A method, device and storage medium for continuous scale image super-resolution reconstruction

Abstract

The present invention discloses a continuous scale image super-resolution reconstruction method, including the following specific operation steps: S1. constructing a training set, a validation set and a verification set, using a bicubic interpolation method to degrade the high-resolution original image to obtain a corresponding low-resolution image; S2. cropping the high-resolution original image and the low-resolution image into paired image blocks according to a preset image size, and performing data enhancement on the image blocks; S3. establishing a deep network model, inputting the low-resolution image for reconstruction processing to obtain a super-resolution reconstructed image; S4. defining a loss function, supervising the final super-resolution reconstructed image through the loss function, iteratively optimizing the learnable parameters of the network model and calculating the minimum loss, and improving the image super-resolution reconstruction quality of the deep network model. The present invention can gradually improve the reconstructed image to reduce the learning cost when the decoder performs pixel mapping on the image coordinates, and improve the reconstruction effect of the large-scale factor reconstruction task.

Description

Technical Field

[0001] The present invention relates to the technical field of image reconstruction, and in particular to a method, device and storage medium for continuous-scale image super-resolution reconstruction. Background Art

[0002] Single image super-resolution reconstruction technology is a classic inverse problem in computer vision tasks. In image super-resolution reconstruction, since the restoration of high-resolution images is not unique, a low-resolution image may correspond to multiple high-resolution images, which is challenging in the underlying visual tasks.

[0003] In traditional image super-resolution reconstruction technology, interpolation or shallow feature learning methods are mainly used to upsample and reconstruct images. However, these methods have limitations. With the gradual development of deep learning, deep learning-based methods have end-to-end deep models with strong feature representation. Therefore, they can extract rich image features from images by learning a large amount of data, and adaptively learn the mapping relationship between low-resolution images and high-resolution images.

[0004] In this research direction, the mainstream method is fixed-scale image super-resolution reconstruction. This method focuses on extracting feature embeddings from low-resolution images and using these embeddings to upsample images with predefined scale factors through learnable deconvolution or sub-pixel convolution to achieve reconstruction effects. However, the methods proposed in this research direction require setting a different deep neural network model for each upsampling scale, and are usually limited to a limited range of integer selections (for example, 2×, 3×, 4×). This limitation limits the potential applications and deployment options of image super-resolution models. Therefore, the existing technology uses a method of continuously upsampling low-resolution images through a single model to overcome this limitation.

[0005] Inspired by the implicit neural representation widely used in the field of three-dimensional vision, by applying implicit neural representation to the image super-resolution reconstruction task, and using a multi-layer perceptron to fit the implicit function and map the two-dimensional space coordinates to RGB values, the limitation of fixed scale factor is removed. Compared with the traditional upsampling method at a specific scale, the method based on implicit neural representation has better adaptability and flexibility. This method can generate reconstructed images of any scale as required, which greatly expands the application scenarios of the model.

[0006] In existing research, the upsampling decoder in the model usually uses a single-step reconstruction method, that is, only two layers of multi-layer perceptrons are used to realize the mapping of image coordinates to image pixel values. However, when facing upsampling with a large scale factor, this mapping method is difficult to directly generate a high-resolution image with good reconstruction quality in only one step. At the same time, the reconstructed image often introduces artifacts that affect the visual effect.

[0007] To this end, the present application proposes a continuous-scale image super-resolution reconstruction method to solve the above technical problems. Summary of the invention

[0008] The main purpose of the present invention is to provide a continuous-scale image super-resolution reconstruction method, which reduces the learning cost of the decoder when performing pixel mapping on the image coordinates by gradually improving the reconstructed image, improves the reconstruction effect of large-scale factor reconstruction tasks, and solves the technical problems raised in the background technology.

