Image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion

Abstract

The invention discloses an image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion, and the method comprises the steps: firstly constructing a training data set of a low-resolution-high-resolution image pair, constructing an adaptive noise prediction network based on a deep neural network, enabling the network to take a low-resolution image, a time step and a gradient map as joint input, and carrying out the deep reinforcement learning adaptive noise inversion. Outputting a noise map required by inversion of the diffusion model; designing a loss function based on deep reinforcement learning, wherein the loss function is composed of pixel-level loss, perception loss and wavelet loss; carrying out joint training on the noise prediction network and the diffusion model by adopting a training data set and a loss function, and optimizing an adaptive noise inversion strategy to obtain a trained noise prediction network and a trained diffusion model; and performing super-resolution reconstruction on the low-resolution image to be processed by adopting the trained noise prediction network and the diffusion model to obtain a reconstructed high-resolution image.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to an image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion. Background Technology

[0002] Image super-resolution technology, as one of the core tasks in the field of computer vision, aims to reconstruct high-resolution images from low-resolution images using algorithmic methods. It is widely used in remote sensing imaging, medical diagnosis, video surveillance, and high-definition display. With the development of large-scale model technology, diffusion-based super-resolution methods have gradually become a research hotspot due to their powerful generation capabilities and high-fidelity performance.

[0003] Existing diffusion model super-resolution methods primarily map low-resolution images to high-resolution images through a noise inversion process. The core idea is to find the optimal noise set and reconstruct the high-resolution image through the inverse process of the diffusion model. However, existing technologies suffer from the following key drawbacks: First, the noise inversion process relies heavily on iterative optimization with fixed steps, resulting in low inference efficiency and difficulty meeting the demands of real-time applications. Second, the input to the noise prediction network only contains low-resolution images and time-step information, lacking the ability to capture detailed image features, leading to problems such as blurred edges and texture distortion in the reconstruction results. Third, the knowledge expansion cost of diffusion models is high, and general models struggle to adapt to complex and unknown degradation models in real-world scenarios. Fourth, loss function designs often employ pixel-level or perceptual losses, failing to adequately constrain the multi-scale frequency features of the image. This leads to the "illusion accumulation" phenomenon, where the diffusion model generates non-existent false details due to its iterative nature. Fifth, while deep learning-based methods outperform in inference speed and the angle of "illusion" generation, their generalization ability is limited due to a lack of prior knowledge. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an image super-resolution reconstruction method based on deep reinforcement learning-based adaptive noise inversion. This method improves the model's detail control by introducing image gradient information; designs a noise estimation model to enhance the model's adaptability; and, based on deep reinforcement learning, utilizes wavelet loss constraints to reduce diffusion model illusions, while employing a marginal sampling strategy to improve inference speed and reduce illusion accumulation, thus balancing reconstruction quality and inference efficiency.

[0005] This invention is achieved through the following technical solution: This invention provides an image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion, comprising the following steps: Step S1: Construct a training dataset of low-resolution-high-resolution image pairs; Step S2: Construct an adaptive noise prediction network based on a deep neural network. This network takes a low-resolution image, a time step, and its gradient map as joint inputs and outputs a noise map required for diffusion model inversion. Step S3: Design a loss function based on deep reinforcement learning. The loss function consists of pixel-level loss, perceptual loss, and wavelet loss. The weighted sum of the three losses serves as the reward signal for deep reinforcement learning, guiding the diffusion model to adaptively learn the optimal noise inversion strategy. Step S4: Use the training dataset from Step S1 and the loss function from Step S3 to jointly train the noise prediction network and the diffusion model, optimize the adaptive noise inversion strategy, and obtain the trained noise prediction network and diffusion model. Step S5: Using the noise prediction network and diffusion model trained in step S4, perform super-resolution reconstruction on the low-resolution image to be processed to obtain the reconstructed high-resolution image.

[0006] In the above technical solution, in step S1, high-resolution images of various scenes are acquired as original images, and the original high-resolution images are subjected to denoising, normalization and downsampling processing to obtain low-resolution-high-resolution image pair data.

[0007] In the above technical solution, step S2 includes the following: Step S2-1: For low-resolution images, use the Sobel operator to calculate their horizontal and vertical gradient maps; Step S2-2: Concatenate the low-resolution image with its horizontal and vertical gradient maps to form a joint input feature map, which is then input into the adaptive noise prediction network. Step S2-3: The feature extraction module of the adaptive noise prediction network extracts features from the input feature map; Step S2-4: The attention enhancement module of the adaptive noise prediction network performs attention enhancement processing on the features extracted by the feature extraction module to obtain the enhanced extracted features; Step S2-5: The feature fusion and output module of the adaptive noise prediction network fuses and adjusts the dimensions of the enhanced extracted features, and finally outputs a noise map with the same size as the low-resolution image.

