An image super-resolution balancing learning method for pixel distribution imbalance in an image

Abstract

The application discloses an image super-resolution balancing learning method for pixel distribution imbalance in an image, introduces static image structure priori and learnable image structure priori, jointly indicates pixel distribution of the image in a training process, and reweights each pixel based on the indication, gives a loss function value of a minority of high-frequency area pixels a greater learning weight, simultaneously reduces a learning weight of a loss function value of a majority of low-frequency area pixels, effectively balances a fitting degree of an image super-resolution neural network model to pixels in different areas, and thus obtains a more balanced image super-resolution neural network model, and the image super-resolution neural network model outputs a super-resolution result, which can effectively relieve a twin fitting problem encountered in image super-resolution.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a super-resolution balancing learning method for images with unbalanced pixel distribution. Background Technology

[0002] Single-image super-resolution (SIR) aims to reconstruct a corresponding high-resolution image from a low-resolution image and is an important technique in image processing and computer vision. Over the past few decades, numerous SIR methods have been proposed and widely implemented in real-world applications, such as video surveillance, face recognition, and medical imaging. Currently, deep learning-based SIR methods have gradually become mainstream due to their superior performance. These methods primarily use neural network models to model the mapping from a low-resolution image to its corresponding high-resolution image, learning and optimizing this neural network model from a large number of natural images. Different methods employ different neural network models. To reconstruct better high-resolution images from input low-resolution images, current SIR methods are mainly developing in two directions: first, designing more powerful general neural network models to better learn the mapping. For example, SWinIR introduces the state-of-the-art Transformer model to model local concerns and long-range dependencies, thereby learning the mapping better; second, since the main difference between low-resolution and high-resolution images lies in image details, special neural network models are designed to enhance image details while learning the mapping, thus obtaining better high-resolution images. For example, PCL proposed a neural network model based on contrastive learning to convert low-resolution images into high-resolution images with clear details and realistic appearance.

[0003] A common observation in image super-resolution is that low-frequency regions of an image are naturally easier to super-resolution than high-frequency regions. To achieve better super-resolution results in high-frequency regions, existing methods either design larger and stronger general neural network models or design special neural network models to enhance high-frequency regions of the image. Although these methods have achieved good performance, they all ignore a natural factor from the data level: the imbalance in pixel distribution in the image. Specifically, in natural images, the number of pixels in low-frequency regions is far greater than the number of pixels in high-frequency regions. Since image super-resolution neural network models are usually trained using pixel-wise loss between the output image and the target image, the majority of pixels in low-frequency regions will overwhelm the minority of pixels in high-frequency regions, thus dominating the model learning process. This results in the learned model performing better in low-frequency regions of the image, i.e., the model overfits to low-frequency regions and underfits to high-frequency regions. Here, this invention refers to the model overfitting / underfitting problem caused by the imbalance in pixel distribution in the image as the twin fitting problem.

[0004] Given a high-resolution image y and its corresponding low-resolution image x, the image super-resolution neural network model f net (·) is usually optimized using a pixel-wise loss function, specifically the following formula:

[0005]

[0006] Where i, I, and p are the pixel-wise index, the total number of pixels, and the loss function exponent, respectively. Because this function treats each pixel's loss function value equally, the loss function values ​​for most low-frequency region pixels will be much larger than those for a few high-frequency region pixels. This causes the model to tend to learn image super-resolution oriented towards low-frequency regions, resulting in the aforementioned Siamese fitting problem. To solve this problem, the loss function needs to be reweighted pixel-wise, giving greater learning weight to the loss function values ​​of a few high-frequency region pixels while reducing the learning weight of the loss function values ​​of most low-frequency region pixels. Since the image frequency domain is local and continuous, and there are no clear boundaries between high and low frequency regions, it is impossible to directly classify a given pixel into a high-frequency or low-frequency region. Summary of the Invention

[0007] To address the aforementioned shortcomings in existing technologies, this invention provides an image super-resolution balancing learning method for images with unbalanced pixel distribution, which solves the problem of twin fitting encountered in image super-resolution.

[0008] To achieve the aforementioned objectives, the present invention employs the following technical solution: a super-resolution balancing learning method for images with imbalanced pixel distribution, comprising the following steps:

[0009] S1. Obtain the low-resolution image corresponding to the high-resolution image, perform image interpolation on the low-resolution image to obtain its corresponding interpolated image, and calculate the static image structure prior.

