Super-resolution reconstruction method and system based on wavelet transform and differential sliding window feature extraction

Abstract

The present invention discloses a super-resolution reconstruction method based on wavelet transform and differential sliding window feature extraction. First, an input image is preprocessed to obtain standardized image data; then the image is input into a differential mode image generator to extract an inverse original image, and the forward and reverse original images constitute a set of differential mode image data; then a two-dimensional discrete wavelet transform is performed on the image; then, four sets of differential mode wavelet subband data are respectively sent to four corresponding trained differential sliding window feature extraction models constructed based on a SwinIR model, the differential sliding window feature extraction model comprising a layer feature extraction layer, a deep feature extraction layer and a high-quality image wavelet subband reconstruction layer, and the deep feature extraction layer comprises six residual blocks and a convolution layer; finally, the four sets of wavelet subband data of the output forward and reverse high-resolution images are combined through a differential operation to remove common mode noise, and a discrete wavelet inverse transform is performed to finally generate a high-quality high-resolution image.

Description

Technical Field

[0001] The present invention relates to the technical field of image quality processing, and in particular to a super-resolution reconstruction method and system based on wavelet transform and differential sliding window feature extraction. Background Art

[0002] The research on super-resolution technology stems from the problem of insufficient image resolution. In particular, in fields such as medical imaging, satellite remote sensing, and surveillance, high-resolution images are important for detail analysis and decision-making. Traditional super-resolution methods, such as interpolation algorithms, are computationally simple but can easily lead to image blurring and loss of detail [1]. Although signal processing methods have been improved, their effectiveness in super-resolution reconstruction of single images is limited and cannot meet the needs of practical applications.

[0003] With the rapid development of deep learning technology, the field of super-resolution has ushered in significant technological breakthroughs. In 2014, the SRCNN model proposed a super-resolution reconstruction method based on convolutional neural networks (CNN), marking the first application of deep learning in this field [2]. Since then, models such as VDSR [3], ESPCN [4], and SRGAN [5] have been introduced. Deep learning has gradually become the mainstream method in the field of super-resolution due to its powerful adaptive feature extraction capabilities and complex nonlinear mapping processing capabilities. Compared with traditional methods, deep learning can better preserve detail information and reduce blurring during image reconstruction, thereby significantly improving image quality. However, deep learning super-resolution methods still face many challenges.

[0004] First, while deep learning methods have achieved promising results in image reconstruction, the resulting high-resolution images are often accompanied by artifacts and unnatural textures in high-frequency details. In particular, the models often introduce spurious structures around image edges and detailed areas, compromising image quality. Second, the generalization ability of deep learning models depends heavily on the diversity and scale of the training data. When data is limited, the generalization ability of the models is weak. Furthermore, because deep learning methods typically involve a large number of computational operations, the models' computational complexity is high, limiting their application in resource-constrained environments.

[0005] To address these problems, scholars have proposed a variety of improved methods. Generative adversarial networks (GANs) combined with perceptual loss functions can effectively generate more realistic detailed images, but they are still prone to introducing artifacts in high-frequency areas [6]. Regularization techniques, such as total variation (TV) regularization, can effectively suppress noise, but may cause excessive smoothing of the image, thereby losing important detail information [7]. Multi-scale and multi-view learning methods help improve image reconstruction effects by extracting features at different scales and angles, but these methods have high computational resource requirements, which limits their application in real-time processing and low-resource environments. The adaptive attention mechanism can enhance the detail representation and global consistency of the image by intelligently selecting important areas, but it increases the complexity and computational burden of the network [8].

[0006] While existing technologies have improved the quality of super-resolution image reconstruction to some extent, achieving high-quality and low-cost super-resolution image reconstruction remains difficult in scenarios with small sample datasets or limited computing resources. Therefore, reducing computational complexity while improving the model's generalization capabilities and effectively reducing artifacts and noise remain key challenges in super-resolution research.

[0007] [1] Gonzalez, RC,&Woods, RE (2002). Digital Image Processing(2nd ed.). Prentice Hall.

