Super-resolution image reconstruction method, system and device based on dynamic frequency domain adaptive coding and contrast constraint optimization, and medium

Abstract

The invention discloses a super-resolution image reconstruction method, system and device based on dynamic frequency domain adaptive coding and contrast constraint optimization, and a medium. The method comprises the following steps: creating an original training data set; image data features and application scenes are considered, and feature extraction is performed on image data by adopting a modular design method; in combination with discrete cosine transform and inverse transform thereof, adaptive transform coding is carried out on different frequency components in the image features through a dynamic weight adjustment mechanism; the U-Net and Transform architectures are combined, isotropic design is optimized, a diffusion model and a feature fusion and reconstruction technology are applied, and diffusion process reconstruction is achieved; by introducing a contrast loss function, contrast learning is carried out between images with different resolutions, and the feature learning ability is enhanced. The image reconstruction method is more intelligent and efficient, and can effectively deal with frequency feature recovery and detail reconstruction problems in a complex image scene.

Description

Technical Field

[0001] The present invention relates to the field of AI computer vision technology, and in particular to a super-resolution image reconstruction method, system, device and medium based on dynamic frequency domain adaptive coding and contrast constraint optimization. Background Art

[0002] In recent years, with the continuous advancement of image processing technology and the rapid growth of digitalization needs, image super-resolution technology has gradually become an important means to improve image quality, especially in the fields of mobile photography, security monitoring, medical imaging, etc. High-quality images have become a common demand of people. Clear and delicate image quality can not only improve the visual effect of photographic works, but also improve recognition accuracy in security monitoring and have decisive significance in medical diagnosis. However, due to the limitations of sensor resolution, environmental noise and information loss during the acquisition process, images often have low resolution problems, resulting in blurred details and affecting the use value of the image. Therefore, the development of effective super-resolution technology is of great significance to the improvement of image quality.

[0003] In the field of image super-resolution (SR), traditional methods mainly include interpolation-based methods and learning-based deep neural network methods. Interpolation methods, such as bilinear interpolation and bicubic interpolation, usually generate higher-resolution images by spatially interpolating the pixels of low-resolution images. However, such methods have limited ability to restore details and often result in blurred images.

[0004] In recent years, with the rapid development of deep learning technology, super-resolution methods have gradually become the mainstream in the field of image processing, especially in improving image details and quality, and have made significant breakthroughs. Among them, the application of diffusion models and transformers is particularly prominent, which has promoted the development of image super-resolution technology.

[0005] Diffusion models demonstrate powerful reconstruction capabilities by treating the generation process of high-resolution images as a gradual denoising process, using multiple time steps to repair and enhance details of images. These models have shown significant advantages in image generation tasks and can generally be divided into two categories: methods trained from scratch and methods based on prior knowledge. Methods trained from scratch learn image generation directly from noise. This method has high flexibility and adaptability and is suitable for a variety of task scenarios. Methods based on prior knowledge rely on existing generation experience to accelerate the training process and improve generation quality.

[0006] SR3 (Image Super-Resolution via Iterative Refinement) is one of the first studies to successfully apply diffusion models to image super-resolution tasks. It achieves gradual restoration and detail enhancement of low-resolution images through multiple iterative denoising processes. ResShift (Efficient Diffusion Model for Image Super-resolution by Residual Shifting) makes innovative improvements to the diffusion process, shortening the traditional Markov chain and reducing the denoising steps to 15, which significantly improves the inference speed and computational efficiency.

[0007] Inspired by the great success of Stable Diffusion in text-to-image generation tasks, some super-resolution methods have begun to adopt generative priors to improve the super-resolution performance of real-world images. These methods combine the advantages of diffusion models and pre-trained models to effectively improve the reconstruction quality and generation effect. DiT-SR (Diffusion Transformer for Super-Resolution) is a state-of-the-art diffusion transformer architecture that combines multi-scale feature extraction and isotropic design of U-shaped architecture, and adopts a unified transformer block structure for multi-stage processing, thereby achieving effective recovery of image details without relying on pre-trained models. The design of DiT-SR adopts an adaptive frequency modulation (Adaptive Frequency Modulation, AdaFM) module to dynamically adjust the feature frequency at different time steps, thereby enhancing the recovery of high-frequency details. However, DiT-SR has limitations in encoding different frequency features, especially in the encoding process, the fixed encoding scheme cannot adapt to the diversity of input image frequency characteristics.

[0008] Although existing image super-resolution technologies have made some progress, they still face many challenges, including:

[0009] Insufficient frequency feature perception: Although the existing DiT-SR introduces a frequency-adaptive time step adjustment module, its ability to restore high-frequency details is still insufficient, especially when dealing with complex textures and detail areas. The fixed encoding scheme is difficult to fully capture high-frequency information; this results in the problem of blurring or missing details in the high-frequency area of ​​the generated image;

[0010] The model lacks flexibility in frequency encoding: The existing frequency-adaptive time-step adjustment module mainly performs feature modulation in the channel dimension, but fails to dynamically balance the encoding of different frequency features. This makes the model lack sufficient flexibility and adaptability when processing images of different complexities, and it is unable to effectively adjust the contribution of frequency features at each time step.

[0011] Insufficient high-quality generation capabilities: In the process of generating high-resolution images, most existing diffusion models rely on the process of gradually restoring low-frequency to high-frequency information. This approach makes the model fail to fully utilize the detail information in the high-resolution image in different stages of denoising, resulting in low-quality generation results and detailed features that are difficult to achieve visual satisfaction. Summary of the invention

[0012] In the image super-resolution task, the existing diffusion model is usually difficult to achieve ideal results in the restoration and generation quality of high-frequency details. In particular, when processing images with complex textures and details, the model's perception and flexibility of high-frequency features are insufficient, resulting in the problems of detail loss, blurring and low visual quality in the generated results. The purpose of the present invention is to provide a super-resolution image reconstruction method, system, device and medium based on dynamic frequency domain adaptive coding and contrast constraint optimization to address the above problems.

[0013] To achieve the above purpose, the technical solution provided by the present invention is:

[0014] The first aspect of the present application provides a super-resolution image reconstruction method based on dynamic frequency domain adaptive coding and contrast constraint optimization, comprising the following steps:

[0015] Create an original training data set, use data enhancement methods to improve sample diversity for image data in the original training data set, and process image data using lightweight data preprocessing methods;

[0016] Considering the characteristics and application scenarios of different types of image data, a modular design method that can be flexibly adjusted and expanded is used to extract features from image data;

[0017] Combining discrete cosine transform and its inverse transform, adaptive transform coding is performed on different frequency components in image features through dynamic weight adjustment mechanism;

[0018] Combining U-Net and Transformer architecture, optimizing isotropic design, applying diffusion model and feature fusion and reconstruction technology, and realizing diffusion process reconstruction;

[0019] By introducing a contrast loss function to perform contrast learning between images of different resolutions, the feature learning ability is enhanced.

[0020] In the feature extraction step, by adjusting the depth, width and configuration of the residual block of the convolutional layer, it is ensured that useful features can be efficiently extracted in a variety of application scenarios; according to the specific application scenario, different activation functions or normalization strategies are customized to adapt to specific task requirements.

[0021] In the feature extraction function, different levels adjust the size of the convolution kernel according to the feature extraction requirements; ReLU is used as the activation function; batch normalization is applied after each convolution layer; in the residual connection design, cross-layer connection is used in each residual block to add the input feature map to the convolution output; the upsampling method uses transposed convolution; through the U-shaped encoder-decoder structure, rich feature information is extracted at multiple scales; the ability to understand complex image structures is enhanced through multi-level feature representation; the depth and breadth of the network are improved by introducing residual blocks, and the detail capture ability of feature extraction is enhanced; the consumption of computing resources is controlled through downsampling and upsampling operations; and computing efficiency is optimized through residual connections and batch normalization.

[0022] In the adaptive transform coding step, the transformation matrix is ​​used to transform the input image features into the frequency domain, converting the spatial domain information into a frequency domain representation; combining the features of the input image data, a frequency mask is dynamically generated to selectively enhance or suppress specific frequency components; the generated frequency mask is applied to perform targeted encoding and decoding processing on each frequency component in the frequency domain; the features processed in the frequency domain are transformed back to the spatial domain through an inverse DCT transform and fused with the original features.

[0023] In the diffusion process reconstruction step, the U-Net architecture is combined with the Transformer model to form an enhanced multi-scale Transformer structure, including: Encoder: composed of multiple layers of down-sampling Transformer modules, gradually extracting image features of different scales; Decoder: composed of multiple layers of up-sampling Transformer modules, fusing the features of the corresponding encoder layers through jump connections to restore high-resolution images; Multi-scale feature fusion: using jump connections and multi-scale attention mechanisms to achieve feature fusion and enhance the ability to capture details; Transformer optimization: using layered and lightweight design to improve the efficiency and effect of multi-scale feature extraction;

[0024] The optimized isotropic design includes: the Transformer module structure and parameters of each stage are consistent to achieve parameter sharing and cross-scale feature transfer; all Transformer modules have unified channel dimensions and number of layers to simplify feature conversion and improve computational efficiency; high-resolution layers are preferentially allocated more computing resources, low-resolution layers are appropriately reduced in resources, and overall computing allocation is optimized; channel dimensions are standardized according to different resolution stages to balance high-frequency detail capture and computational complexity.