[0010] S2. Obtain the super-resolution result corresponding to the low-resolution image through the image super-resolution neural network model, and calculate the learnable image structure prior;

[0011] S3. Convert the static image structure prior and the learnable image structure prior into pixel-wise weights, and weight the image loss function to obtain the weighted loss function.

[0012] S4. Optimize the image super-resolution neural network model according to the weighted loss function, and obtain the super-resolution result corresponding to the low-resolution image through the optimized image super-resolution neural network model to complete the image super-resolution balance learning.

[0013] Furthermore, in S1, the image interpolation method is specifically the Bicubic image interpolation method.

[0014] Further: In S1, the static image structure prior y is calculated. sp The specific expression is:

[0015] y sp =|yf bic (x)|

[0016] In the formula, y represents the high-resolution image, x represents the low-resolution image, and f bic (x) is the interpolated image corresponding to the low-resolution image x, f bic (·) is the general formula for Bicubic image interpolation.

[0017] Further: S2 includes the following sub-steps:

[0018] S21. Input the low-resolution image into the image super-resolution neural network model to obtain the super-resolution result corresponding to the low-resolution image;

[0019] S22. Based on the super-resolution results and interpolated images corresponding to the low-resolution images, learnable image structure priors are obtained.

[0020] Furthermore: In step S22, a learnable image structure prior y is obtained. lp The specific expression is:

[0021] y lp =|yf bic (x)-f net (x)|

[0022] In the formula, f net (x) represents the super-resolution result corresponding to the low-resolution image x.

[0023] The beneficial effects of the above-mentioned further solutions are as follows: This invention introduces a static image structure prior y based on the target image. sp and learnable image structure prior y lp It can effectively indicate the pixel distribution of an image during the training process.

[0024] Further: S3 includes the following sub-steps:

[0025] S31. Normalize the static image structure prior and the learnable image structure prior using the normalization formula;

[0026] S32. Convert the normalized static image structure prior and the learnable image structure prior into pixel-wise weights.

[0027] S33. The loss function of the image is weighted according to the weight of each pixel to obtain the weighted loss function.

[0028] Furthermore: In S31, the expression for the normalized general formula g(z) is specifically as follows:

[0029]

[0030] In the formula, z is a static image structure prior or a learnable image structure prior;

[0031] In step S32, the pixel-wise weights include the static image structure prior pixel-wise weights and the learnable image structure prior pixel-wise weights. The specific expression for obtaining the pixel-wise weight W(z; α, γ) is as follows:

[0032] W(z;α,γ)=α*exp(γ*g(z))

[0033] In the formula, α is the first transformation parameter and γ is the second transformation parameter;

[0034] In S33, the weighted loss function is obtained. The specific expression is:

[0035]

[0036] In the formula, i is the pixel index, I is the total number of pixels, p is the exponent of the loss function, and W(y) sp ;α1,γ1) i For pixel-by-pixel static image structure prior y sp The weights, W(y) lp ;α2,γ2) i For pixel-by-pixel learnable image structure prior y lp The weights, α1 is the static image structure prior y sp The first transformation parameter, γ1, is the static image structure prior y. sp The second transformation parameter, α2, is the learnable image structure prior y. lp The first transformation parameter, γ2, is the learnable image structure prior y. lp The second transformation parameter, f bic (x) i f is the interpolated image corresponding to the pixel-by-pixel low-resolution image x. net () i For each pixel of the low-resolution image x, the super-resolution result is y. i This is a pixel-by-pixel high-resolution image.

[0037] The beneficial effects of the above-mentioned further scheme are as follows: the pixel-wise weights W(z; α, γ) can reweight the loss function pixel by pixel. Based on the weighted loss function, the image super-resolution neural network model can focus on learning the high-frequency regions of the image, thereby achieving a balance in the fitting degree of pixels in different regions and thus obtaining better super-resolution results.

[0038] Furthermore: S4 specifically refers to:

[0039] The image super-resolution neural network model is optimized based on the weighted loss function. Based on the optimized image super-resolution neural network model and the image interpolation method, the super-resolution result corresponding to the low-resolution image is obtained, and the image super-resolution balance learning is completed.