[0008] [2] Dong, C., Loy, CC, He, K.,&Tang, X. (2014). Learning a deepconvolutional network for image super-resolution. In Proceedings of the IEEEconference on computer vision and pattern recognition (CVPR) (pp. 1848-1857).

[0009] [3] Kim, J., Kwon Lee, J.,&Mu Lee, K. (2016). Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEEconference on computer vision and pattern recognition (CVPR) (pp. 1646-1654)。

[0010] [4] Shi, W., Caballero, J., Ledig, C., Zeng, X.,&Wang, Z. (2016).Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference oncomputer vision and pattern recognition (CVPR) (pp. 1874-1883)。

[0011] [5] Ledig, C., Theis, L., Huszár, F., Caballero, J., Totz, J.,&Shi,W. (2017). Photo-realistic single image super-resolution using a generativeadversarial network. In Proceedings of the IEEE conference on computer visionand pattern recognition (CVPR) (pp. 4681-4690)。

[0012] [6] Johnson, J., Alahi, A.,&Fei-Fei, L. (2016). Perceptual losses forreal-time style transfer and super-resolution. In European conference oncomputer vision (ECCV) (pp. 694-711).

[0013] [7] Chambolle, A. (2004). An algorithm for total variation minimization and applications. Journal of mathematical imaging and vision, 20(1-2), 89-97.

[0014] [8] Hu, J., Shen, L.,&Sun, G. (2018). Squeeze-and-excitationnetworks. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR) (pp. 7132-7141). Summary of the Invention

[0015] In order to solve the problems of insufficient noise suppression and limited image quality improvement in existing super-resolution image reconstruction methods when processing low-resolution images, the present invention proposes a super-resolution reconstruction method and system based on wavelet transform and differential sliding window feature extraction. Through wavelet transform and differential operation, common mode noise is effectively removed, and artifact noise in super-resolution images is significantly reduced, thereby improving image reconstruction quality and achieving higher clarity and detail restoration effects.

[0016] According to one aspect of the present invention, a super-resolution reconstruction method based on wavelet transform and differential sliding window feature extraction is provided, comprising:

[0017] Based on the input low-resolution image, the forward original image is obtained by pixel replication;

[0018] Based on the forward original image, generating a reverse original image by differential modeling;

[0019] Performing discrete wavelet transform based on the forward original image and the reverse original image to obtain multiple groups of differential mode wavelet subband data;

[0020] Inputting the plurality of groups of differential mode wavelet subband data into the corresponding trained differential sliding window feature extraction model respectively to obtain a plurality of wavelet subbands of the high-resolution image;

[0021] An inverse wavelet transform is performed based on the plurality of wavelet sub-bands of the high-resolution image to obtain a reconstructed high-resolution image.

[0022] As a further technical solution, the differential sliding window feature extraction model is constructed based on the SwinIR model, including a shallow feature extraction layer, a deep feature extraction layer and a high-quality image wavelet subband reconstruction layer; wherein, the shallow feature extraction layer includes a convolutional network layer and a LeakyReLU activation layer, the number of input channels of the convolutional network layer corresponds to the number of channels of the original image wavelet subband, and the number of output channels is the dimension of the shallow feature tensor; the deep feature extraction layer includes several residual modules and a convolutional network layer; the high-quality image wavelet subband reconstruction layer includes a jump connection module, an upsampling module and a convolutional network layer, the jump connection module directly adds the shallow feature tensor and the deep feature tensor, the number of input channels of the convolutional network layer is the dimension of the shallow feature tensor, and the number of output channels is the number of channels of the high-resolution wavelet subband.

[0023] As a further technical solution, the deep feature extraction layer further includes:

[0024] An offset window multi-head self-attention mechanism is introduced in the residual module to adjust the window position offset in adjacent layers.

[0025] As a further technical solution, the training of the differential sliding window feature extraction model includes:

[0026] Construct a dataset containing low-resolution images and high-resolution target images;

[0027] Obtaining the forward original images of the low-resolution image and the high-resolution target image in the data set, performing differential mode operations on the forward original images of the two, generating corresponding reverse original images, and forming forward and reverse original image data of the differential mode and forward and reverse high-resolution target image data of the differential mode;

[0028] Perform discrete wavelet transform on the forward and reverse original images respectively to obtain multiple sets of differential mode wavelet subband data;

[0029] Inputting the plurality of groups of wavelet subbands of the differential mode and the wavelet subbands of the corresponding high-resolution image into corresponding plurality of differential sliding window feature extraction models for training, wherein the differential sliding window feature extraction models are constructed based on the SwinIR model;

[0030] Use the back-propagation algorithm to continuously adjust and optimize the model parameters and output the trained differential sliding window feature extraction model.

