A method and system for super-resolution image reconstruction

Abstract

The present invention provides an image super-resolution reconstruction method and system, comprising: obtaining a low-resolution image; inputting the low-resolution image into a generator and a discriminator in sequence to obtain a reconstructed image; wherein the generator uses a convolutional layer to extract a shallow feature map of the low-resolution image, the shallow feature map passes through a residual dense branch to obtain a first feature map, the shallow feature passes through a minimum perceptible error branch to obtain a second feature map, the result of multiplying the first feature map and the second feature map and the second feature map are input into a residual block, after a residual operation is performed, a third feature map is obtained, and the third feature map is input into an upsampling block and a reconstruction block in sequence to obtain a preliminary reconstructed image. The image restoration quality is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image reconstruction, and in particular relates to an image super-resolution reconstruction method and system. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Image super-resolution (SR) reconstruction is to improve the resolution of an image, which is a process from a low-resolution (LR) image to a high-resolution (HR) image. The current super-resolution algorithms can be divided into three categories: interpolation-based SR reconstruction, reconstruction-based SR reconstruction, and learning-based SR reconstruction. In recent years, with the continuous development of deep learning, super-resolution reconstruction methods based on deep learning in learning-based SR reconstruction have made rapid progress. Deep learning uses convolutional neural networks to greatly improve the reconstruction effect. At present, super-resolution algorithms based on deep learning are mainly divided into the following categories: algorithms based on linear networks, algorithms based on residual networks, algorithms based on dense networks, algorithms based on recurrent networks, algorithms based on GAN networks, and algorithms based on attention networks.

[0004] The end-to-end super-resolution algorithm SRCNN based on CNN structure has driven the development of deep convolutional neural networks. For end-to-end networks, peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are mostly used to evaluate the SR reconstruction quality. However, these PSNR-oriented methods often output overly smoothed results without sufficient high-frequency details, because the PSNR metric is fundamentally inconsistent with the subjective evaluation of human observers. Summary of the invention

[0005] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides an image super-resolution reconstruction method and system, which reconstructs a more realistic high-resolution image and improves the image restoration quality.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A first aspect of the present invention provides an image super-resolution reconstruction method, comprising:

[0008] Get low-resolution images;

[0009] Input the low-resolution image into the generator and discriminator in sequence to obtain the reconstructed image;

[0010] Among them, the generator uses a convolutional layer to extract the shallow feature map of the low-resolution image. The shallow feature map is passed through the residual dense branch to obtain the first feature map, and the shallow feature is passed through the minimum perceptible error branch to obtain the second feature map. The result of multiplying the first feature map and the second feature map is input into the residual block with the second feature map. After the residual operation, the third feature map is obtained. After the third feature map is successively input into the upsampling block and the reconstruction block, a preliminary reconstructed image is obtained.

[0011] Furthermore, the minimum perceptible error branch is used to calculate the minimum perceptible error of each pixel in the feature map after the convolution of the shallow feature map to obtain a second feature map.

[0012] Furthermore, the minimum perceptible error is calculated as follows:

[0013] Calculating the visibility threshold and spatial masking effect of each pixel;

[0014] Take the minimum value of visibility threshold and spatial masking effect;

[0015] The difference between the sum of the visibility threshold and the spatial masking effect and the product of the gain parameter and the minimum value is the minimum perceptible error for each pixel.

[0016] Furthermore, the visibility threshold is related to the background brightness of each pixel.

[0017] Furthermore, the spatial masking effect is related to the brightness contrast of each pixel.

[0018] Furthermore, the residual dense branch is composed of a plurality of dense blocks, and each dense block is connected to a residual scaling block.

[0019] Furthermore, the residual dense branch is composed of multiple convolutional layers connected in sequence, and the output of each convolutional layer is used as the input of all subsequent convolutional layers.