[0025] The diffusion model generates a high-resolution image by gradually denoising, including: gradually adding noise in the forward process, and gradually restoring the image through the denoiser in the reverse process; constructing a residual Markov chain between high-resolution and low-resolution images, reducing the number of diffusion steps and speeding up the generation speed;

[0026] The feature fusion and reconstruction technology includes: extracting and encoding multi-scale features of low-resolution images through the AdaTCM encoder to model global context information; adopting feature fusion methods, including using feature alignment strategies to unify the spatial and channel dimensions of features of different scales; using weighted fusion strategies to fuse high- and low-resolution features through learnable weight parameters to highlight high-frequency information; using multi-layer fusion strategies to integrate detail information at different levels at multiple levels; adopting high-frequency feature enhancement methods, including using frequency domain conversion strategies to emphasize high-frequency components through frequency domain filters, and using residual learning strategies to directly introduce high-frequency information using residual connections to avoid information loss; adopting reconstruction process optimization methods, including parallel computing and lightweight module design to improve computing efficiency, combining residual shifting methods to shorten the number of reconstruction steps, and performing module integration and dynamic adjustment to ensure adaptation to different image characteristics.

[0027] The contrastive learning steps include: using a feature extraction method to input a low-resolution image and generate a high-resolution image through a super-resolution generator; using a feature extraction network to extract features from the low-resolution image, the generated high-resolution image and the target high-resolution image to obtain their respective feature representations; using a positive and negative sample construction method to construct positive sample pairs and negative sample pairs; using a contrastive loss calculation method to measure the relationship between positive sample pairs and negative sample pairs by calculating the similarity between features and using a contrastive loss function to guide model optimization.

[0028] A second aspect of the present application provides a super-resolution image reconstruction system based on dynamic frequency domain adaptive coding and contrast constraint optimization, characterized in that it includes:

[0029] The data preprocessing module is used to create the original training data set, use the data enhancement method to improve the diversity of samples in the image data of the original training data set, and process the image data through the lightweight data preprocessing method;

[0030] The multi-scale feature extraction module is used to consider the characteristics and application scenarios of different types of image data, and extract features from image data using a modular design method that can be flexibly adjusted and expanded;

[0031] Adaptive transform coding module, used to combine discrete cosine transform and its inverse transform, and perform adaptive transform coding on different frequency components in image features through dynamic weight adjustment mechanism;

[0032] The diffusion process reconstruction module is used to combine the U-Net and Transformer architectures, optimize the isotropic design, and apply the diffusion model and feature fusion and reconstruction technology to achieve diffusion process reconstruction;

[0033] Contrastive learning module, which is used to enhance feature learning ability by introducing contrast loss function to perform contrast learning between images of different resolutions;

[0034] The third aspect of the present application provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the super-resolution image reconstruction method based on dynamic frequency domain adaptive coding and contrast constraint optimization is implemented.

[0035] A fourth aspect of the present application provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the above-mentioned super-resolution image reconstruction method based on dynamic frequency domain adaptive coding and contrast constraint optimization.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The present invention provides a more intelligent and efficient image reconstruction system, which can effectively deal with the frequency feature recovery and detail reconstruction problems in complex image scenes; by introducing an adaptive transform coding module and a contrast learning module, combined with multi-scale feature extraction and diffusion process reconstruction technology, the generation quality of image super-resolution is significantly improved, and the shortcomings of the prior art in high-frequency detail recovery and frequency coding flexibility are solved. The solution of the present invention has the following advantages:

[0038] Enhance the ability to restore high-frequency details: Introduce adaptive transform coding technology and contrastive learning technology to dynamically adjust and efficiently extract image features of different frequencies, enhance the diffusion model's perception of different frequency features and improve the quality of generation, significantly improve the capture and restoration of high-frequency details, and reduce detail loss and blurring. Adaptive transform coding technology uses dynamic weight adjustment to encode different frequency features, thereby enhancing the model's frequency perception ability and avoiding the limitations of fixed coding schemes; contrastive learning technology introduces contrastive learning loss between low-resolution images and generated high-resolution images, allowing the model to more effectively learn the detailed features in high-resolution images, thereby significantly improving the visual quality of image reconstruction.

[0039] Improve the visual quality of generated images: Combine diffusion process reconstruction technology with contrast learning technology to optimize feature fusion and detail restoration capabilities in the image generation process, ensuring that the generated high-resolution images reach a high level in detail retention and overall visual quality.

[0040] Improve the flexibility and scalability of the model: Through the modularly designed multi-scale feature extraction method and adaptive transform coding technology, it supports flexible adjustment and expansion according to different application requirements, adapts to the processing requirements of various types of images, and enhances the applicability and scalability of the system.

[0041] Optimize computing efficiency: Through isotropic design optimization, residual shifting method and frequency domain processing strategy, reduce computing resource consumption, improve overall computing efficiency, and ensure the real-time and high efficiency of the module in high-resolution image processing.

[0042] Enhance the generalization ability of the model: Through multi-task loss function and contrastive learning mechanism, the generalization ability of the model in different data sets and application scenarios is improved to ensure excellent performance in a variety of complex image scenes.

[0043] In summary, the present invention significantly improves the generation quality of image super-resolution through the organic combination of multi-scale feature extraction, adaptive transform coding, diffusion process reconstruction and contrast learning, solves the shortcomings of the prior art in high-frequency detail recovery and frequency coding flexibility, and provides an innovative solution for the development of image super-resolution technology; this method can not only significantly improve the denoising and reconstruction effects, but also ensure the naturalness and consistency of the image while retaining the image details, and has broad application prospects and significant practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 : A structural schematic diagram of a super-resolution image reconstruction system based on dynamic frequency domain adaptive coding and contrast constraint optimization of the present invention.

[0045] Figure 2 : Schematic diagram of the overall framework of the super-resolution image reconstruction system based on dynamic frequency domain adaptive coding and contrast constraint optimization of the present invention.

[0046] Figure 3 : Schematic diagram of the Transformer module structure.

[0047] Figure 4 : Schematic diagram of the adaptive transform coding module structure. DETAILED DESCRIPTION

[0048] The above contents of the present invention are further described in detail below in the form of embodiments, but this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments, and all technologies realized based on the above contents of the present invention belong to the scope of the present invention.

[0049] Some of the terms in the present invention are explained as follows:

[0050] Diffusion models: Diffusion models are a type of generative model that generates new data samples by gradually adding noise to the data during training and then gradually removing the noise during generation; this type of model was originally used for unsupervised learning, but in recent years it has been increasingly used in the field of image generation and enhancement, especially in tasks such as denoising, super-resolution, and image restoration.

[0051] Stable Diffusion: Stable Diffusion is a text-to-image generation model based on the diffusion model. It gradually transforms noisy images into clear images through the diffusion process. The generated results are of high quality and suitable for multiple application scenarios, such as image generation, restoration, super-resolution, etc.

[0052] DCT: Discrete Cosine Transform. DCT is a transform method that converts signals to the frequency domain and is commonly used in image compression (such as JPEG) and signal processing. DCT can compress image data and reduce visual loss by decomposing the image into a series of cosine waves.

[0053] Residual Shifting: Residual shifting is a technique for optimizing a model by adjusting the residuals of different layers in a neural network. It can reduce the bias in feature maps, enhance the representation ability of the network, and improve the effect of image reconstruction tasks, especially in super-resolution and denoising tasks.

[0054] RealESRGAN: RealESRGAN is a generative adversarial network (GAN) model for image super-resolution, which can reconstruct low-resolution images into high-definition images. The design goal of RealESRGAN is to cope with complex scenes of real-world images and generate more natural details, which is suitable for tasks such as image restoration and enlargement.

[0055] LSDIR dataset: LSDIR is a large-scale dataset for image restoration tasks, covering various types of degraded images. The dataset provides rich training and test data for studying image restoration algorithms, covering tasks such as denoising, deblurring, and artifact removal.

[0056] DIV2K dataset: DIV2K is a high-quality dataset for image super-resolution tasks, containing 1,000 high-resolution images (2K resolution); this dataset is widely used in research and competitions in the field of image super-resolution, and provides a variety of image degraded versions for model training and testing.

[0057] DIV8K dataset: DIV8K is an extended version of DIV2K and contains high-quality images with 8K resolution. This dataset provides higher resolution image resources and is suitable for research on super-resolution, image enhancement, and high-precision visual tasks.

[0058] Flicker2K dataset: Flickr2K is a dataset of 2,000 high-resolution images collected from the Flickr website. It is commonly used in image super-resolution tasks. It provides diverse scene images as a supplement to the DIV2K dataset for training and evaluating super-resolution algorithms.

[0059] FFHQ dataset: The FFHQ dataset is a high-quality face image dataset containing 70,000 face photos of different ages, races, and styles with a resolution of up to 1024×1024; this dataset is widely used in research related to generative adversarial networks (GANs), especially face generation and editing tasks.

[0060] The present invention provides a super-resolution image reconstruction method based on dynamic frequency domain adaptive coding and contrast constraint optimization, comprising the following steps:

[0061] Create an original training data set, use data enhancement methods to improve sample diversity for image data in the original training data set, and process image data using lightweight data preprocessing methods;

[0062] Considering the characteristics and application scenarios of different types of image data, a modular design method that can be flexibly adjusted and expanded is used to extract features from image data;

[0063] Combining discrete cosine transform and its inverse transform, adaptive transform coding is performed on different frequency components in image features through dynamic weight adjustment mechanism;

[0064] Combining U-Net and Transformer architecture, optimizing isotropic design, applying diffusion model and feature fusion and reconstruction technology, and realizing diffusion process reconstruction;

[0065] By introducing a contrast loss function to perform contrast learning between images of different resolutions, the feature learning ability is enhanced.