[0031] As a further technical solution, after inputting the low-resolution image, the following is further included:

[0032] The input low-resolution image is normalized to generate a forward original image based on the normalized low-resolution image.

[0033] As a further technical solution, before performing discrete wavelet transform on the inverse original image, the following is also included:

[0034] The pixel values of the reversed original image are The mapping relationships are converted one by one.

[0035] As a further technical solution, discrete wavelet transform is performed based on the forward original image and the reverse original image, including:

[0036] The forward original image and the reverse original image are subjected to a two-dimensional discrete wavelet transform using Daubechies3 wavelet to obtain four wavelet subbands.

[0037] As a further technical solution, after obtaining multiple wavelet sub-bands of the high-resolution image, the following is further included:

[0038] The difference operation is performed on multiple wavelet subbands of the obtained high-resolution image.

[0039] According to one aspect of the present invention, a super-resolution reconstruction system based on wavelet transform and differential sliding window feature extraction is provided, comprising:

[0040] The first main module is used to obtain a positive original image by pixel replication based on the input low-resolution image;

[0041] A second main module is configured to generate a reverse original image by differential modeling based on the forward original image;

[0042] A third main module is configured to perform discrete wavelet transform based on the forward original image and the reverse original image to obtain multiple sets of differential mode wavelet subband data;

[0043] The fourth main module is used to input the plurality of groups of differential mode wavelet subband data into the corresponding trained differential sliding window feature extraction model to obtain a plurality of wavelet subbands of the high-resolution image;

[0044] The fifth main module is configured to perform inverse wavelet transform based on the plurality of wavelet sub-bands of the high-resolution image to obtain a reconstructed high-resolution image.

[0045] According to one aspect of the present invention, a computing device is provided, comprising: one or more processors; a storage device 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 implement the steps of the super-resolution reconstruction method based on wavelet transform and differential sliding window feature extraction.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The present invention uses a super-resolution reconstruction method based on wavelet transform and differential sliding window feature extraction to perform super-resolution reconstruction on low-resolution images. First, image preprocessing is used to normalize and replicate pixels on the input low-resolution image to obtain standardized image data and forward original image data, thereby providing stable input for subsequent processing. Next, a differential operation is performed to extract the reverse original image data. Then, a discrete wavelet transform is performed on the forward and reverse original images. Subsequently, the obtained multiple sets of differential mode wavelet subband data are respectively input into a trained differential sliding window feature extraction model to obtain wavelet subbands of the high-resolution image. Finally, the multiple sets of wavelet subband data of the forward and reverse high-resolution images are combined through a differential operation to remove common-mode noise, and an inverse discrete wavelet transform is performed to ultimately generate a high-quality high-resolution image. The present invention effectively improves the quality of super-resolution image reconstruction and effectively suppresses noise and artifacts.

[0048] The present invention constructs a differential sliding window feature extraction model based on the SwinIR model, combining the deep feature extraction capability of the SwinIR model, the multi-resolution characteristics of the wavelet transform and the noise suppression advantage of the differential operation. Through the differential processing of the wavelet subband data of the differential mode image, the image quality is effectively improved, the artifact noise is removed and the image detail performance is enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, a brief introduction will be given below to the drawings used in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0051] Figure 2 This is a structural diagram of the windowed self-attention network image restoration model (SwinIR) according to an embodiment of the present invention;

[0052] Figure 3This is a training flowchart of the windowed self-attention network image restoration model (SwinIR) according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, 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 part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of 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. In addition, the technical features in the various embodiments or single embodiments provided by the present invention are arbitrarily combined with each other to form a new technical solution. This combination is not restricted by the sequence of steps and / or structural composition mode, but must be based on the ability of ordinary technicians in this field to implement it. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that this combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0054] The embodiment of the present invention takes a given image dataset to be tested as an example to further illustrate the technical solution of the present invention. Figure 1 The embodiment of the present invention provides a super-resolution reconstruction method based on wavelet transform and differential sliding window feature extraction, comprising the following steps:

[0055] Step S1: Image preprocessing. Input a low-resolution image, perform normalization preprocessing on the input image, and then use pixel replication to make the image reach the target high resolution to obtain the forward original image. .