[0009] The present invention adopts the following technical solutions to solve the above technical problems:

[0010] A continuous scale image super-resolution reconstruction method includes the following specific operation steps:

[0011] S1. Construct training set, validation set and verification set, use bicubic interpolation method to degrade the high-resolution original image and obtain the corresponding low-resolution image;

[0012] S2. Crop the high-resolution original image and the low-resolution image into pairs of image blocks according to a preset image size, and perform data augmentation on the image blocks by random flipping and random rotation;

[0013] S3. Establish a deep network model of a continuous image super-resolution reconstruction method, input a low-resolution image and extract deep image features, and obtain a final super-resolution reconstructed image after reconstruction using the deep network model, wherein a scale-aware encoder and a dynamic reconstruction decoder are set in the deep network model;

[0014] S4. Define the loss function of the deep network model, supervise the final super-resolution reconstructed image through the loss function, continuously iterate and optimize the learnable parameters of the deep network model and calculate the minimum loss to improve the image super-resolution reconstruction quality of the deep network model. The loss function expression is:

[0015]

[0016] Where N represents the total number of pixels in an image. represents the pixel value at the i-th position in the high-resolution image, Represents the pixel value at the i-th position in the super-resolution reconstructed image.

[0017] Preferably, the specific processing steps of using the deep network model reconstruction process in step S3 include:

[0018] S31. Construct a scale-aware encoder to extract deep image features of the input low-resolution image;

[0019] S32. Construct a dynamic reconstruction decoder to obtain image residuals through implicit neural representation and dynamic reconstruction mechanism for deep image features;

[0020] S33. Perform bicubic upsampling on the low-resolution image to obtain a high-resolution image;

[0021] S34. Superimpose the high-resolution image and the image residual to obtain the final super-resolution reconstructed image.

[0022] Preferably, the specific operation steps of step S33 include:

[0023] L1. Extract shallow image features of a given low-resolution image and input the shallow image features into the scale-aware encoder;

[0024] L2. Use a scale-aware encoder to refine image features and establish scale dependence on the trained upsampling scale factor and model learnable parameters, including 6 layers of scale-aware refinement residual blocks, where each layer of scale-aware refinement residual blocks includes 6 layers of Transformer blocks and 1 convolution layer with a kernel size of 3×3;

[0025] L3. Decompose the large-scale factor of image super-resolution into a sequence subset containing multiple scale factors. During the large-scale factor decomposition process, if the large-scale factor is greater than 2, the scale factor is recursively divided by 2 until the scale factor of the current recursive step is less than or equal to 2. According to the different scale factors of the reconstruction task, a dynamic scale factor sequence subset is realized;

[0026] L4. The dynamic reconstruction decoder receives the scale-dependent image features refined by the scale-aware encoder, converts the scale-dependent image features into two-dimensional coordinates through implicit neural representation, selects a single scale factor in sequence according to the sequence subset order, and upsamples the two-dimensional coordinates of the low-resolution image features to the two-dimensional coordinates of the target image of the selected scale factor;

[0027] L5. Use a multi-layer perceptron to map the two-dimensional coordinates of the target scale image to the RGB values ​​of the corresponding pixel coordinate positions to obtain an intermediate high-resolution image feature;

[0028] L6. The dynamic reconstruction decoder receives the intermediate high-resolution image features and repeats steps L4 and L5 to obtain the high-resolution image features of the final target scale. The specific high-resolution image feature reconstruction process is expressed as:

[0029]

[0030] Where F represents the image features, N represents the total number of reconstructions, and G represents the multi-layer perceptron. represents a two-dimensional coordinate mapping function, ↑ represents bilinear interpolation upsampling, and S represents the scale factor in the sequence subset;

[0031] L7. Reconstruct the high-resolution image features of the final target scale into an image residual with a channel value of 3 through a multi-layer perceptron with an output dimension of 3;

[0032] L8. Add the image residual to the high-resolution image obtained by upsampling the low-resolution image through bilinear interpolation, so that the deep network model can learn the high-frequency details of the image and obtain the final reconstructed high-resolution image.