[0066] Among them, data preprocessing is performed first, including the training stage and the inference stage:

[0067] Training phase: During the training phase, we provide an original training dataset for the super-resolution image reconstruction model, denoted as D. In order to enhance the generalization ability of the model in complex scenes and ensure that it can cope with low-resolution images of different types and qualities, we carefully designed a series of data augmentation strategies. These strategies include but are not limited to operations such as image rotation, flipping, scaling, and random cropping. Through a variety of augmentation methods, the diversity of training samples is significantly improved, thereby improving the generalization ability and robustness of the model. These data augmentation measures enable the model to better learn useful features when facing complex scenes and different types of low-resolution images, thereby improving the effect of super-resolution reconstruction.

[0068] Inference stage: The design goal of the data preprocessing module in the inference stage is to efficiently process the input image and maximize the utilization of hardware resources to ensure that the inference stage achieves the best balance between computing speed and output image quality. To this end, we introduced some lightweight data preprocessing steps in the inference stage to reduce the overhead of redundant calculations. These processes not only optimize the quality of the input data, but also effectively accelerate the inference process, thereby improving the response speed and stability of the model in practical applications while ensuring the image super-resolution reconstruction effect.

[0069] Feature extraction adopts a modular multi-scale feature extraction method, which is convenient for seamless integration with other image processing modules. The modular structure supports flexible adjustment and expansion according to specific application requirements, such as adding more coding levels or adjusting the number of filters. The module design takes into account the feature requirements of different types of images (such as natural images, medical images, satellite images, etc.). By adjusting the depth and width of the convolution layer and the configuration of the residual block, it is ensured that useful features can be efficiently extracted in a variety of application scenarios. In addition, different activation functions or normalization strategies can be customized according to specific application scenarios to meet specific task requirements. The modular design enhances the flexibility and scalability of the system and facilitates the integration and optimization of subsequent functional modules. Whether it is customized adjustment for specific application scenarios or combined with other advanced technologies, the module can achieve efficient adaptation. Through multi-task loss functions and advanced optimization algorithms, the module can effectively improve the accuracy and generalization ability of feature extraction during training, ensuring excellent performance in different data sets and application scenarios.

[0070] In terms of technical implementation details, first of all, for the selection of convolution kernel size and stride, different levels adjust the size of convolution kernel according to feature extraction requirements. Generally, 3×3 convolution kernel is used to balance computational complexity and feature capture capability. The stride is usually set to 1 to maintain the spatial resolution of the feature map, unless it is adjusted during downsampling and upsampling operations. In terms of activation function and normalization, ReLU is mainly used as the activation function because it is simple to calculate and can effectively alleviate the gradient vanishing problem. In addition, batch normalization is applied after each convolution layer to stabilize the training process and speed up convergence. In the residual connection design, cross-layer connections are used in each residual block to add the input feature map to the convolution output. Such a design improves the trainability of the network and enhances the ability to transfer and reuse features. The upsampling method mainly uses transposed convolution because it can learn the weights of upsampling to generate higher quality feature maps.

[0071] The multi-scale feature extraction module also has a rich multi-level feature representation. Through the U-shaped encoder-decoder structure, it can extract rich feature information at multiple scales to meet the image detail requirements at different levels. The multi-level feature representation enhances the model's ability to understand complex image structures and improves the performance of subsequent processing modules. The introduction of residual blocks not only improves the depth and breadth of the network, but also significantly enhances the detail capture capability of feature extraction, ensuring that the tiny features and details of the image are fully preserved and expressed during the multi-scale feature extraction process. Through downsampling and upsampling operations, the consumption of computing resources is effectively controlled. At the same time, the computing efficiency is optimized through residual connections and batch normalization, ensuring the real-time and high efficiency of the module in high-resolution image processing.

[0072] In the adaptive transform coding step, the discrete cosine transform (DCT) and its inverse transform are combined. Through the dynamic weight adjustment mechanism, the different frequency components in the image features are efficiently encoded, realizing the adaptive processing capability of each frequency component in the feature map.

[0073] By introducing a learnable transformation matrix through adaptive transform coding (AdaTCM), the weight of each frequency component can be dynamically adjusted according to the specific content of the input feature map. This mechanism enables the model to more effectively enhance the reconstruction ability of high-frequency details during the feature diffusion process, thereby improving the fineness and clarity of the image. The core workflow is as follows:

[0074] (1) Frequency domain transformation: First, AdaTCM uses the transformation matrix to transform the input image features into the frequency domain, converting the spatial domain information into a frequency domain representation so that different frequency components can be processed independently.

[0075] (2) Dynamic frequency mask generation: Based on the features of the input image, AdaTCM dynamically generates a frequency mask to selectively enhance or suppress specific frequency components. The frequency mask adaptively adjusts the weight distribution of different frequencies according to the complexity and detail requirements of the image content.

[0076] (3) Frequency encoding and decoding: By applying the generated frequency mask, AdaTCM performs targeted encoding and decoding on each frequency component in the frequency domain. This process ensures that high-frequency details and complex textures can be fully reconstructed, while low-frequency information is effectively preserved, avoiding image distortion caused by over-enhancement.

[0077] (4) Inverse transformation and feature fusion: The features processed in the frequency domain are transformed back to the spatial domain through inverse DCT and fused with the original features to further improve the overall quality and detail performance of the image.

[0078] Through the above process, AdaTCM realizes adaptive frequency encoding of input image features, which not only enhances the reconstruction effect of high-frequency details and complex textures, but also improves the generalization ability of the model under different image contents. The design of this module makes the super-resolution image reconstruction process more intelligent and efficient, and significantly improves the visual quality and detail richness of the reconstructed image. In addition, the AdaTCM module also has the following advantages:

[0079] High flexibility: The learnable transformation matrix enables the module to automatically adjust the frequency encoding strategy according to different image characteristics and adapt to diverse image content requirements.

[0080] Excellent computational efficiency: Through frequency domain processing, redundant computations in the spatial domain are reduced, the overall computational efficiency is improved, and it is suitable for real-time image processing applications.

[0081] Modular design: As an independent module, AdaTCM can be easily integrated into the existing image reconstruction framework to enhance the overall performance of the system.

[0082] In summary, the adaptive transform coding module (AdaTCM) significantly improves the effect of super-resolution image reconstruction through dynamic frequency domain adaptive coding technology, which provides strong technical support for the application of the present invention in the field of image processing.

[0083] In the diffusion process reconstruction process, by combining U-Net and Transformer architecture, optimizing isotropic design, applying diffusion model and efficient feature fusion and reconstruction technology, efficient and high-quality image super-resolution reconstruction is achieved, which has the significant advantages of improving model performance, reducing computing resource consumption and enhancing detail capture capabilities. The core of the diffusion process reconstruction process includes:

[0084] Combination of U-Net and Transformer: The U-Net architecture is combined with the Transformer model to form an enhanced multi-scale Transformer structure to improve image super-resolution performance, including:

[0085] Encoder: It consists of multiple layers of downsampling Transformer modules, which gradually extract image features of different scales, similar to the downsampling path of U-Net.

[0086] Decoder: It consists of multiple layers of upsampling Transformer modules, which fuse the features of the corresponding encoder layers through skip connections to restore high-resolution images.

[0087] Multi-scale feature fusion: Utilize skip connections and multi-scale attention mechanisms to achieve efficient feature fusion and enhance the ability to capture details.

[0088] Transformer optimization: adopts layered and lightweight design to improve the efficiency and effect of multi-scale feature extraction.

[0089] The optimization of isotropic design means that by optimizing the isotropic design, the model has consistent feature processing capabilities at different scales and levels, including:

[0090] Consistent structure: The Transformer module structure and parameters are consistent at each stage, achieving parameter sharing and cross-scale feature transfer.

[0091] Unified channel dimension and depth: All Transformer modules have unified channel dimension and number of layers, which simplifies feature conversion and improves computational efficiency.

[0092] Computing resource optimization: More computing resources are allocated to high-resolution layers first, and resources are appropriately reduced for low-resolution layers to optimize overall computing allocation.

[0093] Standardized channel configuration: Standardize channel dimensions according to different resolution stages to balance high-frequency detail capture and computational complexity.

[0094] The advantages of isotropic design optimization are: improving model capacity, reducing the number of parameters, simplifying the design process, and enhancing scalability.

[0095] In the diffusion module, the diffusion process is defined as follows: the forward process gradually adds noise, and the reverse process gradually restores the image through the denoiser.

[0096] The diffusion module generates high-resolution images by progressive denoising, including:

[0097] Residual Shifting (ResShift): Constructs a residual Markov chain between high-resolution and low-resolution images to reduce the number of diffusion steps and speed up the generation.

[0098] The advantages of the diffusion module are: shortening the number of steps and increasing the generation speed while maintaining high-quality reconstruction.

[0099] Feature fusion and reconstruction technology is a key step to ensure the generation of high-quality high-resolution images, covering:

[0100] (1) AdaTCM encoder: extracts and encodes multi-scale features of low-resolution images and models global context information.

[0101] (2) Feature fusion strategy:

[0102] Feature alignment: unify the spatial and channel dimensions of features of different scales.

[0103] Weighted fusion: Fusion of high- and low-resolution features through learnable weight parameters to highlight high-frequency information.

[0104] Multi-layer fusion: Integrate different levels of detail information at multiple levels.

[0105] (3) High-frequency feature enhancement:

[0106] Frequency domain conversion: Emphasize high-frequency components through frequency domain filters.