[0056] In one embodiment, the normalization preprocessing operation utilizes maximum and minimum value normalization to adjust the pixel value range, ensuring that the image data has a uniform scale and distribution during subsequent processing and suppressing noise and outliers in the input data, providing standardized data input for subsequent feature extraction and differential mode processing, thereby improving the stability and optimization efficiency of model training.

[0057] In one embodiment, the pixel copying method is to simply copy each pixel value in the original image to multiple pixel positions in the new image, so that the input image data reaches the target high resolution, so that the number of terms of subsequent wavelet transform coefficients is the same, which facilitates wavelet transform and its inverse transform.

[0058] Step S2: input a differential mode image generator, and use the differential mode image generator to extract reverse image data from the input forward original image data to obtain a reverse original image.

[0059] In one embodiment, the differential image generator performs differential operation on the normalized forward original image to generate a reverse original image. The reverse original image is obtained by performing differential operation on all pixel values of the forward original image according to The mapping relationships are converted one by one.

[0060] Step S3: performing wavelet transform, performing discrete wavelet transform on the forward and reverse original image data to obtain four sets of differential mode wavelet sub-band data.

[0061] In one embodiment, before performing wavelet transform on the reverse original image, its pixel values need to be converted into The mapping relationship is converted again one by one, so that the forward and reverse original image data present an opposite numerical relationship and a differential mode relationship after wavelet transformation.

[0062] In one embodiment, the discrete wavelet transform is to perform a two-dimensional discrete wavelet transform on the image using Daubechies3 wavelet to obtain four wavelet sub-bands, namely LL, LH, HL, and HH, which respectively contain the main low-frequency and high-frequency features of the image.

[0063] Step S4: The four groups of differential mode wavelet sub-band data are input into four differential sliding window feature extraction models respectively, and four groups of differential mode wavelet sub-band data are output.

[0064] See also Figure 2 The differential sliding window feature extraction model is constructed based on the SwinIR model, including a shallow feature extraction layer, a deep feature extraction layer and a high-quality image wavelet subband reconstruction layer.

[0065] In one embodiment, the shallow feature extraction layer includes a convolutional network layer and a LeakyReLU activation layer. The number of input channels of the convolutional network layer corresponds to the number of channels in the original image wavelet subband, and the number of output channels is the dimension of the shallow feature tensor. The deep feature extraction layer consists of six RSTB modules (residual modules) and one 3×3 convolutional network layer. The high-quality image wavelet subband reconstruction layer includes one skip connection module, one upsampling module, and one convolutional network layer. The skip connection module directly adds the shallow feature tensor to the deep feature tensor. The number of input channels of the convolutional network layer is the dimension of the shallow feature tensor, and the number of output channels is the number of channels in the high-resolution wavelet subband.

[0066] In one embodiment, the deep feature extraction layer processes the following: the input image signal is passed to the offset window self-attention enhanced feature extraction module for processing. This module uses a residual block (RSTB) module to extract features from the signal. First, the input features are processed through a window multi-head self-attention mechanism (W-MSA) to extract details. The image is divided into small 7×7 windows, and the self-attention of each window is independently calculated. Then, the offset window multi-head self-attention mechanism (SW-MSA) is used to adjust the window position by 3×3 in adjacent layers to capture more global information. The input features are then added to the processed features to further enhance the information representation. Finally, after nonlinear transformation by a multi-layer perceptron (MLP) module, the features are integrated through a 3×3 convolutional layer to output deep features, providing richer feature information for subsequent processing.

[0067] See also Figure 3 In one embodiment, the four differential sliding window feature extraction models are pre-trained, and the training process includes the following sub-steps:

[0068] In step S4.1, the DIV2K dataset containing low-resolution images and high-resolution target images is selected as input data.