[0020] A second aspect of the present invention provides an image super-resolution reconstruction system, comprising:

[0021] An image acquisition module is configured to: acquire a low-resolution image;

[0022] An image reconstruction module is configured to: input the low-resolution image into the generator and the discriminator in sequence to obtain a reconstructed image;

[0023] Among them, the generator uses a convolutional layer to extract the shallow feature map of the low-resolution image. The shallow feature map is passed through the residual dense branch to obtain the first feature map, and the shallow feature is passed through the minimum perceptible error branch to obtain the second feature map. The result of multiplying the first feature map and the second feature map is input into the residual block with the second feature map. After the residual operation, the third feature map is obtained. After the third feature map is successively input into the upsampling block and the reconstruction block, a preliminary reconstructed image is obtained.

[0024] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the image super-resolution reconstruction method as described above.

[0025] A fourth aspect of the present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in an image super-resolution reconstruction method as described above are implemented.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] The present invention provides an image super-resolution reconstruction method, which adds a just noticeable difference (JND) branch on the basis of Enhanced Super-Resolution Generative Adversarial Networks (ESRGAN), and arranges the JND branch and the residual dense branch (RRDB) in parallel to reconstruct a high-resolution image that is more realistic and more in line with human eye perception. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The accompanying drawings in the specification, which constitute a part of the present invention, 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 improper limitations on the present invention.

[0029] Figure 1 is a structural diagram of a generator according to a first embodiment of the present invention;

[0030] Figure 2 is a structural diagram of a residual dense branch according to the first embodiment of the present invention;

[0031] Figure 3 It is a schematic diagram of the reconstruction result of the first embodiment of the present invention. DETAILED DESCRIPTION

[0032] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0033] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0034] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0035] Embodiment 1

[0036] This embodiment provides an image super-resolution reconstruction method, which adds a just noticeable difference (JND) calculation structure based on an enhanced super-resolution generative adversarial network (ESRGAN), and reconstructs a high-resolution image that is more realistic and more in line with human eye perception based on a low-resolution image input.

[0037] like Figure 1 As shown, the specific steps include:

[0038] Step 1: Obtain a low-resolution image to be reconstructed.

[0039] Step 2: The low-resolution image to be reconstructed is input into the generator and discriminator in sequence. The generator includes the first convolutional layer, the second convolutional layer, the residual dense branch (RRDB), the minimum perceptible error branch (JND), the residual block, the upsampling layer (upsample) and the reconstruction block. The minimum perceptible error branch includes the convolutional layers and the minimum perceptible error blocks connected in sequence.

[0040] Specifically, a convolutional layer (conv) is used to extract a shallow feature map of a low-resolution image. The shallow feature map passes through a residual dense branch to obtain a first feature map. The shallow feature passes through a minimum perceptible error branch to obtain a second feature map. The result of multiplying the first feature map and the second feature map is input into a residual block with the second feature map. After a residual operation, a third feature map is obtained. The third feature map is sent to an upsampling block. Finally, the SR image is reconstructed through a reconstruction block (two convolutional layers) to obtain a preliminary reconstructed image. The preliminary reconstructed image is then sent to a discriminator to generate a more realistic SR image, i.e., the final reconstructed image.

[0041] The convolutional layer is used to extract the shallow feature map F0 of the low-resolution image.

[0042] F0=H0(I LR ) (1)

[0043] Among them, I LR Represents the input low-resolution image, H0() represents the convolution function, and the shallow feature map F0 is passed through the residual dense branch to obtain the first feature map F1.

[0044] F1=H RRDB (F0) (2)

[0045] Among them, H RRDB () represents the function of the residual dense branch. The shallow feature F0 is passed through the minimum perceptible error branch to obtain the second feature map F2.

[0046] F2=H JND (F0) (3)

[0047] Among them, H JND () represents the function of the minimum perceptible error branch. The first feature map F1 is multiplied by the second feature map F2 to obtain Fx.

[0048] Fx=F1×F2 (4)

[0049] Fx is then input into the residual block together with the second feature map F2, and after the residual operation, the third feature map F3 is obtained.

[0050] F3=Fx+F2 (5)

[0051] The third feature map F3 is sent to the upsampling block for upsampling operation.

[0052] F4=H US (F3) (6)

[0053] H US It represents the function of the upsampling module, F3 represents the feature after amplification, and finally the SR image is reconstructed by the reconstruction block to obtain a preliminary reconstructed image, which is then sent to the discriminator to generate a more realistic SR image, that is, the final reconstructed image.