[0107] Residual learning: Use residual connections to directly introduce high-frequency information to avoid information loss.

[0108] (4) Reconstruction process optimization:

[0109] Parallel computing and lightweight module design improve computing efficiency.

[0110] Combined with the Residual Shifting method, the number of reconstruction steps can be shortened and the generation can be accelerated.

[0111] Module integration and dynamic adjustment ensure adaptation to different image characteristics and optimize reconstruction effects.

[0112] In the contrastive learning step, by introducing the contrastive loss function, the model's feature learning ability between images of different resolutions is enhanced, thereby improving the quality and detail restoration effect of super-resolution images. This module is integrated in the super-resolution image reconstruction model and mainly includes the following functional units: Feature extraction unit: responsible for extracting the feature representation of low-resolution images, generated high-resolution images, and target high-resolution images. Positive and negative sample construction unit: used to construct positive sample pairs and negative sample pairs required for contrastive learning. Contrast loss calculation unit: calculates the contrastive loss function to guide the optimization direction of the model. Optimization update unit: combines the reconstruction loss and contrast loss to optimize and update the model parameters. Specifically:

[0113] Feature extraction unit:

[0114] The input low-resolution image is passed through the super-resolution generator to generate a high-resolution image.

[0115] A feature extraction network is used to extract features from the low-resolution image, the generated high-resolution image, and the target high-resolution image to obtain their respective feature representations.

[0116] Positive and negative sample building blocks:

[0117] Positive pairs: feature pairs of the generated high-resolution image and the target high-resolution image, which are expected to be highly similar in feature space.

[0118] Negative sample pairs: feature pairs of generated high-resolution images and original low-resolution images, which are expected to be significantly different in feature space.

[0119] Contrastive loss calculation unit:

[0120] By calculating the similarity between features, the contrast loss function is used to measure the relationship between positive sample pairs and negative sample pairs to guide model optimization.

[0121] Optimize update unit:

[0122] A reconstruction loss (such as mean squared error or perceptual loss) is calculated to measure the difference between the generated image and the target image.

[0123] The reconstruction loss is combined with the contrast loss to form a total loss function, and the model parameters are updated through the optimization algorithm to minimize the total loss.

[0124] By introducing the contrastive learning module, the model has been significantly enhanced in feature learning. The contrastive loss function helps the model better learn the feature relationship between images of different resolutions, thereby improving the ability to restore details of the generated image; the introduction of multi-scale feature fusion and attention mechanism enables the model to capture more levels of information and enhance the richness of feature representation; the negative sample expansion strategy and optimized parameter adjustment method further improve the effect of contrastive learning, ensuring that the model has higher performance and efficiency in super-resolution reconstruction tasks. The contrastive learning module (ConLM) significantly improves the reconstruction quality and detail restoration ability of super-resolution images through the collaborative work of multiple functional units such as feature extraction, sample construction, contrastive loss calculation and optimization update, ensuring that the model shows higher performance and efficiency in practical applications.

[0125] The present invention also provides a super-resolution image reconstruction system based on dynamic frequency domain adaptive coding and contrast constraint optimization, such as Figure 1 As shown, including:

[0126] The data preprocessing module is used to create the original training data set, use the data enhancement method to improve the diversity of samples in the image data of the original training data set, and process the image data through the lightweight data preprocessing method;

[0127] The multi-scale feature extraction module is used to consider the characteristics and application scenarios of different types of image data, and extract features from image data using a modular design method that can be flexibly adjusted and expanded;

[0128] Adaptive transform coding module (AdaTCM), which combines discrete cosine transform and its inverse transform to adaptively transform different frequency components in image features through dynamic weight adjustment mechanism;

[0129] The diffusion process reconstruction module is used to combine the U-Net and Transformer architectures, optimize the isotropic design, and apply the diffusion model and feature fusion and reconstruction technology to achieve diffusion process reconstruction;

[0130] The contrastive learning module is used to enhance feature learning capabilities by introducing a contrastive loss function to perform contrastive learning between images of different resolutions.

[0131] Among them, the multi-scale feature extraction module adopts a modular design, which can be seamlessly integrated with other image processing modules and supports flexible adjustment and expansion according to specific application requirements. By adjusting the depth, width and configuration of the residual block of the convolution layer, it ensures efficient extraction of useful features in different types of images such as natural images, medical images, satellite images, etc. The modular structure enhances the flexibility and scalability of the system, and combines the multi-task loss function and advanced optimization algorithm to improve the accuracy and generalization ability of feature extraction.

[0132] The adaptive transform coding module combines discrete cosine transform (DCT) and its inverse transform, and through a dynamic weight adjustment mechanism, efficiently encodes different frequency components in image features, achieving adaptive processing capabilities for each frequency component in the feature map. The learnable transformation matrix enables the module to dynamically adjust the weights of frequency components according to input features, enhance the reconstruction capability of high-frequency details, and improve the fineness and clarity of the image.

[0133] The diffusion process reconstruction module combines U-Net and Transformer architectures, optimizes isotropic design, applies diffusion models and efficient feature fusion and reconstruction technology to achieve efficient and high-quality image super-resolution reconstruction. This module includes an enhanced multi-scale Transformer structure, isotropic design optimization, Residual Shifting (ResShift) method to reduce the number of diffusion steps and speed up generation, as well as feature fusion and reconstruction strategies to ensure high-quality generation and detail capture of high-resolution images.

[0134] The contrastive learning module enhances the model's ability to learn features between images of different resolutions by introducing a contrastive loss function, thereby improving the quality and detail restoration of super-resolution images. The module includes a feature extraction unit, a positive and negative sample construction unit, a contrastive loss calculation unit, and an optimization update unit. The contrastive learning module enhances the model's ability to learn details between low-resolution and high-resolution images through feature extraction, sample construction, and loss calculation, thereby improving detail restoration and overall visual quality of generated images.

[0135] The present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the super-resolution image reconstruction method based on dynamic frequency domain adaptive coding and contrast constraint optimization is implemented.

[0136] The present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the above-mentioned super-resolution image reconstruction method based on dynamic frequency domain adaptive coding and contrast constraint optimization.

[0137] The technical solution of the present invention is further described in detail below in conjunction with specific embodiments:

[0138] The overall framework of the network model of the present invention is as follows Figure 3As shown. The specific embodiments of the present invention will be described in detail below. These embodiments provide detailed implementation methods and operation steps based on the technical solutions of the present invention. However, the scope of application of the present invention is not limited to the embodiments, and any improvement, change or equivalent substitution within the technical framework of the present invention should be included in the protection scope of the present invention.

[0139] This example comprehensively evaluates the proposed model on four real-world super-resolution (SR) tasks. The training datasets include LSDIR, DIV2K, DIV8K, OutdoorSceneTraining, Flicker2K, and the first 10,000 face images selected from the FFHQ dataset. During the training process, we randomly cropped the high-resolution (HR) images in the training set to a size of 256×256, and used the degradation process of RealESRGAN to generate low-resolution (LR) / high-resolution (HR) paired data. The synthesis of these degraded data not only enriches the diversity of training samples, but also enhances the model's learning ability for different types of degraded images. In the test phase, the images in the test set were center-cropped to a size of 512×512 and went through the same degradation process as the training phase to create a synthetic dataset. Through experiments on this test set, we systematically verified the effectiveness and performance of the model in super-resolution reconstruction tasks. The results show that the model has significant advantages in image quality improvement and can effectively cope with a variety of complex low-resolution input scenes.

[0140] The specific steps of this embodiment are as follows:

[0141] 1. Data Preprocessing Module

[0142] Training phase: During the training phase, we provide an original training dataset for the super-resolution image reconstruction model, denoted as D. In order to enhance the generalization ability of the model in complex scenes and ensure that it can cope with low-resolution images of different types and qualities, we carefully designed a series of data augmentation strategies. These strategies include but are not limited to operations such as image rotation, flipping, scaling, and random cropping. Through a variety of augmentation methods, the diversity of training samples is significantly improved, thereby improving the generalization ability and robustness of the model. These data augmentation measures enable the model to better learn useful features when facing complex scenes and different types of low-resolution images, thereby improving the effect of super-resolution reconstruction.

[0143] The specific implementation is as follows:

[0144] Image rotation: We rotated each image at random angles ranging from 0 to 360 degrees. This operation generated images from different perspectives, allowing the model to learn and grasp the characteristics of low-resolution images at different rotation angles. This method effectively enhances the model's ability to cope with a variety of complex scenes, enabling it to accurately perform super-resolution reconstruction when processing low-resolution images in various directions.

[0145] Image flipping: We randomly flipped the images in the training dataset horizontally and vertically. Through the flipping operation, new data samples were generated, allowing the model to learn and recognize the features of low-resolution images in different directions during training. This enhancement strategy not only improves the diversity of training samples, but also effectively improves the robustness of the model in various complex scenarios, helping it better cope with image inputs in different directions.

[0146] Scale scaling: In order to adapt the model to low-resolution images of different resolutions and sizes, we randomly scale the original data. These scaling operations simulate the changes of low-resolution images at different resolutions and scales, allowing the model to better capture the detailed features of these changes, improving its ability to process multi-scale images, thereby enhancing the effect of super-resolution reconstruction.

[0147] Random cropping: During the training process, we randomly cropped the original image into sub-images of different sizes. This operation not only increases the diversity of training samples, but also enables the model to focus on local features in the image and learn more detailed information. This plays an important role in enhancing the performance of the model in super-resolution reconstruction tasks.