[0069] Step S4.2 processes the dataset, including non-overlapping cropping, normalization, and pixel replication of the input image. Non-overlapping cropping divides the image into multiple equal-sized blocks, normalization adjusts the pixel values to a range between 0 and 1, and pixel replication simply stretches the low-resolution image to a high-resolution one. The processed dataset is further divided into training and test sets.

[0070] Step S4.3, generate forward original images for the dataset images, and input the generated forward original images into the differential mode image generator respectively to obtain the forward and reverse original image data of the differential mode and the forward and reverse high-resolution target image data of the differential mode.

[0071] In step S4.4, the forward and reverse images are subjected to two-dimensional discrete wavelet transform based on db3 wavelet, respectively, to obtain four sets of differential mode wavelet subband data, namely LL, LH, HL, and HH.

[0072] In step S4.5, the four groups of wavelet subbands of the four differential modes and the wavelet subbands corresponding to the high-resolution image are respectively input into the corresponding four differential sliding window feature extraction models constructed based on the SwinIR model for training.

[0073] In step S4.6, the back propagation algorithm is used to continuously adjust and optimize the parameters of the model.

[0074] In one embodiment, in step S4.6, during the training process, the performance of the model is evaluated by comparing the difference between the wavelet subband of the forward high-resolution image and the wavelet subband of the target image with the difference between the wavelet subband of the reverse high-resolution image output by the model and the wavelet subband of the reverse target image. The loss value is passed back to each layer of the model through the backpropagation algorithm, and the model parameters are adjusted layer by layer, including the convolution layer weights of the shallow feature extraction module and the weights and biases of the attention mechanism in the deep feature extraction module RSTB. The optimizer updates these parameters according to the set learning rate, so that the generated high-resolution image gradually approaches the target image. This process is repeated until the loss value stabilizes and the performance index on the validation set reaches the best, indicating that the model training is complete and has converged.

[0075] Step S5: removing common mode noise from the differential output signal.

[0076] In one embodiment, in step S5, a differential operation is performed on the four sets of differential mode wavelet subband data output by the differential sliding window feature extraction model to remove common mode noise, thereby obtaining four wavelet subband data of the high-resolution image. The differential operation is performed by subtracting and normalizing the four sets of wavelet subband data of the high-resolution reconstructed differential mode image.

[0077] Step S6: Inverse discrete wavelet transform to output a high-resolution image. Perform inverse discrete wavelet transform on the four wavelet subband data of the high-resolution image to obtain the final high-resolution image.

[0078] In summary, the present invention can achieve super-resolution restoration of low-resolution images, effectively removing image noise and improving image quality. The method can significantly reduce artifact noise and achieve excellent performance in a variety of image restoration tasks. The present invention can also be widely applied to image denoising, image deblurring, medical image enhancement, and other fields, and has broad application prospects.

[0079] The implementation basis of each embodiment of the present invention is achieved through programmed processing by a device with processor functionality. Therefore, in practical engineering, the technical solutions and functions of each embodiment of the present invention are encapsulated into various modules. Based on this reality, and in addition to the above-mentioned embodiments, an embodiment of the present invention provides a super-resolution reconstruction system based on wavelet transform and differential sliding window feature extraction. This system is used to implement the super-resolution reconstruction method based on wavelet transform and differential sliding window feature extraction described in the above-mentioned method embodiments.

[0080] The system includes: a first main module, used to obtain a forward original image by pixel replication based on an input low-resolution image; a second main module, used to generate a reverse original image by differential modeling based on the forward original image; a third main module, used to perform discrete wavelet transform based on the forward original image and the reverse original image to obtain multiple groups of differential mode wavelet subband data; a fourth main module, used to input the multiple groups of differential mode wavelet subband data into the corresponding trained differential sliding window feature extraction model to obtain multiple wavelet subbands of the high-resolution image; and a fifth main module, used to perform inverse wavelet transform based on the multiple wavelet subbands of the high-resolution image to obtain a reconstructed high-resolution image.