[0054] Among them, RRDB combines multi-level residual networks and dense connections. The structure of RRDB is as follows Figure 2 As shown on the left, each RRDB is composed of three dense blocks, with a residual-in-residual structure, using residual learning structures at different levels. In addition, each dense block is connected to a residual scaling block, that is, a residual scaling is also used after each dense block, that is, the residual is reduced by multiplying a constant between 0 and 1, and then added to the main path to prevent instability. The structure of each dense block is as follows Figure 2As shown on the right, a dense block (DenseBlock) uses dense connections. The residual dense branch consists of multiple convolutional layers connected in sequence. The output of each convolutional layer is used as the input of all subsequent convolutional layers. That is, each layer has more than one input (except the first layer), and the output of each layer is also used as the input of all subsequent layers. This dense connection makes the network capacity higher.

[0055] The minimum perceptible error branch is used to calculate the minimum perceptible error (i.e., JND threshold) of each pixel in the feature map after the convolution of the shallow feature map to obtain the second feature map. The minimum perceptible error is used to represent the maximum image distortion that the human eye cannot perceive, reflecting the tolerance of the human eye to image changes. In the field of image processing, JND can be used to measure the sensitivity of the human eye to distortion in different areas of the image. Traditional image or video coding technology rarely takes into account the characteristics and psychological effects of the human visual system, so a large amount of visual redundant data is encoded and transmitted. In order to further improve the efficiency of coding, research on removing visual redundancy is needed. At present, an effective method to characterize visual redundancy is the minimum perceptible distortion model based on psychology and physiology, that is, changes that the human eye cannot perceive. Due to various shielding effects of the human eye, the human eye can only perceive changes that exceed a certain threshold, which is the just perceptible difference of the human eye.

[0056] For any image F (i.e. shallow feature map), take its local area R∈F, the direction of each pixel x∈R is related to the gradient, and take the gradient direction of x as its orientation θ:

[0057]

[0058] Among them, G v (x) represents the gradient change of pixel x along the vertical direction, G h (x) represents the gradient change of pixel x along the horizontal direction. The complexity of each pattern R is directly related to the interaction between θ(x). If a pattern R has many different θ(x), then it is very likely to be an irregular pattern, which has complex interactions and exhibits a strong masking effect. If a pattern R has very similar θ(x), then it is very likely to be a regular pattern, and its masking effect is not so strong. Therefore, the complexity of a pattern can be regarded as the diversity of all θ(x)∈R. It is known that the histogram of regular patterns is sparse, while the histogram of irregular patterns is dense. Therefore, the pattern complexity C of each pixel x p Calculated as the sparsity of the corresponding histogram:

[0059]

[0060] Among them, ||·||0 represents the L0 norm, H k(x) is the histogram corresponding to pixel x:

[0061]

[0062] Where δ(·) is the impulse function:

[0063]

[0064] in, It is θ(x) quantized with T=12° as the interval.

[0065] The pattern masking effect is determined by both brightness contrast and pattern complexity. Therefore, in pattern masking estimation, the JND model considers both brightness contrast C l And the pattern complexity C p , pattern complexity C p As shown in formula (8), the brightness contrast C of each pixel x is l It can be calculated by the following formula:

[0066]

[0067] Among them, G v and G h is the gradient size along the vertical and horizontal directions. Experiments have found that in the network structure, the brightness contrast C is used at the same time. l And the pattern complexity C p Therefore, using brightness contrast C alone will not produce beneficial results. l Or pattern complexity C p , it is found through experiments that when JND is applied to the network structure, the brightness contrast C is used alone l The effect is better.

[0068] Therefore, in the present invention, the method for calculating the minimum perceptible error is as follows: calculate the visibility threshold and the spatial masking effect of each pixel; take the minimum value of the visibility threshold and the spatial masking effect; the difference between the sum of the visibility threshold and the spatial masking effect and the product of the gain parameter and the minimum value is the minimum perceptible error of each pixel. Specifically:

[0069] The total spatial masking effect in JND only considers the contrast masking function, the contrast masking function A cm Contrast with brightness C l Therefore, the total spatial masking effect of each pixel x can be expressed as:

[0070]

[0071] Among them, α and β are two parameters, which are set to 16 and 26 respectively.