[0148] After these data augmentation operations, an extended training dataset D1 is generated. On this basis, all images are further processed to the same size to ensure the consistency of the training data. This standardization process not only improves the training efficiency, but also ensures the stability of the model under different input conditions. Through the strategy of combining data augmentation and standardization, the model can converge faster during training while showing excellent generalization performance.

[0149] Inference stage: The design goal of the data preprocessing module in the inference stage is to efficiently process the input image and maximize the utilization of hardware resources to ensure that the inference stage achieves the best balance between computing speed and output image quality. To this end, we introduced some lightweight data preprocessing steps in the inference stage to reduce the overhead of redundant calculations. These processes not only optimize the quality of the input data, but also effectively accelerate the inference process, thereby improving the response speed and stability of the model in practical applications while ensuring the image super-resolution reconstruction effect.

[0150] The specific implementation is as follows:

[0151] Image reading and normalization: The first step in the inference phase is to read the input image from an external device or storage system. In order to ensure that the input image can be effectively processed by the model, the preprocessing module normalizes the image and adjusts the pixel values ​​to the range expected by the model. This process not only ensures the consistency of data during the inference process, but also effectively avoids the impact of data deviation on model performance. In addition, the module also supports color space conversion and format adjustment of input images in different formats to ensure that they meet the format and size requirements of the model input layer, providing standardized data input for subsequent inference.

[0152] Precision conversion: An important optimization strategy in the inference stage is to use FP16 (16-bit floating point) precision for calculations. Compared with the traditional FP32 (32-bit floating point) precision, FP16 precision can significantly reduce video memory usage and computing resource consumption, thereby increasing the speed of inference. Although the accuracy is reduced, FP16 can still guarantee sufficient calculation accuracy in most cases. This precision conversion is particularly suitable for resource-constrained hardware platforms, such as mobile devices and embedded systems, so that the model can also achieve efficient inference on these platforms.

[0153] Dynamically adjust input images: Input images may vary greatly in size and resolution. To ensure that the model maintains consistent performance when processing different input images, the data preprocessing module has the function of dynamically adjusting the input image size. The module automatically performs operations such as scaling and cropping based on the actual size of the input image to make the image size meet the input requirements of the model. This dynamic adjustment not only ensures the standardization of input data, but also ensures that the output effect of the model is consistent when processing images of different sizes and resolutions.

[0154] Batch processing function: In order to improve the efficiency of inference, the data preprocessing module supports batch image processing. For large-scale inference tasks, this module can process multiple image samples at a time and automatically complete normalization, precision conversion and other operations. The batch processing capability significantly improves the system throughput, allowing it to maintain stable performance when facing high-load inference tasks, thereby meeting the needs for fast response in practical applications.

[0155] Through the above optimization designs, the data preprocessing module in the inference phase effectively reduces the consumption of computing resources and significantly improves the inference speed while ensuring the quality of the model output image. These optimization measures enable the system to quickly and accurately perform super-resolution image reconstruction on resource-constrained hardware platforms, which is particularly suitable for application scenarios with high requirements for response speed and image quality.

[0156] 2. Multi-scale feature extraction module

[0157] The multi-scale feature extraction module in this invention is designed to efficiently and accurately extract multi-level, multi-scale feature representations from input low-resolution images. The module adopts an advanced U-shaped encoder-decoder (U-Net) architecture, combined with deep convolutional neural network (CNN) technology, and achieves excellent feature extraction capabilities through a series of optimized designs.

[0158] The specific implementation is as follows:

[0159] Step 1: Input preprocessing

[0160] Image input: Receive the input low-resolution image, assuming the image size is H×W×C, where H is the height, W is the width, and C is the number of channels. The input image is normalized to normalize the pixel values ​​to the range of [0, 1] or [-1, 1] to accelerate network training and improve stability.

[0161] Initial convolutional layer: Use a 3×3 convolution kernel, set the number of filters to 64, the stride to 1, and the padding to “same”.

[0162] The activation function uses ReLU (rectified linear unit) to introduce nonlinear characteristics. Batch normalization is performed to accelerate the training process and stabilize the network.

[0163] Step 2: Encoder Part

[0164] The encoder part is responsible for gradually extracting and compressing image features, which is described in multiple levels below.

[0165] First coding level:

[0166] Convolution operations: two consecutive 3×3 convolutional layers, each with 64 filters, stride 1, and padding "same". Each convolutional layer is followed by ReLU activation and batch normalization.

[0167] Downsampling operation: Use 2×2 maximum pooling with a stride of 2 to reduce the feature map size from H×W to H / 2×W / 2.

[0168] Residual block: A residual block is introduced, which contains two 3×3 convolutional layers, each with 64 filters and a stride of 1. Cross-layer connections are added to the residual block to add the input feature map to the convolution output to alleviate the gradient vanishing problem.

[0169] Second coding level:

[0170] Convolution operations: two consecutive 3×3 convolutional layers with 128 filters per layer, stride 1, and padding “same”. Each convolutional layer is followed by ReLU activation and batch normalization.

[0171] Downsampling operation: Use 2×2 maximum pooling with a stride of 2 to reduce the size of the feature map to H / 4×W / 4.

[0172] Residual block: A residual block is introduced, which contains two 3×3 convolutional layers, 128 filters in each layer, and a stride of 1. A cross-layer connection is added to add the input feature map to the convolution output.

[0173] Third coding level:

[0174] Convolution operations: two consecutive 3×3 convolutional layers with 256 filters per layer, stride 1, and padding “same”. Each convolutional layer is followed by ReLU activation and batch normalization.

[0175] Downsampling operation: Use 2×2 maximum pooling with a stride of 2 to reduce the size of the feature map to H / 8×W / 8.

[0176] Residual block: A residual block is introduced, which contains two 3×3 convolutional layers, 256 filters in each layer, and a stride of 1. A cross-layer connection is added to add the input feature map to the convolution output.

[0177] Step 3: Bottleneck Layer

[0178] The bottleneck layer between the encoder and decoder further deepens the feature extraction.

[0179] Convolutional operations: Use three consecutive 3×3 convolutional layers with 512 filters per layer, stride 1, and padding “same”. Each convolutional layer is followed by ReLU activation and batch normalization.

[0180] Attention mechanism: Introduce the channel attention mechanism to enhance the expressiveness of important features, suppress irrelevant features, and improve the accuracy of feature representation.

[0181] Step 4: Decoder Part

[0182] The decoder part is responsible for gradually restoring the spatial resolution of the feature map and fusing it with the features in the encoder.

[0183] First decoding level:

[0184] Upsampling operation: Use transposed convolution with a convolution kernel of 2×2 and a stride of 2 to restore the feature map size from H / 8×W / 8 to H / 4×W / 4.

[0185] Feature fusion: Concatenate the upsampled feature map with the feature map of the second level of the encoder to form a new feature map with the sum of the channels of the upsampled feature map and the encoder feature map.

[0186] Convolution operations: two consecutive 3×3 convolutional layers with 256 filters per layer, stride 1, and padding “same”. Each convolutional layer is followed by ReLU activation and batch normalization.

[0187] Residual block: A residual block is introduced, which contains two 3×3 convolutional layers, 256 filters in each layer, and a stride of 1. A cross-layer connection is added to add the input feature map to the convolution output.

[0188] Second decoding level:

[0189] Upsampling operation: Use transposed convolution with a 2×2 kernel and a stride of 2 to restore the feature map size from H / 4×W / 4H / 4 to H / 2×W / 2.

[0190] Feature fusion: The upsampled feature map is concatenated with the feature map of the first level of the encoder to form a new feature map with 128+128=256 channels.

[0191] Convolution operations: two consecutive 3×3 convolutional layers with 128 filters per layer, stride 1, and padding “same”. Each convolutional layer is followed by ReLU activation and batch normalization.

[0192] Residual block: A residual block is introduced, which contains two 3×3 convolutional layers, 128 filters in each layer, and a stride of 1. A cross-layer connection is added to add the input feature map to the convolution output.

[0193] Third decoding level:

[0194] Upsampling operation: Use transposed convolution with a kernel of 2×2 and a stride of 2 to restore the feature map size from H / 2×W / 2H to H×W.

[0195] Feature fusion: The upsampled feature map is concatenated with the feature map of the initial convolutional layer to form a new feature map with 64+64=128 channels.

[0196] Convolution operations: two consecutive 3×3 convolutional layers with 64 filters each, stride 1, and padding “same”. Each convolutional layer is followed by ReLU activation and batch normalization.

[0197] Residual block: A residual block is introduced, which contains two 3×3 convolutional layers, each with 64 filters and a stride of 1. A cross-layer connection is added to add the input feature map to the convolution output.

[0198] Step 5: Features are passed to the adaptive transform coding module

[0199] Feature integration: The feature map of the last layer of the decoder is used as the output of the multi-scale feature extraction module. The output feature map has the same size as the input image (H×W×64) and contains rich multi-level and multi-scale information.

[0200] Interface design:

[0201] Design a standardized interface to pass the feature map to the subsequent adaptive transform coding module. Ensure that the format and size of the feature map are consistent with the input requirements of the adaptive transform coding module.

[0202] 3. Adaptive Transform Coding Module

[0203] Adaptive Transformation Coding Module (AdaTCM) is one of the core innovations in this invention to improve the performance of super-resolution image reconstruction. Figure 4 As shown in the figure. This module combines the Discrete Cosine Transform (DCT) and its inverse transform, and through a dynamic weight adjustment mechanism, efficiently encodes the different frequency components in the image features, realizing the adaptive processing capability of each frequency component in the feature map. AdaTCM not only enhances the reconstruction effect of high-frequency details and complex textures, but also improves the generalization ability of the model under different image contents.