[0040] The beneficial effects of this invention are as follows: This invention provides an image super-resolution balancing learning method for images with unbalanced pixel distribution. By introducing static image structure priors and learnable image structure priors, the pixel distribution of the image during training is jointly indicated, and each pixel is reweighted based on this. The loss function values ​​of a few high-frequency region pixels are given greater learning weights, while the learning weights of the loss function values ​​of most low-frequency region pixels are reduced. This effectively balances the fitting degree of the image super-resolution neural network model to pixels in different regions, thereby obtaining a more balanced image super-resolution neural network model. The super-resolution results output by the image super-resolution neural network model according to this invention can effectively alleviate the twin fitting problem encountered in image super-resolution. Attached Figure Description

[0041] Figure 1 This is a flowchart of an image super-resolution balancing learning method for images with unbalanced pixel distribution, according to the present invention. Detailed Implementation

[0042] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0043] Example 1:

[0044] like Figure 1 As shown, in one embodiment of the present invention, an image super-resolution balancing learning method for images with imbalanced pixel distribution includes the following steps:

[0045] S1. Obtain the low-resolution image corresponding to the high-resolution image, perform image interpolation on the low-resolution image to obtain its corresponding interpolated image, and calculate the static image structure prior.

[0046] S2. Obtain the super-resolution result corresponding to the low-resolution image through the image super-resolution neural network model, and calculate the learnable image structure prior;

[0047] S3. Convert the static image structure prior and the learnable image structure prior into pixel-wise weights, and weight the image loss function to obtain the weighted loss function.

[0048] S4. Optimize the image super-resolution neural network model according to the weighted loss function, and obtain the super-resolution result corresponding to the low-resolution image through the optimized image super-resolution neural network model to complete the image super-resolution balance learning.

[0049] The present invention proposes an image super-resolution balancing learning method for imbalanced pixel distribution in natural images, which can effectively balance the fitting degree of the model to pixels in different regions, thereby obtaining a more balanced image super-resolution neural network model.

[0050] To address the twin fitting problem, this invention reweights the loss function pixel-by-pixel, giving greater learning weight to the loss function values ​​of a few high-frequency region pixels while reducing the learning weight of the loss function values ​​of most low-frequency region pixels. Since the image frequency domain is local and continuous, and lacks clear boundaries between high and low frequency regions, it is impossible to directly classify a given pixel into a high-frequency or low-frequency region. Therefore, this invention introduces a static image structure prior y based on the target image. sp and learnable image structure prior y lp The pixel distribution of the image during the joint instruction training process is determined, and each pixel is reweighted based on this.

[0051] In S1, the image interpolation method is specifically the Bicubic image interpolation method.

[0052] In S1, the static image structure prior y is calculated. sp The specific expression is:

[0053] y sp =|yf bic ()|

[0054] In the formula, y represents the high-resolution image, x represents the low-resolution image, and f bic () represents the interpolated image corresponding to the low-resolution image x, f bic (·) is the general formula for Bicubic image interpolation.

[0055] fbic () typically achieves better super-resolution performance in the low-frequency region than in the high-frequency region. This invention utilizes a static image structure prior y sp The value of y reflects the high and low frequency distribution of pixels. sp The larger the value at the corresponding pixel position, the more likely that pixel is located in a high-frequency region; sp The smaller the value at the corresponding pixel position, the more likely that pixel is located in the low-frequency region.

[0056] S2 includes the following steps:

[0057] S21. Input the low-resolution image into the image super-resolution neural network model to obtain the super-resolution result corresponding to the low-resolution image;

[0058] S22. Based on the super-resolution results and interpolated images corresponding to the low-resolution images, learnable image structure priors are obtained.

[0059] In step S22, a learnable image structure prior y is obtained. lp The specific expression is:

[0060] y lp =|yf bic (x)- net ()|

[0061] In the formula, f net () represents the super-resolution result corresponding to the low-resolution image x.

[0062] Due to the static image structure prior y sp The value of f reflects the high and low frequency distribution of pixels, but it cannot dynamically reflect the degree of fitting of the current model to pixels in different regions during the learning process. Furthermore, in the Siamese fitting problem, pixels in different regions are in competition with each other. Therefore, this invention will... bic () is introduced as a residual term into the image super-resolution neural network model f. net (·), thus obtaining the learnable image structure prior y lp This makes the image super-resolution neural network model f net (·) It can focus more on high-frequency regions and pixels that are difficult to fit, y lp The larger the value at the corresponding pixel position, the more difficult it is to fit that pixel; lp The smaller the value at the corresponding pixel position, the easier it is to fit that pixel.