[0008] In the above technical solution, in step S2, the feature extraction module consists of multiple depthwise separable convolutional layers connected in series. Each depthwise separable convolutional layer includes depthwise convolution and pointwise convolution, which are used to efficiently extract texture, edge and noise-related information from the input feature map.

[0009] In the above technical solution, in step S2, the attention enhancement module consists of multiple cross-attention layers connected in parallel, which respectively process the joint features of low-resolution image-horizontal gradient and low-resolution image-vertical gradient. The attention output is added to the input features through residual connection and then normalized to avoid gradient vanishing.

[0010] In the above technical solution, in step S2, the feature fusion and output module fuses and adjusts the dimensions of the extracted features through multiple ordinary convolutional layers and average pooling, and finally outputs a noise map with the same size as the low-resolution image. This noise map conforms to a Gaussian distribution and is used to initialize the inversion process of the diffusion model.

[0011] In the above technical solution, the loss function in step S3 is expressed as: ;in, , , These are pixel-level loss, perceptual loss, and wavelet loss, respectively. , , These are the corresponding weights; Pixel-level loss Represented as: ;in, For true high-resolution images, To reconstruct a high-resolution image, H, W, and C represent the image's height, width, and number of channels, respectively. i , j , k This is an index variable representing the position of a pixel in the image data structure; Perceived loss A pre-trained VGG network is used to extract perceptual features of the image, and the distance between the real image and the reconstructed image in the feature domain is calculated to improve the visual consistency of the reconstructed image. Wavelet loss Stationary wavelet transforms are performed on the real high-resolution image and the reconstructed image respectively. The decomposition level is 1, resulting in one low-frequency sub-band LL and three high-frequency sub-bands LH, HL, and HH. The loss of each of the four sub-bands is calculated and weighted summed to constrain the consistency of the frequency domain features.

[0012] In the above technical solution, step S4 includes: Step S4-1: Initialize the diffusion model and noise prediction network; Step S4-2: Training iteration, set the number of training iterations, batch size, optimizer, initial learning rate, and use the learning rate decay strategy; Step S4-3: Noise inversion training; Step S4-4: Calculate the reconstruction error based on the loss function, update the parameters of the noise prediction network and the fine-tuning layer of the diffusion model through backpropagation, and guide the model to learn the optimal noise inversion strategy; Step S4-5: Evaluate model performance using the validation set. If the perceived loss on the validation set decreases after a certain number of consecutive iterations... If there is no decrease, stop training and save the optimal model parameters.

[0013] In the above technical solution, step S5 includes: Step S5-1: Preprocessing: Denoise, normalize and calculate gradients on the input low-resolution image to obtain joint features that meet the model input requirements; Step S5-2: Noise Prediction: Input the joint features into the noise prediction network and adaptively output a noise map; Step S5-3: Diffusion Inversion Reconstruction: Input the noise map into the diffusion model and start the inverse denoising process. Adaptively adjust the number of sampling steps according to the image degradation type. For noise-dominated degradation images, use 1 sampling step; for blur-dominated degradation images, use 3 to 5 sampling steps. Step S5-4: Post-processing: The pixel values ​​of the reconstructed image output by the diffusion model are denormalized to obtain the final high-resolution image.

[0014] Another aspect of the present invention provides an electronic device comprising: one or more processors; and a memory for storing one or more programs, wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the method as described above.

[0015] Another aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the method described above.

[0016] Another aspect of the present invention provides a computer program product comprising computer-executable instructions which, when executed, are used to implement the method described above.

[0017] The advantages and beneficial effects of this invention are as follows: The present invention provides an image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion. This method improves the model's ability to control details by introducing image gradient information; it designs a noise estimation model to improve the model's adaptability; based on the idea of ​​deep reinforcement learning, it uses wavelet loss constraints to reduce the illusion of the diffusion model, and at the same time uses a marginal sampling strategy to improve inference speed and reduce the accumulation of illusions, thus balancing reconstruction quality and inference efficiency. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion according to the present invention.

[0020] Figure 2 This is a flowchart illustrating the execution process of step S5 in the image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion of the present invention. Detailed Implementation

[0021] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the invention. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the invention for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0022] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0023] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0024] Embodiments of the present invention provide an image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion, see appendix. Figure 1 This includes the following steps: Step S1: Image data acquisition and preprocessing, constructing a training dataset.