[0204] The specific implementation is as follows:

[0205] Step 1: Input feature extraction and preprocessing

[0206] Input feature acquisition: Extract low-resolution feature maps from the previous layer of network or input image, denoted as F in .

[0207] Feature Standardization: F in Normalization is performed to stabilize the numerical calculations in the subsequent transformation process.

[0208] Feature blocking: As needed, the feature map is divided into several small blocks to facilitate local frequency domain processing, improve processing efficiency and detail capture capabilities.

[0209] Step 2: Frequency domain transformation

[0210] DCT transformation: Use the learnable transformation matrix W to transform the input feature F in Perform discrete cosine transform (DCT) to convert the spatial domain information into frequency domain representation and obtain the frequency domain feature F freq :

[0211] F freq =W·F in

[0212] Frequency domain feature decomposition: F freq Decompose according to different frequency components to obtain the low-frequency part and the high-frequency part respectively, which is convenient for subsequent independent processing.

[0213] Step 3: Dynamic Frequency Mask Generation

[0214] Feature analysis: The input feature F is analyzed through the convolutional neural network module. in Further analysis is performed to extract high-level semantic information and generate contextual features C.

[0215] Mask Generation Network: Design a special Mask Generation Network to use the context feature C to generate a dynamic frequency mask M, which is used to control the enhancement or suppression of different frequency components:

[0216] M=MaskGen(C)

[0217] Mask application: Apply the generated frequency mask M to the frequency domain features F freq , to achieve dynamic adjustment of different frequency components:

[0218] F masked =F freq ⊙M

[0219] Where ⊙ represents an element-by-element multiplication operation.

[0220] Step 4: Frequency encoding and decoding

[0221] Frequency encoding: frequency domain feature F after mask processing masked Perform encoding processing and use learnable weight parameters to adjust and optimize each frequency component:

[0222] F encoded =E·F masked

[0223] Among them, E is the encoding matrix, which is obtained through training.

[0224] Frequency decoding: The encoded frequency domain feature F encoded Decode to recover the frequency domain information:

[0225] F decoded =D·F encoded

[0226] Among them, D is the decoding matrix, corresponding to the encoding matrix E.

[0227] Step 5: Inverse transformation and feature fusion

[0228] Inverse DCT transform: Use the inverse discrete cosine transform (IDCT) to transform the decoded frequency domain features F decoded Convert back to the spatial domain and get the restored spatial domain feature F out :

[0229] F out =W -1 ·F decoded

[0230] Feature fusion: The restored feature F out With the original input feature F in Fusion is performed to enhance the overall quality and detail of the image:

[0231] F fused =Fuse(F in ,F out )

[0232] Among them, Fuse represents a fusion operation.

[0233] Step 6: Output Generation and Feedback Mechanism

[0234] Output generation: The fused feature F fused Input to subsequent network layers to generate high-resolution image output.

[0235] Feedback mechanism: The parameters of the AdaTCM module are optimized through the loss function, so that the module can continuously adjust the frequency encoding strategy during the training process to adapt to the needs of different image contents.

[0236] 4. Diffusion Process Reconstruction Module

[0237] The diffusion process reconstruction module combines the U-Net and Transformer architectures to give full play to the advantages of both in image processing and optimize the isotropic design to achieve a more balanced reconstruction capability. This module utilizes the progressive generation characteristics of the diffusion model and combines it with efficient feature fusion and reconstruction technology to significantly improve computational efficiency while ensuring image quality. Compared with traditional methods, this solution not only performs well in super-resolution reconstruction tasks, but also greatly reduces the consumption of computing resources. At the same time, it shows stronger expressiveness in capturing image details, providing a solid guarantee for the overall improvement of model performance.

[0238] The specific implementation is as follows:

[0239] Combination of U-Net and Transformer

[0240] The U-Net architecture has become a popular choice in image generation and restoration tasks due to its excellent multi-scale feature extraction capabilities and symmetrical encoder-decoder structure. U-Net effectively combines feature information at different resolutions through skip connections, thereby achieving efficient feature fusion while maintaining high-resolution details. The present invention innovatively combines the U-Net architecture with the Transformer model to form an enhanced multi-scale Transformer structure to improve the performance of image super-resolution tasks. The Transformer module of the present invention is as follows: Figure 3 shown.

[0241] The specific implementation is as follows:

[0242] Integration of encoder-decoder framework:

[0243] Encoder part: The encoder consists of multiple downsampling Transformer modules, each layer is responsible for extracting image features at different scales. At each layer, the input feature map is processed through a series of Transformer blocks, gradually reducing the spatial resolution of the feature map while increasing its channel dimension. This process is similar to the downsampling path in U-Net, which captures global information by gradually extracting higher-level abstract features.

[0244] Decoder: The decoder consists of multiple upsampling Transformer modules, each of which is responsible for restoring low-resolution feature maps to higher resolutions. At each layer, the decoder restores the spatial resolution of the feature map through upsampling operations and fuses the feature maps of the corresponding encoder layer through skip connections. This skip connection ensures that high-resolution detail information can be effectively passed to the decoder, thereby improving the detail quality of the reconstructed image.

[0245] Multi-scale feature fusion: In the combined architecture of U-Net and Transformer, multi-scale feature fusion is the key to achieving efficient image reconstruction. Specific methods include:

[0246] Skip Connections: The output feature map of each encoder layer is directly passed to the input of the corresponding decoder layer and fused by splicing or weighting. This method not only retains low-level detail information, but also combines high-level semantic information, improving the reconstruction ability of the model.

[0247] Multi-scale attention mechanism: In each decoder layer, a multi-scale attention mechanism is introduced to dynamically adjust the weight distribution between features of different scales through self-attention and cross-attention mechanisms. In this way, the model can more flexibly integrate feature information at different scales and enhance the ability to capture key details.

[0248] Optimization of Transformer module: In order to better adapt to the multi-scale characteristics of the U-Net architecture, the present invention optimizes the design of the Transformer module:

[0249] Hierarchical Transformer Design: The Transformer modules in each encoder and decoder layer adopt a hierarchical design, which is divided into local and global self-attention layers. The local self-attention layer is responsible for capturing local detail information, while the global self-attention layer is responsible for modeling global contextual relationships. This hierarchical design improves the efficiency and effectiveness of Transformer in multi-scale feature extraction.

[0250] Lightweight Transformer structure: In order to reduce computational complexity, this paper adopts a lightweight Transformer structure, such as reducing the number of heads and dimensions, while maintaining its powerful feature expression capabilities. This design greatly improves the computational efficiency of the model while ensuring performance.

[0251] Optimization of Isotropic Designs

[0252] Isotropic design refers to the use of consistent structure and parameter configuration in each part of the model, so that the model has a unified feature processing capability at different scales and levels. This paper further improves the performance of the model in multi-scale feature extraction and image reconstruction by optimizing the isotropic design of the Transformer module. The specific implementation is as follows:

[0253] (1) Consistency of Transformer modules at each stage:

[0254] In the multi-scale architecture combining U-Net and Transformer, the Transformer modules at each stage (i.e., encoder and decoder layers at different resolutions) maintain consistent structure and parameter configuration. This consistent design brings the following advantages:

[0255] Parameter sharing and reuse: By using the same Transformer module in different stages, parameters can be effectively shared and reused, reducing the total number of model parameters while improving the consistency and stability of the model.

[0256] Cross-scale feature transfer: The consistent Transformer module design enables consistent feature processing at different scales, which facilitates seamless transfer and fusion of features between different scales and enhances the model's multi-scale feature expression capabilities.

[0257] (2) Unified channel dimension and depth:

[0258] To achieve isotropic design, the present invention unifies the channel dimension and depth parameters in all Transformer modules:

[0259] Unified channel dimension: The channel dimension of the feature map remains consistent in all stages of the Transformer module. This design simplifies the conversion and fusion process between features of different scales and improves computational efficiency.

[0260] Depth consistency: The depth (i.e., number of layers) of the Transformer modules at each stage remains consistent, ensuring the depth consistency of feature processing and enhancing the overall expressiveness of the model.

[0261] (3) Optimal allocation of computing resources:

[0262] Isotropic design not only focuses on structural consistency, but also includes the rational allocation of computing resources. The present invention optimizes the allocation of computing resources in the following ways:

[0263] Prioritize resource allocation for high-resolution layers: Considering the importance of high-resolution layers in capturing high-frequency details of images, the present invention allocates more computing resources in these layers; for example, at the high-resolution stage, the number or width of Transformer modules is increased to improve its processing capability.

[0264] Resource control of low-resolution layers: In low-resolution layers, due to the small spatial size of the feature maps, the required computing resources are relatively small; therefore, the present invention appropriately reduces the number or width of Transformer modules in these layers to save computing resources.

[0265] (4) Standardized channel dimension configuration:

[0266] In order to further optimize the isotropic design, the present invention standardizes the channel dimensions at each stage:

[0267] High-resolution stage: In the high-resolution stage, the standardized channel dimension is set larger to enhance the model's ability to capture high-frequency details. This configuration ensures that the model can make full use of rich feature information when processing high-resolution images.

[0268] Low-resolution stage: In the low-resolution stage, the standardized channel dimension is relatively small to reduce the computational complexity while maintaining sufficient feature expression capability.

[0269] Intermediate stage: In the intermediate resolution stage, the channel dimension is balanced between the high-resolution and low-resolution stages to ensure balanced feature processing capabilities between each scale.