[0033] Preferably, the step of establishing the scale-aware refined residual block in the L2 step includes:

[0034] L21. Define a set of convolution kernel sequences as feature extraction pools with different upsampling scale factors, and use sine-cosine encoding to map the scalar value r of the upsampling scale factor to a higher-dimensional latent space through a linear layer;

[0035] L22. Multiply the mapped scale-dependent vector with the same feature dimension as the initialized convolution kernel by the convolution kernel in the scale-aware refinement residual block to obtain the final scale-dependent convolution kernel;

[0036] L23. Establish the dependency between the refined residual block and the scale factor and form a scale-aware refined residual block. The specific relationship of the scale-aware process is expressed as:

[0037]

[0038] where F out represents the output image features, represents the scale mapping function, represents the weight parameter of the convolutional layer, F in represents the input image features, represents the bias parameter in the convolutional layer, and i is the preset sine-cosine encoding vector.

[0039] On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0040] On the other hand, the present invention further discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0041] It can be seen from the above technical solution that the present invention provides a continuous scale image super-resolution reconstruction method. Compared with the prior art, the present invention has the following advantages:

[0042] 1. The present invention explicitly introduces the scale factor into the deep network model through a scale-aware encoder, which can establish a dependency relationship between the upsampling scale factor and the model, thereby facilitating high-dimensional mapping of continuous-scale images after multi-feature dimensional mapping.

[0043] 2. By setting a dynamic reconstruction decoder, the present invention can adaptively establish the number of steps of implicit neural representation to reconstruct the image according to different upsampling scale factors, thereby realizing the model's dynamic reconstruction capability for different upsampling scale factors and reducing the difficulty of the task of reconstruction by single-step large-scale factors in the past.

[0044] 3. The present invention introduces a dynamic reconstruction mechanism in the decoder, which can gradually improve the reconstructed image to reduce the learning cost of the decoder when performing pixel mapping on the image coordinates, thereby improving the reconstruction effect of large-scale factor reconstruction tasks.

[0045] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present invention, nor is it intended to limit the scope of the present invention. Other features of the present invention will become easy to understand through the following description. Of course, it is not necessary to achieve all of the advantages described above simultaneously for any product implementing the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The drawings constituting a part of the present application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0047] Figure 1 It is a schematic diagram of the overall steps of the method of the present invention;

[0048] Figure 2 A schematic diagram of the process of obtaining high-resolution images using the deep network model of the present invention;

[0049] Figure 3 A schematic diagram of a network framework for obtaining high-resolution images using a deep network model of the present invention;

[0050] Figure 4Schematic diagram of the structural module of the scale-aware refinement residual block of the present invention. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. In the absence of conflict, the embodiments in this application and the features in the embodiments can be combined with each other. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0052] In the embodiment, see Figures 1 to 4 .

[0053] like Figure 1 A continuous scale image super-resolution reconstruction method proposed in an embodiment of the present invention includes the following specific operation steps:

[0054] S1. Construct training set, validation set and verification set, use bicubic interpolation method to degrade the high-resolution original image and obtain the corresponding low-resolution image;

[0055] S2. Crop the high-resolution original image and the low-resolution image into pairs of image blocks according to a preset image size, and perform data augmentation on the image blocks by random flipping and random rotation;

[0056] S3. Establish a deep network model for the continuous image super-resolution reconstruction method, input a low-resolution image and extract deep image features, and use the deep network model to reconstruct the image to obtain the final super-resolution reconstructed image, wherein the deep network model is provided with a scale-aware encoder and a dynamic reconstruction decoder, wherein the specific processing steps of using the deep network model for reconstruction are as follows: Figure 3 As shown, including:

[0057] S31. Construct a scale-aware encoder to extract deep image features of the input low-resolution image;

[0058] S32. Construct a dynamic reconstruction decoder to obtain image residuals through implicit neural representation and dynamic reconstruction mechanism for deep image features;

[0059] S33. Perform bicubic upsampling on the low-resolution image to obtain a high-resolution image. The specific operation steps are as follows: Figure 2 and Figure 3 As shown, including:

[0060] L1. Extract shallow image features of a given low-resolution image and input the shallow image features into the scale-aware encoder to refine the deep features of the image;

[0061] L2. Use the scale-aware encoder to refine the image features and establish the scale dependence of the trained upsampling scale factor and the model learnable parameters, including 6 layers of scale-aware refinement residual blocks, such as Figure 4 As shown in the figure, each layer of scale-aware refinement residual block consists of 6 layers of Transformer blocks and 1 convolutional layer with a kernel size of 3×3;