[0081] The super-resolution reconstruction system based on wavelet transform and differential sliding window feature extraction provided by the embodiment of the present invention addresses the problems of insufficient noise suppression and limited image quality improvement in existing super-resolution image reconstruction methods when processing low-resolution images. It adopts the aforementioned several modules to effectively remove common mode noise through wavelet transform and differential operations, significantly reducing artifact noise in super-resolution images, thereby improving image reconstruction quality and achieving higher clarity and detail restoration effects.

[0082] It should be noted that the system embodiments provided by the present invention are not only used to implement the methods in the above-mentioned method embodiments, but also used to implement the methods in other method embodiments provided by the present invention. The only difference lies in the setting of corresponding functional modules, and the principles thereof are basically the same as the principles of the above-mentioned system embodiments provided by the present invention. As long as those skilled in the art refer to the specific technical solutions in other method embodiments on the basis of the above-mentioned system embodiments, obtain corresponding technical means and technical solutions composed of these technical means by combining technical features, and on the premise of ensuring the practicality of the technical solutions, improve the modules in the above-mentioned system embodiments to obtain corresponding system class embodiments for implementing the methods in other method class embodiments.

[0083] Based on the same inventive concept as the aforementioned embodiment, an embodiment of the present invention also provides a computing device, comprising: one or more processors; a storage device for storing one or more programs, which, when executed by the one or more processors, enables the one or more processors to implement the steps of the super-resolution reconstruction method based on wavelet transform and differential sliding window feature extraction.

[0084] In an embodiment of the present invention, the memory may be a non-volatile memory, such as a hard disk drive (HDD) or a solid-state drive (SSD), or a volatile memory (volatile memory), such as a random-access memory (RAM). The memory is any other medium that can be used to carry or store desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory in an embodiment of the present invention may also be a circuit or any other device that can implement a storage function, for storing program instructions and / or data.

[0085] In the embodiments of the present invention, the processor may be a general-purpose processor, a digital signal processor, an application-specific integrated circuit, a field-programmable gate array or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component, and may implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. A general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of the present invention may be directly implemented and executed by a hardware processor, or by a combination of hardware and software modules within the processor.

[0086] In summary, the present invention first preprocesses the input image to obtain standardized image data; then inputs the image into a differential image generator to extract the reverse original image, and the forward and reverse original images constitute a set of differential image data; then performs a two-dimensional discrete wavelet transform on it; then respectively inputs the four sets of differential wavelet subband data into four corresponding trained differential sliding window feature extraction models constructed based on the SwinIR model, the differential sliding window feature extraction model includes a layer feature extraction layer, a deep feature extraction layer, and a high-quality image wavelet subband reconstruction layer, the deep feature extraction layer includes six residual blocks and a convolution layer; finally, the four sets of wavelet subband data of the output forward and reverse high-resolution images are combined by a differential operation to remove common mode noise, and a discrete wavelet inverse transform is performed to finally generate a high-quality high-resolution image. The present invention combines the deep feature extraction capability of the SwinIR model, the multi-resolution characteristics of the wavelet transform, and the noise suppression advantage of the differential operation, effectively improving the quality of super-resolution image reconstruction and effectively suppressing noise and artifacts.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. A super-resolution reconstruction method based on wavelet transform and differential sliding window feature extraction, characterized in that: include: Based on the input low-resolution image, the forward original image is obtained by pixel replication; Based on the forward original image, generating a reverse original image by differential modeling; Performing discrete wavelet transform based on the forward original image and the reverse original image to obtain multiple groups of differential mode wavelet subband data; inputting the plurality of sets of differential mode wavelet subband data into the corresponding trained differential sliding window feature extraction model respectively, performing differential operations on the plurality of sets of differential mode wavelet subband data output by the differential sliding window feature extraction model to remove common mode noise, thereby obtaining a plurality of sets of wavelet subband data of a high-resolution image; An inverse wavelet transform is performed based on the plurality of wavelet sub-bands of the high-resolution image to obtain a reconstructed high-resolution image.