[0025] High-resolution images from various scenes are acquired as raw high-resolution images, covering different types such as natural images, medical images, and remote sensing images to ensure the diversity and representativeness of the dataset. The raw high-resolution images are preprocessed to construct low-resolution-high-resolution image pairs. Specifically, the preprocessing operations include the following:

[0026] Step S1-1: Image denoising Preferably, all images are lightly denoised using a 3×3 Gaussian filter (standard deviation σ=0.5) to remove random salt-and-pepper noise, truncate pixel values ​​to [0,255], and retain true structural information.

[0027] Step S1-2: Normalization Preferably, the image pixel values ​​are linearly mapped from [0,255] to the [0,1] interval to eliminate the interference of brightness differences on training.

[0028] Steps S1-3: Downsampling to generate low-resolution images Preferably, for the normalized image, bicubic interpolation is used to perform downsampling at ×2, ×4, and ×8 respectively to simulate the degradation of the real resolution.

[0029] Step S2: Construct an adaptive noise prediction network based on a deep neural network. This network takes a low-resolution image, a time step, and its gradient map as joint inputs and outputs the optimal noise map required for diffusion model inversion.

[0030] Furthermore, step S2 specifically includes the following: Step S2-1: For low-resolution images, use the Sobel operator to calculate their horizontal and vertical gradient maps, and normalize the gradient values ​​to [0,1] as supplementary inputs for subsequent noise prediction networks.

[0031] Step S2-2: Concatenate the low-resolution image with its horizontal and vertical gradient maps to form a joint input feature map, which is then input into the adaptive noise prediction network.

[0032] Step S2-3: The feature extraction module of the adaptive noise prediction network extracts features from the input feature map.

[0033] The feature extraction module consists of multiple depthwise separable convolutional layers connected in series. Each depthwise separable convolutional layer contains depthwise convolution and pointwise convolution, which are used to efficiently extract texture, edge and noise-related information from the input feature map.

[0034] Step S2-4: The attention enhancement module of the adaptive noise prediction network performs attention enhancement processing on the features extracted by the feature extraction module to obtain the enhanced extracted features.

[0035] The attention enhancement module consists of multiple parallel cross-attention layers that process the joint features of low-resolution image-horizontal gradient and low-resolution image-vertical gradient respectively. The attention output is added to the input features through residual connections and then normalized to avoid gradient vanishing.

[0036] Step S2-5: The feature fusion and output module of the adaptive noise prediction network fuses and adjusts the dimensions of the enhanced extracted features, and finally outputs a noise map with the same size as the low-resolution image.

[0037] The feature fusion and output module fuses and adjusts the dimensions of the extracted features through multiple ordinary convolutional layers and average pooling, and finally outputs a noise map with the same size as the low-resolution image. This noise map conforms to a Gaussian distribution and is used to initialize the inversion process of the diffusion model.

[0038] Step S3: Design a loss function based on deep reinforcement learning. The loss function consists of pixel-level loss, perceptual loss, and wavelet loss. The weighted sum of the three losses serves as the reward signal for deep reinforcement learning, guiding the diffusion model to adaptively learn the optimal noise inversion strategy.

[0039] The principle of the loss function is shown in equation (1).

[0040] (1) in, , , These are pixel-level loss, perceptual loss, and wavelet loss, respectively. , , These are the corresponding weights.

[0041] Specifically, pixel-level loss The principle is shown in equation (2).

[0042] (2) in, For true high-resolution images, To reconstruct a high-resolution image, H, W, and C represent the image's height, width, and number of channels, respectively. i , j , k This is an index variable representing the position of a pixel in the image data structure.

[0043] Specifically, perceived loss A pre-trained VGG network is used to extract perceptual features from the image, and the distance between the real image and the reconstructed image in the feature domain is calculated to improve the visual consistency of the reconstructed image. Specifically, a VGG16 network is used, and the output features of two convolutional layers are extracted. The L2 loss is calculated separately for each layer and then summed with a weight of 0.5.

[0044] Specifically, wavelet loss The real high-resolution image and the reconstructed image are subjected to stationary wavelet transform respectively. The wavelet basis is "db4" and the number of decomposition layers is 1. The decomposition yields 1 low-frequency sub-band LL and 3 high-frequency sub-bands LH, HL and HH. The L1 loss of each of the four sub-bands is calculated and weighted and summed to constrain the consistency of frequency domain features and reduce the illusion of diffusion model. The principle is shown in Equation (3).

[0045] (3) in, , , , For weights.