[0072] In addition, the human visual system (HVS) has different sensitivities to different background brightnesses, and the visibility thresholds under different brightness backgrounds are different. Therefore, brightness adaptation also needs to be considered. The visibility threshold L of the brightness adaptation of each pixel x is ada Modeled as:

[0073]

[0074] Wherein, B(x) is the background brightness of pixel x, which is the average brightness value of the surrounding area where x is located.

[0075] Finally, the JND threshold T for each pixel x is JND (x) can be calculated as:

[0076] T JND (x) = L ada (x)+A cm (x)-C·min{L ada (x), A cm (x)} (14)

[0077] Where C is a gain parameter, which is set to 0.3.

[0078] The discriminator uses the relative discriminator in ESRGAN, which is used to predict the probability that the real image xr is more real than the fake image xf, where the real image xr refers to a real high-resolution image, and the fake image xf refers to a high-resolution image generated by the generator (that is, preliminarily reconstructed). The principle of the relative discriminator is to predict the probability that the real image is real relative to the fake image. In the standard GAN, the discriminator is used to estimate the probability that the input sample is a real sample, and the generator is used to generate realistic samples to increase the probability of being identified as a real sample by the discriminator. The relative discriminator used in the present invention is used to estimate the probability that the real sample is more real than the fake sample. The core idea is that when the predicted probability value of the fake image xf increases (that is, the samples generated by the generator become more realistic), the predicted probability value of the real image xr should decrease.

[0079] The loss function L of the generator of the present invention is G The loss function in ESRGAN is adopted as follows:

[0080]

[0081] Among them, λ and η are coefficients that balance different loss terms. The first term L percepIt is a perceptual loss that extracts features before activation, that is, the VGG features before activation are used to replace the VGG features after activation in SRGAN, thereby obtaining a more effective perceptual loss; the third term L1 is used to calculate the restored image G(xi) and the ground truth value x r The 1-norm distance between the content loss:

[0082]

[0083] in, represents the operation of the loss function, G(x i ) is the restored image, x r is a real image.

[0084] The second term is the GAN adversarial loss:

[0085]

[0086] Among them, x f =G(x i ), x i is the input LR image, represents the operation of taking the average of all real data (i.e. xr) in the mini-batch processing, D Ra It represents the relative average discriminator (also called relative discriminator) used in the present invention, x r refers to the real high-resolution image, x f refers to the high-resolution image generated by the generator (that is, the preliminary reconstruction), It is an operation to take the average of all the false data (that is, xf) in the small batch processing.

[0087] The loss of the discriminator of the present invention adopts the loss in ESRGAN:

[0088]

[0089] In order to verify the reconstruction effect of the method of the present invention, four evaluation indicators, PSNR, SSIM, PI and RMSE, were used:

[0090] PSNR: Most of the current methods are PSNR-oriented. Peak signal-to-noise ratio (PSNR) is one of the most popular quality metrics for lossy transformation reconstruction (e.g., image compression, image restoration). For image super-resolution, PSNR is defined by the maximum pixel value (denoted as L) and the mean square error (MSE) between images. Given two images X and Y of the same size of m×n, the mean square error (MSE) is defined as:

[0091]

[0092] Then, PSNR (dB) is defined as:

[0093]

[0094] in, The maximum possible pixel value of the image.

[0095] In general, L represented by 8 bits is equal to 255. From the formula, it can be seen that PSNR is only related to the pixel-level MSE, so it only cares about the difference between corresponding pixels rather than visual perception, which often leads to poor reproduction quality in real scenes. However, PSNR is still the most widely used SR model evaluation standard.

[0096] SSIM: Considering that the human visual system (HVS) is highly adapted to extract image structure, the structural similarity index (SSIM) is proposed to measure the structural similarity between images based on independent comparisons of brightness and contrast. Given two images x and y, the structural similarity of the two images can be calculated as follows:

[0097]

[0098] Among them, μ x is the mean value of x, μ y is the mean value of y, is the variance of x, is the variance of y, σ xy is the covariance of x and y; c1=(k1L) 2 、c2=(k2L) 2 is a constant used to maintain stability; L is the dynamic range of pixel values; k1=0.01, k2=0.03.

