0072] The sampled image and the original high-resolution input image to be enlarged are color-corrected, and then the complete image is reassembled based on the records and coordinates.

[0073] Wavelet domain color correction: Discrete wavelet transform (DWT, e.g., using Haar wavelets) is performed on each magnified high-quality image tile T'_{i,j}, decomposing it into low-frequency components (LL) and high-frequency components (LH, HL, HH). The low-frequency components mainly carry color and brightness information. To eliminate cross-block color differences and brightness unevenness that may occur due to independent block processing, we use the wavelet low-frequency components of the original input image (simply upsampled to the target size) at the corresponding (i,j) position as a global reference. The mean and variance of the low-frequency components LL_{i,j} of T'_{i,j} are aligned to this reference, while the high-frequency components generated by the model, which contain rich details and textures, are preserved. Finally, the color-corrected tiles T''_{i,j} are reconstructed using the corrected low-frequency components and the original high-frequency components through inverse discrete wavelet transform (IDWT). These corrected tiles will be used for the next stage of stitching.

[0074] Weighted fusion and stitching: Initialize a zero-weighted image canvas (Canvas) of size W_aligned × H_aligned to accumulate tile pixel values; and a zero-weighted weighted canvas (WeightMap) of the same size to accumulate the coverage weight of each pixel. Iterate through each color-corrected tile T''_{i,j} and its corresponding top-left corner coordinates (x_i, y_j):

[0075] a. Generate a weighted mask m_{tile} of the same size as the tile. This mask has a value of 1 in the center region of the tile, and its value decreases linearly from 1 to 0 in the edge overlapping region (e.g., a strip region with a width of overlap_rate * tile size / 2);

[0076] b. Determine the ROI (Region Area) of the tile coverage on the canvas: ROI = [x_i : x_i+w, y_j : y_j+h];

[0077] c. Multiply the pixel value of T''_{i,j} by its weight mask and add it to the corresponding position in the Canvas: Canvas[ROI] += T''_{i,j} * m_{tile};

[0078] d. Add the value of the weight mask to the corresponding position in WeightMap: WeightMap[ROI] += m_{tile}.

[0079] Normalized Output: After traversing all tiles, for each pixel of the Canvas, divide by the cumulative weight at the corresponding position in the WeightMap (adding a small constant to prevent division by zero) to complete the normalization. The final image obtained is a high-fidelity, seamlessly stitched, high-resolution output image I_final.

[0080] Please see Figure 1 The comparison of the results shows that, compared with the obvious inter-block color difference, structural distortion and semantic inconsistency produced by the traditional block diffusion amplification method, the result obtained by the method of this invention is highly consistent in global color and brightness, rich in detail and faithful to the style of the original image, and the block boundary transition is natural and smooth.

[0081] The technical solutions provided by the embodiments of the present invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the embodiments of the present invention. The descriptions of the embodiments above are only for helping to understand the principles of the embodiments of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the embodiments of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A super-resolution method for high-resolution image generation, characterized in that, Includes the following steps: S1. Constructing a pairwise training dataset: Collect high-resolution image source datasets and clean them to obtain high-resolution images. Cut the high-resolution images into uniform tiles according to fixed size and overlap rate to obtain high-quality image tiles. Each high-quality image tile is degraded to generate a corresponding low-quality image tile, and a semantic text instruction is assigned to the pair of tiles to form a triplet training sample of low-quality image + semantic text instruction + high-quality image, i.e., a pair training dataset. The low-quality image is the initial image, and the high-quality image is the target image. S2: Conditional generation model training: Based on the DiT architecture, obtain triplet training samples, use the latent features of low-quality image tiles as structural conditions, use the semantic text instructions as semantic conditions, train a conditional generation model, the training objective of the conditional generation model is to learn the mapping relationship from noisy state to the latent representation of the target high-quality image, continuously update the model parameters, and obtain the trained magnified model. S3: Preprocessing of high-resolution input image to be magnified: Receive the high-resolution input image to be magnified, calculate the target size according to the target magnification factor and perform alignment adjustment, and then perform tile cutting on the adjusted image according to the fixed size and overlap rate parameters consistent with step S1 to obtain a set of low-quality image tile datasets to be processed, and record the cutting coordinates of each low-quality image tile. S4: Block-based iterative denoising and magnification: Input the low-quality image tiles obtained in step S3 into the magnification model trained in step S2, perform iterative denoising, and obtain the corresponding high-quality image tiles after magnification. S5: Color correction and fusion stitching: Perform discrete wavelet transform on all high-quality image tiles obtained in step S4, extract low-frequency components, and combine the wavelet method to perform block color correction; Subsequently, based on the recorded cutting coordinates and the weight mask of each tile, the corrected tiles are weighted, fused, and stitched together to generate the final high-resolution output image.

2. The super-resolution method for high-resolution image generation according to claim 1, characterized in that: In step S1, the degradation processing includes downsampling, blurring, color shifting, adding noise, and simulating JPEG compression artifacts; the semantic text instructions are used to describe the degradation type and / or the desired restoration effect.

3. The super-resolution method for high-resolution image generation according to claim 1, characterized in that: Step S2 specifically includes: S21: Perform latent encoding on the high-quality image tiles and low-quality image tiles in the triplet training samples respectively to obtain the target latent representation and the initial latent representation as structural conditions; S22: Inject random noise into the latent representation of the target and construct a training target; S23: At each time step, the current noisy target latent representation and the structured condition latent representation are concatenated for feature concatenation and used as the first layer input of the DiT network; the semantic text instruction is encoded and then integrated into the middle layer of the DiT network through an attention mechanism; S24: The model predicts the noise residual based on the first layer input, time step information and fused semantic conditions, and optimizes it by minimizing the loss function between the predicted residual and the true residual to obtain an amplification model for block image amplification.

4. The super-resolution method for high-resolution image generation according to claim 1, characterized in that: In step S3, the alignment adjustment involves rounding the calculated target size up to an integer multiple of the tile's fixed size to ensure the integrity of the cut.

5. The super-resolution method for high-resolution image generation according to claim 1, characterized in that: In step S5, the low-frequency component color correction in the discrete wavelet transform specifically involves using the low-frequency component of the high-resolution input image to be magnified in the corresponding tile region as a global reference benchmark.

6. The super-resolution method for high-resolution image generation according to claim 1, characterized in that: In step S5, the weight mask is a mask that decays linearly from the center of the tile to the edge, and the decay region corresponds to the overlapping region between tiles.

Images