[0063] S3 includes the following steps:

[0064] S31. Normalize the static image structure prior and the learnable image structure prior using the normalization formula;

[0065] S32. Convert the normalized static image structure prior and the learnable image structure prior into pixel-wise weights.

[0066] S33. The loss function of the image is weighted according to the weight of each pixel to obtain the weighted loss function.

[0067] In S31, the expression for the normalized general formula g(z) is specifically as follows:

[0068]

[0069] In the formula, z is a static image structure prior or a learnable image structure prior;

[0070] The normalization formula g(z) can normalize two structural priors to 0 to 1 and maintain the relative magnitude of the values. At the same time, in order to prevent the normalized structural priors from making the weights of some pixel loss values ​​too small, this invention introduces an exponential function exp(·) and two parameters α and γ to convert the normalized structural priors into pixel-wise weights.

[0071] In step S32, the pixel-wise weights include the static image structure prior pixel-wise weights and the learnable image structure prior pixel-wise weights. The specific expression for obtaining the pixel-wise weight W(z; α, γ) is as follows:

[0072] W(z;α,γ)=*exp(γ*g(z))

[0073] In the formula, α is the first transformation parameter and γ is the second transformation parameter;

[0074] The range of weights W(z; α, γ) is restricted to [α, a*e γ ], where α controls the lower limit of the weights and γ controls the upper limit of the weights, which makes the converted weights and weighting behavior more controllable;

[0075] This invention uses structural prior y sp and y lp The weights are converted to pixel-wise weights W(z; α, γ), and the loss function is reweighted pixel-wise, while f is introduced. bic (x) is used as the residual term to obtain the reweighted loss function.

[0076] In S33, the weighted loss function is obtained. The specific expression is:

[0077]

[0078] In the formula, i is the pixel index, I is the total number of pixels, p is the exponent of the loss function, and W(y) sp ;α1,γ1)i For pixel-by-pixel static image structure prior y sp The weights, W(y) lp ;α2,γ2) i For pixel-by-pixel learnable image structure prior y lp The weights, α1 is the static image structure prior y sp The first transformation parameter, γ1, is the static image structure prior y. sp The second transformation parameter, α2, is the learnable image structure prior y. lp The first transformation parameter, γ2, is the learnable image structure prior y. lp The second transformation parameter, f bic (x) i f is the interpolated image corresponding to the pixel-by-pixel low-resolution image x. net (x) i For each pixel of the low-resolution image x, the super-resolution result is y. i This is a pixel-by-pixel high-resolution image.

[0079] The weighted loss function assigns higher weights to high-frequency regions, especially pixels that are difficult to fit, so that the image super-resolution neural network model can focus on learning them, thereby achieving a balance in the fitting degree of pixels in different regions and thus obtaining better super-resolution results.

[0080] Specifically, S4 is:

[0081] The image super-resolution neural network model is optimized based on the weighted loss function. Based on the optimized image super-resolution neural network model and the image interpolation method, the super-resolution result corresponding to the low-resolution image is obtained, and the image super-resolution balance learning is completed.

[0082] Example 2:

[0083] This embodiment presents a specific experiment on a super-resolution balancing learning method for images with unbalanced pixel distribution.

[0084] To verify the effectiveness of this invention, three validations were performed on three different single-image super-resolution tasks: a 4x single-image super-resolution task, a 2x single-image super-resolution task, and a 4x single-image multi-degradation super-resolution task. Each validation covered five image datasets: Set5, Set14, BSD100, Manga109, and Urban100. PSNR and SSIM values ​​in the RGB color space were used as performance metrics. Since this invention is a super-resolution balancing learning method for images with imbalanced pixel distribution, and is not limited to a specific image super-resolution neural network model, each validation was performed on four to five image super-resolution neural network models: FSRCNN, CARN, SRResNet, MSRN, and SwinIR.