[0099] Since the human visual system is complex and judging the difference between two images involves many aspects, namely the texture and flow within the image, methods that pursue absolute similarity difference (PSNR / SSIM) do not always perform well. Although distortion measurement has been widely used, the improvement of reconstruction accuracy is not always accompanied by an improvement in visual quality. In fact, in some cases, distortion and perceived quality are inconsistent. Image The perceptual quality of an image is defined as the degree to which it looks like a natural image, independent of its similarity to any reference image.

[0100] Natural Image Quality Evaluator (NIQE): NIQE is a completely blind image quality assessment method. NIQE does not require knowledge about training examples and the corresponding human opinion scores in the form of scores. NIQE only uses measurable deviations from statistical regularities observed in natural images. It extracts a set of local (quality-aware) features from the image based on the Natural Scene Statistics (NSS) model, and then fits the feature vector to a Multivariate Gaussian (MVG) model. The quality of the test image is then predicted by the distance between the MVG model learned from natural images and the MVG model:

[0101]

[0102] Among them, v1, v2 and Σ1, Σ2 are the mean vector and covariance matrix of the MVG model of HR and SR images. The higher the NQIE index, the lower the image perception quality.

[0103] The Perceptual Index (PI) is a combination of the no-reference image quality measures Ma and NIQE:

[0104]

[0105] Among them, M a It is a learning-based no-reference image quality assessment; the lower the PI, the better the perceived quality.

[0106] RMSE: Root mean square error, also called standard error, is the square root of the ratio of the sum of the squares of the deviations between the observed value and the true value to the number of observations m. It is used to measure the deviation between the observed value and the true value.

[0107]

[0108] Among them, m×n is the size of the two images X and Y.

[0109] Figure 3 This is the reconstruction effect diagram of the present invention. It can be seen that the image restoration effect is very good. The method of the present invention restores the details and texture of the image very well.

[0110] Table 1 is a comparison of the reconstruction effects of the present invention and ESRGAN. Although the method of the present invention pays more attention to the perception quality of the human eye and the quality of the image, at the same time, the present invention does not ignore the values ​​of PSNR and SSIM. It can be seen from Table 1 that the present invention has not only improved the PI index, but also improved the PSNR and SSIM indexes to varying degrees.

[0111] Table 1 Reconstruction effect comparison table

[0112]

[0113]

[0114] Embodiment 2

[0115] This embodiment provides an image super-resolution reconstruction system, which specifically includes the following modules:

[0116] An image acquisition module is configured to: acquire a low-resolution image;

[0117] An image reconstruction module is configured to: input the low-resolution image into the generator and the discriminator in sequence to obtain a reconstructed image;

[0118] Among them, the generator uses a convolutional layer to extract the shallow feature map of the low-resolution image. The shallow feature map is passed through the residual dense branch to obtain the first feature map, and the shallow feature is passed through the minimum perceptible error branch to obtain the second feature map. The result of multiplying the first feature map and the second feature map is input into the residual block with the second feature map. After the residual operation, the third feature map is obtained. After the third feature map is successively input into the upsampling block and the reconstruction block, a preliminary reconstructed image is obtained.

[0119] It should be noted here that each module in this embodiment corresponds to each step in Example 1 one by one, and the specific implementation process is the same, which will not be repeated here.

[0120] Embodiment 3

[0121] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the image super-resolution reconstruction method described in the first embodiment are implemented.

[0122] Embodiment 4

[0123] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the image super-resolution reconstruction method described in the first embodiment are implemented.

[0124] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.