2. The super-resolution reconstruction method based on wavelet transform and differential sliding window feature extraction according to claim 1, characterized in that: The differential sliding window feature extraction model is constructed based on the SwinIR model, and includes a shallow feature extraction layer, a deep feature extraction layer and a high-quality image wavelet subband reconstruction layer; wherein, the shallow feature extraction layer includes a convolutional network layer and a LeakyReLU activation layer, the number of input channels of the convolutional network layer corresponds to the number of channels of the original image wavelet subband, and the number of output channels is the dimension of the shallow feature tensor; the deep feature extraction layer includes several residual modules and a convolutional network layer; the high-quality image wavelet subband reconstruction layer includes a skip connection module, an upsampling module and a convolutional network layer, the skip connection module directly adds the shallow feature tensor and the deep feature tensor, the number of input channels of the convolutional network layer is the dimension of the shallow feature tensor, and the number of output channels is the number of channels of the high-resolution wavelet subband.

3. The super-resolution reconstruction method based on wavelet transform and differential sliding window feature extraction according to claim 2, characterized in that: The deep feature extraction layer further includes: An offset window multi-head self-attention mechanism is introduced in the residual module to adjust the window position offset in adjacent layers.

4. The super-resolution reconstruction method based on wavelet transform and differential sliding window feature extraction according to claim 2, characterized in that: The training of the differential sliding window feature extraction model includes: Construct a dataset containing low-resolution images and high-resolution target images; Obtaining the forward original images of the low-resolution image and the high-resolution target image in the data set, performing differential mode operations on the forward original images of the two, generating corresponding reverse original images, and forming forward and reverse original image data of the differential mode and forward and reverse high-resolution target image data of the differential mode; Perform discrete wavelet transform on the forward and reverse original images respectively to obtain multiple sets of differential mode wavelet subband data; Inputting the plurality of groups of wavelet subbands of the differential mode and the wavelet subbands of the corresponding high-resolution image into corresponding plurality of differential sliding window feature extraction models for training, wherein the differential sliding window feature extraction models are constructed based on the SwinIR model; Use the back-propagation algorithm to continuously adjust and optimize the model parameters and output the trained differential sliding window feature extraction model.

5. The super-resolution reconstruction method based on wavelet transform and differential sliding window feature extraction according to claim 1, characterized in that: After inputting the low-resolution image, it also includes: The input low-resolution image is normalized to generate a forward original image based on the normalized low-resolution image.

6. The super-resolution reconstruction method based on wavelet transform and differential sliding window feature extraction according to claim 1, characterized in that: Before performing discrete wavelet transform on the inverse original image, it also includes: The pixel values of the reversed original image are The mapping relationships are converted one by one.

7. The super-resolution reconstruction method based on wavelet transform and differential sliding window feature extraction according to claim 1, characterized in that: Performing discrete wavelet transform based on the forward original image and the reverse original image includes: The forward original image and the reverse original image are subjected to a two-dimensional discrete wavelet transform using Daubechies3 wavelet to obtain four wavelet subbands.

8. The super-resolution reconstruction method based on wavelet transform and differential sliding window feature extraction according to claim 1, characterized in that: After obtaining multiple wavelet subbands of high-resolution images, it also includes: The difference operation is performed on multiple wavelet subbands of the obtained high-resolution image.

9. A super-resolution reconstruction system based on wavelet transform and differential sliding window feature extraction, characterized in that: include: The first main module is used to obtain a positive original image by pixel replication based on the input low-resolution image; A second main module is configured to generate a reverse original image by differential modeling based on the forward original image; A third main module is configured to perform discrete wavelet transform based on the forward original image and the reverse original image to obtain multiple sets of differential mode wavelet subband data; a fourth main module, configured to input the plurality of sets of differential mode wavelet subband data into the corresponding trained differential sliding window feature extraction model, perform differential operations on the plurality of sets of differential mode wavelet subband data output by the differential sliding window feature extraction model, remove common mode noise, and obtain a plurality of sets of wavelet subband data for a high-resolution image; The fifth main module is configured to perform inverse wavelet transform based on the plurality of wavelet sub-bands of the high-resolution image to obtain a reconstructed high-resolution image.

10. A computing device, characterized in that include: one or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, enables the one or more processors to implement the steps of the super-resolution reconstruction method based on wavelet transform and differential sliding window feature extraction as described in any one of claims 1 to 8.

Images