[0046] Step S4: Use the training dataset from Step S1 and the loss function from Step S3 to jointly train the noise prediction network and the diffusion model, optimize the adaptive noise inversion strategy, and obtain the trained noise prediction network and diffusion model.

[0047] The specific steps are as follows: Step S4-1: Initialize the settings, load the pre-trained Stable Diffusion model as the backbone network of the diffusion model, fix the first 80% of the Stable Diffusion layers, and only fine-tune the top UNet residual module to retain the image prior knowledge of the pre-trained diffusion model and reduce the training difficulty; initialize the parameters of the noise prediction network using the Xavier initialization method.

[0048] Step S4-2: Training iteration, set the number of training iterations, batch size, optimizer, initial learning rate, and use the learning rate decay strategy.

[0049] Step S4-3: Noise inversion training, including: Step S4-3-1: For each training batch of low-resolution images, input the low-resolution image and gradient map into the noise prediction network, and output the noise map prediction. .

[0050] Step S4-3-2: Construct intermediate states of the diffusion model The principle is shown in equation (4).

[0051] (4) in, , , LR represents a low-resolution image, which is a preset hyperparameter.

[0052] Step S4-3-3: Inverse denoising of the diffusion model, Input a fine-tuning layer, perform reverse denoising steps T=1 to 5, and output a reconstructed high-resolution image. .

[0053] Step S4-4: Loss Calculation and Parameter Update: Calculate the reconstruction error based on the loss function L, update the parameters of the noise prediction network and the fine-tuning layer of the diffusion model through backpropagation, and guide the model to learn the optimal noise inversion strategy, that is, adaptively adjust the noise prediction results according to the image gradient and frequency domain features.

[0054] Step S4-5: Model Validation and Optimization: After a certain number of iterations, evaluate the model performance using the validation set. Use an early stopping strategy, i.e., if the perceived loss on the validation set exceeds the threshold after a certain number of consecutive iterations... If there is no decrease, stop training and save the optimal model parameters.

[0055] Step S5: Using the noise prediction network and diffusion model trained in step S4, perform super-resolution reconstruction on the low-resolution image to be processed to obtain a high-resolution image.

[0056] like Figure 2 As shown, the specific steps are as follows: Step S5-1: Preprocessing: Denoise, normalize, and calculate gradients on the input low-resolution image to obtain joint features that meet the model input requirements.

[0057] Step S5-2: Noise prediction: Input the joint features into the noise prediction network and adaptively output a noise map.

[0058] Step S5-3: Diffusion Inversion Reconstruction: Input the noise map into the diffusion model and start the inverse denoising process. Adaptively adjust the number of sampling steps according to the image degradation type. For noise-dominated degradation images, use 1 sampling step; for blur-dominated degradation images, use 3 to 5 sampling steps.

[0059] Step S5-4: Post-processing: The pixel values ​​of the reconstructed image output by the diffusion model are denormalized to obtain the final high-resolution image.

[0060] The present invention also provides a computer-readable storage medium carrying one or more programs that, when executed, implement the method according to embodiments of the present invention.

[0061] According to embodiments of the present invention, the computer-readable storage medium may be a non-volatile computer-readable storage medium. Examples include, but are not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In the present invention, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0062] Embodiments of the present invention also include a computer program product comprising a computer program containing program code for performing the methods provided in the embodiments of the present invention. When the computer program product is run on an electronic device, the program code is used to enable the electronic device to implement the image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion provided in the embodiments of the present invention.

[0063] According to embodiments of the present invention, program code for executing the computer programs provided in the embodiments of the present invention can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on a user's computing device, partially on a user's device, partially on a remote computing device, or entirely on a remote computing device or server.

[0064] The embodiments of the present invention have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of the invention. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of the invention, and all such substitutions and modifications should fall within the scope of the invention.

Claims

1. An image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion, characterized in that, Includes the following steps: Step S1: Construct a training dataset of low-resolution-high-resolution image pairs; Step S2: Construct an adaptive noise prediction network based on a deep neural network. This network takes a low-resolution image, a time step, and its gradient map as joint inputs and outputs a noise map required for diffusion model inversion. Step S3: Design a loss function based on deep reinforcement learning. The loss function consists of pixel-level loss, perceptual loss, and wavelet loss. The weighted sum of the three losses serves as the reward signal for deep reinforcement learning, guiding the diffusion model to adaptively learn the optimal noise inversion strategy. Step S4: Use the training dataset from Step S1 and the loss function from Step S3 to jointly train the noise prediction network and the diffusion model, optimize the adaptive noise inversion strategy, and obtain the trained noise prediction network and diffusion model. Step S5: Using the noise prediction network and diffusion model trained in step S4, perform super-resolution reconstruction on the low-resolution image to be processed to obtain the reconstructed high-resolution image.