[0085] (1) 4x single image super-resolution task:

[0086] The DIV2K dataset was used as the training dataset, containing 800 2K resolution images. For each image, a bicubic downsampling method was first applied, downsampling the original 2K image by scaling factors of 0.9, 0.8, 0.7, and 0.6 to generate a high-resolution image. Then, a downsampling by scaling factor 0.25 was applied to obtain the corresponding low-resolution image. All low-resolution and high-resolution images were densely cropped to sizes of 32×32 and 128×128, respectively, to obtain 1.59 million low-resolution and high-resolution image patch pairs. Finally, all image patches were further augmented during training by flipping and rotating. The experimental results are as follows:

[0087]

[0088]

[0089] As shown in the table above, the method of the present invention can improve the performance of various image super-resolution neural network models on 4x single image super-resolution tasks on multiple test datasets. In other words, it can effectively alleviate the twin fitting problem encountered in image super-resolution, namely the problem of the model overfitting the low-frequency region and underfitting the high-frequency region, so that the model can achieve a better balance between fitting the low-frequency region and fitting the high-frequency region, thereby achieving better performance.

[0090] (2) 2x single image super-resolution task:

[0091] The DIV2K dataset was used as the training dataset, containing 800 2K resolution images. For each image, a bicubic downsampling method was first applied, downsampling the original 2K image by scaling factors of 0.9, 0.8, 0.7, and 0.6 to generate a high-resolution image. Then, a downsampling by scaling factor 0.5 was applied to obtain the corresponding low-resolution image. All low-resolution and high-resolution images were densely cropped to sizes of 64×64 and 128×128, respectively, to obtain 1.59 million low-resolution and high-resolution image patch pairs. Finally, all image patches were further augmented during training by flipping and rotating. The experimental results are as follows:

[0092]

[0093]

[0094] As can be seen from the table, this invention can improve the performance of various image super-resolution neural network models on 2x single image super-resolution tasks on multiple test datasets. In other words, it can effectively alleviate the twin fitting problem encountered in image super-resolution, namely, the model overfits the low-frequency region and underfits the high-frequency region, so that the model can achieve a better balance between fitting the low-frequency region and fitting the high-frequency region, thereby achieving better performance.

[0095] (3) 4x single image multi-degradation super-resolution task:

[0096] The DIV2K dataset was used as the training dataset, containing 800 2K resolution images. For each image, a bicubic downsampling method was first applied, downsampling the original 2K image by scaling factors of 0.9, 0.8, 0.7, and 0.6 to generate a high-resolution image. Then, a downsampling factor of 0.25 was applied to obtain the corresponding low-resolution image. All low-resolution and high-resolution images were densely cropped to sizes of 32×32 and 128×128, respectively, resulting in 1.59 million low-resolution and high-resolution image patch pairs. Simultaneously, the 32×32 low-resolution image patches were blurred using a 3×3 blur kernel with a standard deviation of 5, and Gaussian noise with a standard deviation of 5 was added. Finally, all image patches were further augmented during training by flipping and rotating. Experimental results are as follows:

[0097]

[0098]

[0099] As can be seen from the table, this invention can improve the performance of various image super-resolution neural network models on 4x single-image multi-degradation super-resolution tasks on multiple test datasets. In other words, it can effectively alleviate the twin fitting problem encountered in image super-resolution, namely, the model overfits the low-frequency region and underfits the high-frequency region, so that the model can achieve a better balance between fitting the low-frequency region and fitting the high-frequency region, thereby achieving better performance.

[0100] The beneficial effects of this invention are as follows: This invention provides an image super-resolution balancing learning method for images with unbalanced pixel distribution. By introducing static image structure priors and learnable image structure priors, the pixel distribution of the image during training is jointly indicated, and each pixel is reweighted based on this. The loss function values ​​of a few high-frequency region pixels are given greater learning weights, while the learning weights of the loss function values ​​of most low-frequency region pixels are reduced. This effectively balances the fitting degree of the image super-resolution neural network model to pixels in different regions, thereby obtaining a more balanced image super-resolution neural network model. The super-resolution results output by the image super-resolution neural network model according to this invention can effectively alleviate the twin fitting problem encountered in image super-resolution.

[0101] In the description of this invention, it should be understood that the terms "center," "thickness," "upper," "lower," "horizontal," "top," "bottom," "inner," "outer," and "radial," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying the relative importance or the number of technical features implicitly specified. Therefore, a feature defined by "first," "second," and "third" may explicitly or implicitly include one or more of that feature.