[0062] The steps of establishing the scale-aware refinement residual block include:

[0063] L21. Define a set of convolution kernel sequences as feature extraction pools with different upsampling scale factors, and use sine and cosine encoding (the preset encoding vector is i) to map the scalar value r of the upsampling scale factor to a higher dimensional latent space through a linear layer;

[0064] L22. Multiply the mapped scale-dependent vector with the same feature dimension as the initialized convolution kernel by the convolution kernel in the scale-aware refinement residual block to obtain the final scale-dependent convolution kernel;

[0065] L23. Establish the dependency between the refined residual block and the scale factor and form a scale-aware refined residual block. The specific relationship of the scale-aware process is expressed as:

[0066]

[0067] where F out represents the output image features, represents the scale mapping function, represents the weight parameter of the convolutional layer, F in represents the input image features, represents the bias parameter in the convolutional layer, i is the preset sine and cosine encoding vector;

[0068] In addition, it can be added that, in a further specific embodiment, a scale-aware convolution layer can be explicitly introduced in the feature extraction step, and by modeling a given scale factor and dynamically adjusting the scale-aware convolution layer weights, the model can be provided with scale prior information, effectively reducing the scale error of the model when sampling on continuous scales during training, and further improving the robustness of the model;

[0069] In summary, by explicitly introducing the scale factor into the deep network model through the scale-aware encoder, the dependency relationship between the upsampling scale factor and the model can be established, which facilitates the high-dimensional mapping of continuous scale images after multi-feature dimension mapping;

[0070] L3. Decompose the large-scale factor of image super-resolution into a sequence subset containing multiple scale factors. During the large-scale factor decomposition process, when the large-scale factor is greater than 2, the scale factor is recursively divided by 2 until the scale factor of the current recursive step is less than or equal to 2. According to the different scale factors of the reconstruction task, a dynamic scale factor sequence subset is realized;

[0071] L4. The dynamic reconstruction decoder receives the scale-dependent image features refined by the scale-aware encoder, converts the scale-dependent image features into two-dimensional coordinates through implicit neural representation, selects a single scale factor in sequence according to the sequence subset order, and upsamples the two-dimensional coordinates of the low-resolution image features to the two-dimensional coordinates of the target image of the selected scale factor;

[0072] At this time, by setting up a dynamic reconstruction decoder, the number of steps of the implicit neural representation to reconstruct the image can be adaptively established according to different upsampling scale factors, realizing the model's dynamic reconstruction capability for different upsampling scale factors, reducing the difficulty of the task of reconstruction by single-step large scale factors in the past;

[0073] L5. Use a multi-layer perceptron to map the two-dimensional coordinates of the target scale image to the RGB values ​​of the corresponding pixel coordinate positions to obtain an intermediate high-resolution image feature;

[0074] At this time, by establishing the connection between the previous image features and the current image pixel mapping, the limitations of single-step upsampling can be overcome, and features of different scales can be fully learned and reconstructed to gradually map image coordinates to RGB values ​​in a high-resolution space, and then embedded into the deep network model in an efficient plug-and-play manner.

[0075] L6. The dynamic reconstruction decoder receives the intermediate high-resolution image features and repeats steps L4 and L5 to obtain the high-resolution image features of the final target scale. The specific high-resolution image feature reconstruction process is expressed as:

[0076]

[0077] Where F represents the image features, N represents the total number of reconstructions, and G represents the multi-layer perceptron. represents a two-dimensional coordinate mapping function, ↑ represents bilinear interpolation upsampling, and S represents the scale factor in the sequence subset;

[0078] In summary, by introducing a dynamic reconstruction mechanism in the decoder, the reconstructed image can be gradually improved to reduce the learning cost of the decoder when performing pixel mapping on the image coordinates, and improve the reconstruction effect of large-scale factor reconstruction tasks;

[0079] L7. Reconstruct the high-resolution image features of the final target scale into an image residual with a channel value of 3 through a multi-layer perceptron with an output dimension of 3;