2. The image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion according to claim 1, characterized in that: In step S1, high-resolution images of various scenes are acquired as the original images. The original high-resolution images are then denoised, normalized, and downsampled to obtain low-resolution-high-resolution image pairs.

3. The image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion according to claim 1, characterized in that: Step S2 includes the following: Step S2-1: For low-resolution images, use the Sobel operator to calculate their horizontal and vertical gradient maps; Step S2-2: Concatenate the low-resolution image with its horizontal and vertical gradient maps to form a joint input feature map, which is then input into the adaptive noise prediction network; Step S2-3: The feature extraction module of the adaptive noise prediction network extracts features from the input feature map; Step S2-4: The attention enhancement module of the adaptive noise prediction network performs attention enhancement processing on the features extracted by the feature extraction module to obtain the enhanced extracted features; Step S2-5: The feature fusion and output module of the adaptive noise prediction network fuses and adjusts the dimensions of the enhanced extracted features, and finally outputs a noise map with the same size as the low-resolution image.

4. The image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion according to claim 3, characterized in that: In step S2, the feature extraction module consists of multiple depthwise separable convolutional layers connected in series. Each depthwise separable convolutional layer contains depthwise convolution and pointwise convolution, which are used to efficiently extract texture, edge and noise-related information from the input feature map.

5. The image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion according to claim 3, characterized in that: In step S2, the attention enhancement module consists of multiple parallel cross-attention layers that process the joint features of low-resolution image-horizontal gradient and low-resolution image-vertical gradient respectively. The attention output is added to the input features through residual connections and then normalized to avoid gradient vanishing.

6. The image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion according to claim 3, characterized in that: In step S2, the feature fusion and output module fuses and adjusts the dimensions of the extracted features through multiple ordinary convolutional layers and average pooling, and finally outputs a noise map with the same size as the low-resolution image. This noise map conforms to a Gaussian distribution and is used to initialize the inversion process of the diffusion model.

7. The image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion according to claim 1, characterized in that: In step S3, the loss function is expressed as: ; in, , , These are pixel-level loss, perceptual loss, and wavelet loss, respectively. These are the corresponding weights; Pixel-level loss Represented as: ； in, For true high-resolution images, To reconstruct a high-resolution image, H, W, and C represent the image's height, width, and number of channels, respectively. i , j , k This is an index variable representing the position of a pixel in the image data structure; Perceived loss A pre-trained VGG network is used to extract perceptual features of the image, and the distance between the real image and the reconstructed image in the feature domain is calculated to improve the visual consistency of the reconstructed image. Wavelet loss Stationary wavelet transforms are performed on the real high-resolution image and the reconstructed image respectively. The decomposition level is 1, resulting in one low-frequency sub-band LL and three high-frequency sub-bands LH, HL, and HH. The loss of each of the four sub-bands is calculated and weighted summed to constrain the consistency of the frequency domain features.

8. The image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion according to claim 1, characterized in that: Step S4 includes: Step S4-1: Initialize the diffusion model and noise prediction network; Step S4-2: Training iteration, set the number of training iterations, batch size, optimizer, initial learning rate, and use the learning rate decay strategy; Step S4-3: Noise inversion training; Step S4-4: Calculate the reconstruction error based on the loss function, update the parameters of the noise prediction network and the fine-tuning layer of the diffusion model through backpropagation, and guide the model to learn the optimal noise inversion strategy; Step S4-5: Evaluate model performance using the validation set. If the perceived loss on the validation set decreases after a certain number of consecutive iterations... If there is no decrease, stop training and save the optimal model parameters.

9. The image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion according to claim 1, characterized in that: Step S5 includes: Step S5-1: Preprocessing: Denoise, normalize and calculate gradients on the input low-resolution image to obtain joint features that meet the model input requirements; Step S5-2: Noise Prediction: Input the joint features into the noise prediction network and adaptively output a noise map; Step S5-3: Diffusion Inversion Reconstruction: Input the noise map into the diffusion model and start the inverse denoising process; Step S5-4: Post-processing: The pixel values ​​of the reconstructed image output by the diffusion model are denormalized to obtain the final high-resolution image.

10. A computer-readable storage medium, characterized in that: The device stores computer-executable instructions, which, when executed, are used to implement the image super-resolution reconstruction method based on deep reinforcement learning adaptive noise inversion as described in claim 1.

Images