[0270] (5) Advantages of isotropic design:

[0271] By optimizing the isotropic design, the present invention has achieved significant advantages in the following aspects: Improving model capacity: By allocating more computing resources in the key high-resolution layers, the capacity of the Transformer architecture is significantly improved, enabling it to better handle complex image details. Reducing the number of parameters: The unified Transformer module design and channel dimension configuration reduce the redundant parameters of the model, making the overall model more compact and efficient. Simplifying the design process: Avoiding complex scheduling strategies and diverse module configurations simplifies the model design process and improves design efficiency and maintainability. Enhanced scalability: Standardized and consistent design makes it easier to expand the model to higher or lower resolutions, with good scalability, to meet the needs of different application scenarios.

[0272] Diffusion module:

[0273] The diffusion module is a key component of the present invention. Based on the basic principle of the diffusion model, a high-resolution image is generated by a step-by-step denoising method. The specific implementation is as follows:

[0274] Forward diffusion process:

[0275] Goal: Model the conditional distribution q(x0|y), where y is a low-resolution image and x0 is the corresponding high-resolution image.

[0276] Steps: Forward process definition: Define a T-step forward process to gradually add random noise to the high-resolution image x0 to generate a noisy image x t .

[0277]

[0278] in, x t represents the noisy image at time step t.

[0279] Reparameterization technique: Through the reparameterization technique, the forward process is concisely represented as the addition of noise at each step.

[0280] Modeling of the reverse process:

[0281] Goal: Starting from pure Gaussian noise, gradually generate the image x at the previous time step using the approximate posterior distribution t-1 .

[0282] Steps: Reverse process definition:

[0283] p θ (x t-1 |x t ,y)=N(μ θ (x t ,y,t),∑(x t ,t))

[0284] Among them, ∑(x t ,t) is dependent on The covariance matrix of is usually set to a fixed constant; μ θ (x t ,y,t) by the denoiser ε p (x t ,y,t) parameterization, specifically:

[0285]

[0286] Iterative generation: from x T (pure Gaussian noise), and gradually generate x by iteratively applying the reverse process T-1 ,x T-2 ,…,x0. Application of Residual Shifting (ResShift) method

[0287] Objective: To reduce the length of the Markov chain by constructing a Markov chain between high-resolution (HR) and low-resolution (LR) images, thereby shortening the number of steps in the diffusion process and increasing the generation speed.

[0288] step:

[0289] Residual definition:

[0290] e0=y0-x0

[0291] Among them, y0 is the low-resolution image, x0 is the corresponding high-resolution image, and e0 represents the residual between the two.

[0292] Shift sequence introduction: Introduce a shift sequence {η t}, its value gradually increases from η1→0 to η T →1.

[0293] Forward process redefinition:

[0294] q(x t |x0,y0)=N(x t ; x0+ηt e0,k 2 η t I)

[0295] Here, k is a hyperparameter that controls the noise intensity.

[0296] Reverse process reformulation:

[0297]

[0298] Among them, f θ (x t ,y0,t) is a denoiser and directly predicts x0.

[0299] Advantages: Shorten the number of steps: By constructing a residual Markov chain, the number of diffusion steps required is effectively reduced. Improve speed: The reduced number of steps directly increases the speed of image generation while maintaining high-quality reconstruction results.

[0300] Feature fusion and reconstruction

[0301] Feature fusion and reconstruction are key steps to ensure the generation of high-quality, high-resolution images. The present invention achieves efficient feature fusion and reconstruction through a series of technical means. The specific implementation is as follows:

[0302] AdaTCM is responsible for extracting and encoding multi-scale features of low-resolution images, providing rich information for subsequent feature fusion and reconstruction.

[0303] Function: Multi-scale feature extraction: Through multi-layer Transformer modules, image features of different scales are gradually extracted. Context information modeling: Using the Transformer's self-attention mechanism, the global context information in the image is captured to enhance the feature representation capability.

[0304] Feature fusion strategy: At each time step of the diffusion process, the model needs to effectively fuse features from different stages to ensure accurate reconstruction of high-frequency details.

[0305] step:

[0306] Feature alignment: Ensure that features of different scales are aligned in spatial and channel dimensions to facilitate subsequent fusion.

[0307] Weighted fusion: A weighted mechanism is used to fuse high-resolution features with low-resolution features to highlight high-frequency information.

[0308]

[0309] in, is the learnable weight parameter, F 高 and F 低are high-resolution and low-resolution features, respectively.

[0310] Multi-layer fusion: Feature fusion is performed at multiple levels to ensure that detailed information at different levels can be integrated.

[0311] High frequency feature enhancement

[0312] In order to improve the detail expression ability of the image, the present invention introduces a high-frequency feature enhancement mechanism, which specifically enhances the high-frequency information in the image. Method:

[0313] Frequency domain conversion: Convert spatial domain features to frequency domain and emphasize high-frequency components through frequency domain filters.

[0314] Residual learning: Through residual connections, high-frequency information is directly introduced into the reconstruction process to avoid information loss.

[0315] F 增强 =F 融合 +ResBlock(F 融合 )

[0316] Among them, ResBlock is a residual block, which is responsible for extracting and enhancing high-frequency features.

[0317] Optimization of the reconstruction process

[0318] The efficient reconstruction process ensures that the generated high-resolution images are not only rich in details but also maintain high computational efficiency.

[0319] Parallel computing: The isotropically designed Transformer module is used to parallelize feature processing and accelerate the reconstruction process.

[0320] Lightweight module design: By optimizing the structure of the Transformer module, redundant calculations are reduced and overall efficiency is improved.

[0321] Combination of Residual Shifting: The Residual Shifting method is combined with the feature fusion mechanism to further shorten the number of steps in the reconstruction process and improve the generation speed.

[0322] Specific implementation:

[0323] Module integration: Feature fusion and high-frequency enhancement modules are tightly combined to form a unified reconstruction pipeline.

[0324] Dynamic adjustment: According to the characteristics of the input image, the weights of feature fusion and high-frequency enhancement are dynamically adjusted to ensure that they can adapt to the reconstruction requirements of different images.

[0325] Optimization algorithm: Use advanced optimization algorithms to accelerate the convergence of the model and improve the reconstruction effect.

[0326] 5. Comparative Learning Module

[0327] The present invention relates to a contrastive learning module (ConLM) for super-resolution image reconstruction, which enhances the feature learning ability of the model between images of different resolutions by introducing a contrast loss function, thereby improving the quality and detail restoration ability of the generated image.

[0328] The overall architecture of the module:

[0329] ConLM is integrated into the super-resolution image reconstruction model and mainly includes the following functional units:

[0330] Feature extraction unit: used to extract feature representations of input images, including low-resolution images, generated high-resolution images, and target high-resolution images.

[0331] Positive and negative sample construction unit: used to construct positive sample pairs and negative sample pairs required for contrastive learning.

[0332] Contrastive loss calculation unit: used to calculate the contrastive loss function and guide the optimization direction of the model.

[0333] Optimization update unit: used to optimize and update model parameters by combining reconstruction loss and contrast loss.

[0334] The specific implementation is as follows:

[0335] Feature extraction unit: Input: low resolution image X LR and the target high-resolution image X HR Generate high-resolution images: Use super-resolution generator G to generate X LR Processing is performed to obtain the generated high-resolution image X SR =G(X LR ). Feature extraction: Use the feature extraction network φ(·) to extract X LR , X SR and X HR Perform feature extraction and obtain the corresponding feature representation:

[0336] f LR =φ(X LR ),f SR =φ(X SR ),f HR =φ(X HR )

[0337] Positive and negative sample construction unit: Positive sample pair: composed of f SR and f HR It is composed of, indicating that the model expects that the generated high-resolution image should be highly similar to the target high-resolution image in the feature space. Negative sample pair: composed of fSR and f LR Composition, indicating that the generated high-resolution image should be distinguishable from the original low-resolution image in the feature space.

[0338] Contrastive loss calculation unit: Similarity measurement: Cosine similarity is used to calculate the similarity between feature representations:

[0339]

[0340] Contrastive loss function: Use the contrastive loss function based on InfoNCE:

[0341]

[0342] Among them, τ is the temperature parameter, which controls the smoothness of the similarity distribution.

[0343] Optimization update unit: Reconstruction loss calculation: use L1 loss or perceptual loss l rec Measure X SR and X HR The difference between:

[0344] l rec =||X SR -X HR ||

[0345] Total loss function: Combine reconstruction loss and contrast loss to form a total loss function:

[0346] l=l rec +λl con

[0347] Among them, λ is the trade-off coefficient, which adjusts the impact of the two losses on model training.

[0348] Parameter update: Use the optimization algorithm to update the model parameters to minimize the total loss l.

[0349] Detailed component design:

[0350] Design of feature extraction network φ(·)

[0351] Shared generator features: The feature extraction network can share parameters with some convolutional layers of the generator to ensure consistency in the feature space.

[0352] Independent feature extractor: An independent deep convolutional network can also be designed as a feature extractor to improve the richness of feature expression.

[0353] Multi-scale feature fusion: By fusing features of different scales, multi-level information of the image is captured and the ability to express details is enhanced.

[0354] Introduction of attention mechanism: Attention mechanism is added in the feature extraction process to highlight the feature representation of key areas.

[0355] Negative sample expansion strategy

[0356] In-batch negative samples: Use other low-resolution images or generated images in the same training batch as negative samples to enrich the negative sample set.

[0357] Dynamic negative sample library: Establish a dynamically updated negative sample library to store diverse negative samples collected in historical training to improve the effect of contrastive learning.