[0080] L8. Add the image residual to the high-resolution image obtained by upsampling the low-resolution image through bilinear interpolation, so that the deep network model learns the high-frequency details of the image and obtains the final reconstructed high-resolution image;

[0081] S34. Superimposing the high-resolution image and the image residual to obtain a final super-resolution reconstructed image;

[0082] S4. Determine the loss function of the deep network model, supervise the final super-resolution reconstructed image through the loss function, continuously iterate and optimize the learnable parameters of the deep network model and calculate the minimum loss to improve the image super-resolution reconstruction quality of the deep network model, where the loss function expression is:

[0083]

[0084] Where N represents the total number of pixels in an image. represents the pixel value at the i-th position in the high-resolution image, Represents the pixel value at the i-th position in the super-resolution reconstructed image.

[0085] On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the above method.

[0086] On the other hand, the present invention further discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0087] In another embodiment provided in the present application, a computer program product comprising instructions is also provided, which, when executed on a computer, enables the computer to execute any continuous-scale image super-resolution reconstruction method in the above embodiments.

[0088] It is understandable that the system provided by the embodiment of the present invention corresponds to the method provided by the embodiment of the present invention, and the explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the above method.

[0089] The embodiment of the present application also provides an electronic device, including a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus.

[0090] Memory, used to store computer programs;

[0091] The processor is used to implement the above-mentioned continuous-scale image super-resolution reconstruction method when executing the program stored in the memory.

[0092] The communication bus mentioned in the above electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc.

[0093] The communication interface is used for communication between the above electronic device and other devices.

[0094] The memory may include a random access memory (RAM) or a non-volatile memory (NVM), such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0095] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, and discrete hardware components.

[0096] It should also be noted that electronic devices also include terminal devices, which can also be called terminals, user equipment (UE), mobile stations (MS), mobile terminals (MT), etc. Terminal devices can be mobile phones, smart TVs, wearable devices, tablet computers (Pad), computers with wireless transceiver functions, virtual reality (VR) terminal devices, augmented reality (AR) terminal devices, wireless terminals in industrial control, wireless terminals in self-driving, wireless terminals in remote medical surgery, wireless terminals in smart grids, wireless terminals in transportation safety, wireless terminals in smart cities, wireless terminals in smart homes, etc. The embodiments of the present application do not limit the specific technology and specific device form adopted by the terminal devices.

[0097] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive Solid State Disk (SSD)), etc.

[0098] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

[0099] In addition, it should be noted that if the embodiments of the present invention involve directional indications (such as up, down, left, right, front, back, etc.), the directional indications are only used to explain the relative position relationship, movement status, etc. between the components in a certain specific posture. If the specific posture changes, the directional indication will also change accordingly.

[0100] In addition, if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the meaning of "and / or" appearing in the full text includes three parallel schemes. Taking "A and / or B" as an example, it includes scheme A, or scheme B, or schemes that A and B meet at the same time. In addition, in the embodiments of the present invention, "multiple" refers to more than two. In addition, the technical solutions between the various embodiments can be combined with each other, but it must be based on the ability of ordinary technicians in the field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

Claims

1. A continuous scale image super-resolution reconstruction method, characterized in that: The specific steps are as follows: S1. Construct training set, validation set and verification set, use bicubic interpolation method to degrade the high-resolution original image and obtain the corresponding low-resolution image; S2. Crop the high-resolution original image and the low-resolution image into pairs of image blocks according to a preset image size, and perform data augmentation on the image blocks by random flipping and random rotation; S3. Establish a deep network model of a continuous image super-resolution reconstruction method, input a low-resolution image and extract deep image features, and obtain a final super-resolution reconstructed image after reconstruction using the deep network model, wherein a scale-aware encoder and a dynamic reconstruction decoder are set in the deep network model; S4. Define the loss function of the deep network model, supervise the final super-resolution reconstructed image through the loss function, continuously iterate and optimize the learnable parameters of the deep network model and calculate the minimum loss to improve the image super-resolution reconstruction quality of the deep network model. The loss function expression is: Where N represents the total number of pixels in an image. represents the pixel value at the i-th position in the high-resolution image, Represents the pixel value at the i-th position in the super-resolution reconstructed image.