Claims

1. A super-resolution image balancing learning method for images with imbalanced pixel distribution, characterized in that, Includes the following steps: S1. Obtain the low-resolution image corresponding to the high-resolution image, perform image interpolation on the low-resolution image to obtain its corresponding interpolated image, and calculate the static image structure prior. S2. Obtain the super-resolution result corresponding to the low-resolution image through the image super-resolution neural network model, and calculate the learnable image structure prior; S3. Convert the static image structure prior and the learnable image structure prior into pixel-wise weights, and weight the image loss function to obtain the weighted loss function. S4. Optimize the image super-resolution neural network model according to the weighted loss function, and obtain the super-resolution result corresponding to the low-resolution image through the optimized image super-resolution neural network model to complete the image super-resolution balance learning.

2. The super-resolution balancing learning method for images with imbalanced pixel distribution as described in claim 1, characterized in that, In S1, the image interpolation method is specifically the Bicubic image interpolation method.

3. The super-resolution balancing learning method for images with imbalanced pixel distribution as described in claim 2, characterized in that, In S1, the static image structure prior y is calculated. sp The specific expression is: y sp =|y-f bic ()| In the formula, y represents the high-resolution image, x represents the low-resolution image, and f bic () represents the interpolated image corresponding to the low-resolution image x, f bic (·) is the general formula for Bicubic image interpolation.

4. The super-resolution balancing learning method for images with imbalanced pixel distribution as described in claim 3, characterized in that, S2 includes the following steps: S21. Input the low-resolution image into the image super-resolution neural network model to obtain the super-resolution result corresponding to the low-resolution image; S22. Based on the super-resolution results and interpolated images corresponding to the low-resolution images, learnable image structure priors are obtained.

5. The super-resolution balancing learning method for images with imbalanced pixel distribution as described in claim 4, characterized in that, In step S22, a learnable image structure prior y is obtained. lp The specific expression is: y lp =|y-f bic (x)- net ()| In the formula, f net () represents the super-resolution result corresponding to the low-resolution image x.

6. The super-resolution balancing learning method for images with imbalanced pixel distribution as described in claim 1, characterized in that, S3 includes the following steps: S31. Normalize the static image structure prior and the learnable image structure prior using the normalization formula; S32. Convert the normalized static image structure prior and the learnable image structure prior into pixel-wise weights. S33. The loss function of the image is weighted according to the weight of each pixel to obtain the weighted loss function.

7. The super-resolution balancing learning method for images with imbalanced pixel distribution as described in claim 6, characterized in that, In S31, the expression for the normalized general formula g(z) is specifically as follows: In the formula, z is a static image structure prior or a learnable image structure prior; In step S32, the pixel-wise weights include the static image structure prior pixel-wise weights and the learnable image structure prior pixel-wise weights. The specific expression for obtaining the pixel-wise weight W(z; α, γ) is as follows: W(z;α,γ)=*exp(γ*g(z)) In the formula, α is the first transformation parameter and γ is the second transformation parameter; In S33, the weighted loss function is obtained. The specific expression is: In the formula, i is the pixel index, I is the total number of pixels, p is the exponent of the loss function, and W(y) sp ;α1,γ1) i For pixel-by-pixel static image structure prior y sp The weights, W(y) lp ;α2,γ2) i For pixel-by-pixel learnable image structure prior y lp The weights, α1 is the static image structure prior y sp The first transformation parameter, γ1, is the static image structure prior y. sp The second transformation parameter, α2, is the learnable image structure prior y. lp The first transformation parameter, γ2, is the learnable image structure prior y. lp The second transformation parameter, f bic (x) i f is the interpolated image corresponding to the pixel-by-pixel low-resolution image x. net (x) i For each pixel of the low-resolution image x, the super-resolution result is y. i This is a pixel-by-pixel high-resolution image.

8. The super-resolution balancing learning method for images with imbalanced pixel distribution as described in claim 1, characterized in that, Specifically, S4 is: The image super-resolution neural network model is optimized based on the weighted loss function. Based on the optimized image super-resolution neural network model and the image interpolation method, the super-resolution result corresponding to the low-resolution image is obtained, and the image super-resolution balance learning is completed.

Images