[0358] Parameter adjustment and optimization

[0359] Setting of temperature parameter τ: Through experimental verification, the most suitable temperature parameter is selected to balance the similarity between positive and negative sample pairs.

[0360] Selection of the trade-off coefficient λ: Adjust the value of λ to ensure that the impact of reconstruction loss and contrast loss on model training is optimally balanced.

[0361] Learning rate strategy: Use dynamic learning rate adjustment strategies, such as cosine annealing or exponential decay, to promote faster and better model convergence.

[0362] The effects that can be achieved by the above embodiments are as follows:

[0363] Improve generalization and robustness: Through a variety of data augmentation strategies, enhance the model's ability to process low-resolution images of different types and qualities in complex scenes, and improve the effect of super-resolution reconstruction.

[0364] Enhanced feature extraction and detail capture:

[0365] The multi-scale feature extraction module and residual connection design ensure efficient extraction of useful features in a variety of application scenarios, capture the tiny details and complex structures of the image, and improve the delicacy and clarity of the reconstructed image.

[0366] Optimizing frequency domain processing:

[0367] The adaptive transform coding module enhances the reconstruction capability of high-frequency details through dynamic frequency masking and frequency domain coding, while retaining low-frequency information, avoiding image distortion and significantly improving the visual quality of the image.

[0368] Faster reconstruction and more efficient calculations:

[0369] The diffusion process reconstruction module combines the U-Net and Transformer architectures, as well as the Residual Shifting method, to reduce the number of diffusion steps, speed up the generation speed, optimize the allocation of computing resources, and improve the real-time and efficiency of the model in high-resolution image processing.

[0370] Enhanced feature learning and detail restoration:

[0371] The contrastive learning module improves the model's feature learning ability between images of different resolutions through feature extraction, positive and negative sample construction, and contrast loss calculation, enhances the detail restoration effect of super-resolution images, and improves the quality and fineness of reconstructed images.

[0372] Modular and flexible design:

[0373] The modular design and flexible structural configuration of each module facilitate integration and expansion, adapt to diverse image processing needs, and enhance the overall performance and adaptability of the system.

[0374] Thus, the present invention has significantly improved the effect of super-resolution image reconstruction through the collaborative work of various modules, optimized the computational efficiency and model performance, enhanced the ability to restore image details, and has broad application prospects and significant technical advantages.

[0375] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any technician familiar with the profession, without departing from the scope of the technical solution of the present invention, according to the technical essence of the present invention, any simple modification, equivalent replacement and improvement made to the above embodiment still falls within the protection scope of the technical solution of the present invention.

Claims

1. A super-resolution image reconstruction method based on dynamic frequency domain adaptive coding and contrast constraint optimization, characterized in that: The following steps are involved: Create an original training data set, use data enhancement methods to improve sample diversity for image data in the original training data set, and process image data using lightweight data preprocessing methods; Considering the characteristics and application scenarios of different types of image data, a modular design method that can be flexibly adjusted and expanded is used to extract features from image data; Combining discrete cosine transform and its inverse transform, adaptive transform coding is performed on different frequency components in image features through dynamic weight adjustment mechanism; Combining U-Net and Transformer architecture, optimizing isotropic design, applying diffusion model and feature fusion and reconstruction technology, and realizing diffusion process reconstruction; By introducing a contrast loss function to perform contrast learning between images of different resolutions, the feature learning ability is enhanced.

2. The super-resolution image reconstruction method based on dynamic frequency domain adaptive coding and contrast constraint optimization according to claim 1, characterized in that: In the feature extraction step, by adjusting the depth, width and configuration of the residual block of the convolutional layer, it is ensured that useful features can be efficiently extracted in a variety of application scenarios; according to the specific application scenario, different activation functions or normalization strategies are customized to adapt to specific task requirements.

3. The super-resolution image reconstruction method based on dynamic frequency domain adaptive coding and contrast constraint optimization according to claim 1, characterized in that: In the feature extraction function, different levels adjust the size of the convolution kernel according to the feature extraction requirements; ReLU is used as the activation function; batch normalization is applied after each convolution layer; in the residual connection design, cross-layer connection is used in each residual block to add the input feature map to the convolution output; the upsampling method uses transposed convolution; through the U-shaped encoder-decoder structure, rich feature information is extracted at multiple scales; the ability to understand complex image structures is enhanced through multi-level feature representation; the depth and breadth of the network are improved by introducing residual blocks, and the detail capture ability of feature extraction is enhanced; the consumption of computing resources is controlled through downsampling and upsampling operations; and computing efficiency is optimized through residual connections and batch normalization.

4. The super-resolution image reconstruction method based on dynamic frequency domain adaptive coding and contrast constraint optimization according to claim 1, characterized in that: In the adaptive transform coding step, the transformation matrix is ​​used to transform the input image features into the frequency domain, converting the spatial domain information into a frequency domain representation; combining the features of the input image data, a frequency mask is dynamically generated to selectively enhance or suppress specific frequency components; the generated frequency mask is applied to perform targeted encoding and decoding processing on each frequency component in the frequency domain; the features processed in the frequency domain are transformed back to the spatial domain through an inverse DCT transform and fused with the original features.

5. The super-resolution image reconstruction method based on dynamic frequency domain adaptive coding and contrast constraint optimization according to claim 1, characterized in that: In the diffusion process reconstruction step, the U-Net architecture is combined with the Transformer model to form an enhanced multi-scale Transformer structure, including: Encoder: It consists of multiple layers of downsampling Transformer modules, which gradually extract image features of different scales; Decoder: It consists of multiple layers of upsampling Transformer modules, which fuse the features of the corresponding encoder layers through skip connections to restore high-resolution images; Multi-scale feature fusion: Use skip connections and multi-scale attention mechanisms to achieve feature fusion and enhance the ability to capture details; Transformer optimization: adopts layered and lightweight design to improve the efficiency and effect of multi-scale feature extraction; The optimized isotropic design includes: The Transformer module structure and parameters are consistent at each stage, achieving parameter sharing and cross-scale feature transfer; All Transformer modules have unified channel dimensions and number of layers, simplifying feature conversion and improving computational efficiency; The high-resolution layer is allocated more computing resources first, and the low-resolution layer is appropriately reduced in resources to optimize the overall computing allocation; The channel dimensions are standardized according to different resolution stages to balance the capture of high-frequency details and computational complexity.

6. The super-resolution image reconstruction method based on dynamic frequency domain adaptive coding and contrast constraint optimization according to claim 1, characterized in that: The diffusion model generates a high-resolution image by gradually denoising, including: gradually adding noise in the forward process, and gradually restoring the image through the denoiser in the reverse process; constructing a residual Markov chain between high-resolution and low-resolution images, reducing the number of diffusion steps and speeding up the generation speed; The feature fusion and reconstruction technology includes: extracting and encoding multi-scale features of low-resolution images through the AdaTCM encoder to model global context information; adopting feature fusion methods, including using feature alignment strategies to unify the spatial and channel dimensions of features of different scales; using weighted fusion strategies to fuse high- and low-resolution features through learnable weight parameters to highlight high-frequency information; using multi-layer fusion strategies to integrate detail information at different levels at multiple levels; adopting high-frequency feature enhancement methods, including using frequency domain conversion strategies to emphasize high-frequency components through frequency domain filters, and using residual learning strategies to directly introduce high-frequency information using residual connections to avoid information loss; adopting reconstruction process optimization methods, including parallel computing and lightweight module design to improve computing efficiency, combining residual shifting methods to shorten the number of reconstruction steps, and performing module integration and dynamic adjustment to ensure adaptation to different image characteristics.

7. The super-resolution image reconstruction method based on dynamic frequency domain adaptive coding and contrast constraint optimization according to claim 1, characterized in that: The contrastive learning steps include: using a feature extraction method to input a low-resolution image and generate a high-resolution image through a super-resolution generator; using a feature extraction network to extract features from the low-resolution image, the generated high-resolution image and the target high-resolution image to obtain their respective feature representations; using a positive and negative sample construction method to construct positive sample pairs and negative sample pairs; using a contrastive loss calculation method to measure the relationship between positive sample pairs and negative sample pairs by calculating the similarity between features and using a contrastive loss function to guide model optimization.

8. A super-resolution image reconstruction system based on dynamic frequency domain adaptive coding and contrast constraint optimization, characterized in that: include: The data preprocessing module is used to create the original training data set, use the data enhancement method to improve the diversity of samples in the image data of the original training data set, and process the image data through the lightweight data preprocessing method; The multi-scale feature extraction module is used to consider the characteristics and application scenarios of different types of image data, and extract features from image data using a modular design method that can be flexibly adjusted and expanded; Adaptive transform coding module, used to combine discrete cosine transform and its inverse transform, and perform adaptive transform coding on different frequency components in image features through dynamic weight adjustment mechanism; The diffusion process reconstruction module is used to combine the U-Net and Transformer architectures, optimize the isotropic design, and apply the diffusion model and feature fusion and reconstruction technology to achieve diffusion process reconstruction; Contrastive learning module, which is used to enhance feature learning ability by introducing contrast loss function to perform contrast learning between images of different resolutions; The system is used to execute the super-resolution image reconstruction method based on dynamic frequency domain adaptive coding and contrast constraint optimization as described in any one of claims 1 to 7.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the super-resolution image reconstruction method based on dynamic frequency domain adaptive coding and contrast constraint optimization as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the super-resolution image reconstruction method based on dynamic frequency domain adaptive coding and contrast constraint optimization according to any one of claims 1 to 7.

Images