2. The continuous scale image super-resolution reconstruction method according to claim 1, characterized in that: The specific processing steps of using the deep network model reconstruction process in step S3 include: S31. Construct a scale-aware encoder to extract deep image features of the input low-resolution image; S32. Construct a dynamic reconstruction decoder to obtain image residuals through implicit neural representation and dynamic reconstruction mechanism for deep image features; S33. Perform bicubic upsampling on the low-resolution image to obtain a high-resolution image; S34. Superimpose the high-resolution image and the image residual to obtain the final super-resolution reconstructed image.

3. The continuous scale image super-resolution reconstruction method according to claim 2, characterized in that: The specific operation steps of step S33 include: L1. Extract shallow image features of a given low-resolution image and input the shallow image features into the scale-aware encoder; L2. Use a scale-aware encoder to refine image features and establish scale dependence on the trained upsampling scale factor and model learnable parameters, including 6 layers of scale-aware refinement residual blocks, where each layer of scale-aware refinement residual blocks includes 6 layers of Transformer blocks and 1 convolution layer with a kernel size of 3×3; L3. Decompose the large-scale factor of image super-resolution into a sequence subset containing multiple scale factors; L4. The dynamic reconstruction decoder receives the scale-dependent image features refined by the scale-aware encoder, converts the scale-dependent image features into two-dimensional coordinates through implicit neural representation, selects a single scale factor in sequence according to the sequence subset order, and upsamples the two-dimensional coordinates of the low-resolution image features to the two-dimensional coordinates of the target image of the selected scale factor; L5. Use a multi-layer perceptron to map the two-dimensional coordinates of the target scale image to the RGB values ​​of the corresponding pixel coordinate positions to obtain an intermediate high-resolution image feature; L6. The dynamic reconstruction decoder receives the intermediate high-resolution image features and repeats steps L4 and L5 to obtain the high-resolution image features of the final target scale. The specific high-resolution image feature reconstruction process is expressed as: Where F represents the image features, N represents the total number of reconstructions, and G represents the multi-layer perceptron. represents a two-dimensional coordinate mapping function, ↑ represents bilinear interpolation upsampling, and S represents the scale factor in the sequence subset; L7. Reconstruct the high-resolution image features of the final target scale into an image residual with a channel value of 3 through a multi-layer perceptron with an output dimension of 3; L8. Add the image residual to the high-resolution image obtained by upsampling the low-resolution image through bilinear interpolation, so that the deep network model can learn the high-frequency details of the image and obtain the final reconstructed high-resolution image.

4. The continuous scale image super-resolution reconstruction method according to claim 3, characterized in that: The steps of establishing the scale-aware refinement residual block in the L2 step include: L21. Define a set of convolution kernel sequences as feature extraction pools with different upsampling scale factors, and use sine-cosine encoding (the encoding vector is i) to map the scalar value r of the upsampling scale factor to a higher-dimensional latent space through a linear layer; L22. Multiply the mapped scale-dependent vector with the same feature dimension as the initialized convolution kernel by the convolution kernel in the scale-aware refinement residual block to obtain the final scale-dependent convolution kernel; L23. Establish the dependency between the refined residual block and the scale factor and form a scale-aware refined residual block. The specific relationship of the scale-aware process is expressed as: where F out represents the output image features, represents the scale mapping function, represents the weight parameter of the convolutional layer, F in represents the input image features, represents the bias parameter in the convolutional layer, and i is the preset sine-cosine encoding vector.

5. The continuous scale image super-resolution reconstruction method according to claim 3, characterized in that: During the large-scale factor decomposition process in the L3 step, if the large-scale factor is greater than 2, the scale factor is divided by 2 in a recursive manner until the scale factor of the current recursive step is less than or equal to 2.

6. The continuous scale image super-resolution reconstruction method according to claim 5, characterized in that: During the large-scale factor decomposition process in the L3 step, a dynamic scale factor sequence subset is implemented according to different scale factors of the reconstruction task.

7. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the processor is caused to execute the steps of the method according to any one of claims 1 to 6.

8. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 6.

Images