[0125] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0126] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0127] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0128] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0129] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for super-resolution image reconstruction, characterized in that: include: Get low-resolution images; Input the low-resolution image into the generator and discriminator in sequence to obtain the reconstructed image; The generator uses a convolutional layer to extract a shallow feature map of a low-resolution image. The shallow feature map is passed through a residual dense branch to obtain a first feature map. The shallow feature map is passed through a minimum perceptible error branch to obtain a second feature map. The result of multiplying the first feature map and the second feature map is input into a residual block with the second feature map. After a residual operation, a third feature map is obtained. The third feature map is sequentially input into an upsampling block and a reconstruction block to obtain a preliminary reconstructed image. The residual dense branch is composed of a plurality of dense blocks, each of which is connected to a residual scaling block; the residual dense branch is composed of a plurality of convolutional layers connected in sequence, and the output of each convolutional layer is used as the input of all subsequent convolutional layers; Among them, the convolution layer is used to extract the shallow feature map F0 of the low-resolution image, F0 = H0 (I LR ), where I LR Represents the input low-resolution image, H0() represents the convolution function; the shallow feature map F0 is obtained through the residual dense branch to obtain the first feature map F1, F1 = H RRDB (F0), where H RRDB () represents the function of the residual dense branch; the shallow feature F0 is branched through the minimum perceptible error to obtain the second feature map F2, F2 = H JND (F0), where H JND () represents the function of the minimum perceptible error branch; the first feature map F1 is multiplied by the second feature map F2 to obtain Fx, Fx = F1 × F2; Fx is then input into the residual block together with the second feature map F2, and after the residual operation, the third feature map F3 is obtained, F3 = Fx + F2.

2. The image super-resolution reconstruction method according to claim 1, characterized in that: The minimum perceptible error branch is used to calculate the minimum perceptible error of each pixel in the feature map after the convolution of the shallow feature map to obtain a second feature map.

3. The image super-resolution reconstruction method according to claim 2, characterized in that: The minimum perceptible error is calculated as follows: Calculating the visibility threshold and spatial masking effect of each pixel; Take the minimum value of visibility threshold and spatial masking effect; The difference between the sum of the visibility threshold and the spatial masking effect and the product of the gain parameter and the minimum value is the minimum perceptible error for each pixel.

4. The image super-resolution reconstruction method according to claim 3, characterized in that: The visibility threshold is related to the background brightness of each pixel.

5. The image super-resolution reconstruction method according to claim 3, characterized in that: The spatial masking effect is related to the brightness contrast of each pixel.

6. An image super-resolution reconstruction system, characterized in that: include: An image acquisition module is configured to: acquire a low-resolution image; An image reconstruction module is configured to: input the low-resolution image into the generator and the discriminator in sequence to obtain a reconstructed image; The generator uses a convolutional layer to extract a shallow feature map of a low-resolution image. The shallow feature map is passed through a residual dense branch to obtain a first feature map. The shallow feature map is passed through a minimum perceptible error branch to obtain a second feature map. The result of multiplying the first feature map and the second feature map is input into a residual block with the second feature map. After a residual operation, a third feature map is obtained. The third feature map is sequentially input into an upsampling block and a reconstruction block to obtain a preliminary reconstructed image. The residual dense branch is composed of a plurality of dense blocks, each of which is connected to a residual scaling block; the residual dense branch is composed of a plurality of convolutional layers connected in sequence, and the output of each convolutional layer is used as the input of all subsequent convolutional layers; Among them, the convolution layer is used to extract the shallow feature map F0 of the low-resolution image, F0 = H0 (I LR ), where I LR Represents the input low-resolution image, H0() represents the convolution function; the shallow feature map F0 is obtained through the residual dense branch to obtain the first feature map F1, F1 = H RRDB (F0), where H RRDB () represents the function of the residual dense branch; the shallow feature F0 is branched through the minimum perceptible error to obtain the second feature map F2, F2 = H JND (F0), where H JND () represents the function of the minimum perceptible error branch; the first feature map F1 is multiplied by the second feature map F2 to obtain Fx, Fx = F1 × F2; Fx is then input into the residual block together with the second feature map F2, and after the residual operation, the third feature map F3 is obtained, F3 = Fx + F2.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the image super-resolution reconstruction method as described in any one of claims 1 to 5 are implemented.

8. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the image super-resolution reconstruction method according to any one of claims 1 to 5 are implemented.

Images