An image super-resolution reconstruction method and system based on data uncertainty

Abstract

The application discloses an image super-resolution reconstruction method and system based on data uncertainty, and the method comprises the following steps: acquiring a low-resolution image needing to be reconstructed; inputting the low-resolution image into an uncertainty optimization model to generate an uncertainty map and a corresponding first image super-resolution reconstruction result; calculating a loss weight function of the uncertainty optimization model according to the uncertainty map; completing the calculation of a mask uncertainty driving loss for optimizing a basic model; inputting the low-resolution image into the basic model to obtain a corresponding second image super-resolution reconstruction result; fusing the first and second image super-resolution reconstruction results, and obtaining a fusion weight according to the loss weight function, and finally obtaining a super-resolution image. The method can better utilize uncertainty to improve the expression effect of the super-resolution reconstruction task, and provide an uncertainty label for the super-resolution reconstruction result in an actual use real scene, so that the image quality can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of digital image processing, and more specifically to an image super-resolution reconstruction method and system based on data uncertainty. Background Technology

[0002] Super-resolution reconstruction (SMR) is a method of converting a low-resolution image into a high-resolution image to improve its visual quality. It is a key technology in image processing and provides crucial technical support for fields such as image detection and image segmentation. However, in real-world applications, the low-resolution and high-resolution images are often not in one-to-one correspondence, making it difficult to effectively apply the reconstructed SMR results to subsequent tasks and assess their reliability. Data uncertainty can effectively measure the reliability of a region; therefore, utilizing uncertainty maps to provide effective guidance for the SMR process can significantly promote its application in practical scenarios.

[0003] Data uncertainty can be used to describe the degree of fluctuation in image data. In practical applications, difficult samples with complex texture variations typically exhibit greater fluctuations during the learning process, resulting in higher uncertainty in the learned data. Conversely, simple samples, such as white walls, are easier to learn to achieve optimal solutions, thus exhibiting lower uncertainty in these areas. Difficult samples are generally more challenging to learn than simple samples, requiring more attention from the network to obtain better overall results. Therefore, utilizing uncertainty to provide the network with attention to these difficult samples is crucial.

[0004] Current super-resolution reconstruction techniques mostly rely on more convolutional operations to improve the overall model's expressive effect, but they neglect the reliability of the reconstruction results. This makes it impossible for most current techniques to be effectively applied in real-world scenarios such as remote sensing image processing and medical image processing, and the reconstruction results are not widely accepted.

[0005] Although some researchers have begun to study the uncertainty problem in the super-resolution reconstruction process, most of their studies only use the uncertain pixel intensity to observe the reliability of the results. The generated uncertainty is difficult to effectively guide the effective learning of deep learning-based super-resolution reconstruction networks. Furthermore, due to the increased cost of modeling the generated uncertainty, the learning effect of the network is poor. This has also prevented most previous uncertainty learning methods from being applied on a large scale to super-resolution reconstruction tasks.

[0006] While techniques such as uncertainty-weighted loss have been proposed to enhance focus on critical regions, low-uncertainty regions tend to learn more easily during the learning process, while high-uncertainty regions only show relatively better learning results later. Therefore, the learning levels of these two types of regions are often inconsistent during reconstruction, leading to a persistent learning discrepancy. This presents a significant challenge to the effectiveness of uncertainty-weighted loss, hindering deep learning networks from effectively learning these two types of uncertain regions and making it difficult to obtain satisfactory reconstruction results. Therefore, an effective method to utilize uncertainty remains urgently needed in single-frame super-resolution reconstruction tasks. Summary of the Invention

[0007] In view of this, the present invention proposes an image super-resolution reconstruction method and system based on data uncertainty, which is applied to the super-resolution reconstruction task of natural images. By effectively separating regions with high and low uncertainty, the problem of mutual contradiction between the two regions in the learning process is effectively solved.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] In a first aspect, embodiments of the present invention provide an image super-resolution reconstruction method based on data uncertainty, comprising the following steps:

[0010] S1. Obtain a low-resolution image that needs to be reconstructed;

[0011] S2. Input the low-resolution image into the uncertainty optimization model to generate an uncertainty map and the corresponding first image super-resolution reconstruction result; the uncertainty optimization model is optimized using uncertainty loss.

[0012] S3. Calculate the loss weight function of the uncertainty optimization model based on the uncertainty diagram.

[0013] S4. Calculate the mask uncertainty-driven loss based on the loss weight function, which is used to optimize the basic model;

[0014] S5. Input the low-resolution image into the optimized base model, and after processing by the convolution and residual modules, obtain the corresponding second image super-resolution reconstruction result.

[0015] S6. The first image super-resolution reconstruction result and the second image super-resolution reconstruction result are fused together, and the corresponding fusion weights are obtained according to the loss weight function to finally obtain the super-resolution image.

[0016] Further, step S2 specifically includes:

[0017] Construct a super-resolution reconstruction network, assuming the low-resolution image input to the network is represented as x, the high-resolution image as y, and the reconstructed super-resolution image as... And assume that y follows a Gaussian distribution. ,in The mean of the sample to be reconstructed. The variance of the sample to be reconstructed;

[0018] After m random samplings, the conditional likelihood function Expressed using the following formula:

[0019]

[0020] in, Let be the parameters of the given Gaussian prior network; i represents the index of the random sample. This represents the uncertainty in the model's prediction during the i-th sampling. Indicates the parameters used during the i-th sampling. Reconstruction results generated by the optimized network;

[0021] By maximizing the log-likelihood in the above formula, we obtain the loss function used to reduce data uncertainty, as shown below:

[0022]

[0023] In the formula, This represents the uncertainty loss used in the uncertainty optimization model, where N represents the number of sample points. This represents the generated uncertainty graph.

[0024] Furthermore, step S3 specifically includes:

[0025] The uncertainty plot is used to select non-negative linear functions as weighting coefficients for the loss function, as shown in the following form:

[0026]

[0027]

[0028] in, This represents the uncertainty graph, where s represents the result of the uncertainty transformation function. This represents a non-negative linear scaling function. This represents the value at each point in the result of the uncertainty transformation function;

[0029] Calculate the average value of the non-negative linear scaling function. The specific form is as follows:

[0030]

[0031] Where h and w represent the height and width of the uncertainty transformation function graph, respectively. The region with relatively low uncertainty level is selected as a filtering threshold, and a new set of weight coefficients is generated to adjust the original L1 loss function.

[0032] New weighting coefficients The specific form is as follows:

[0033]

[0034] in, This indicates the pixel position of the corrected weight map.

[0035] Furthermore, step S4 specifically includes:

[0036] The mask uncertainty-driven loss is calculated based on the loss weight function, and the uncertainty-driven loss mudl is as follows:

[0037]

[0038] In the formula, This indicates that uncertainty drives the loss.

[0039] Further, step S6 specifically includes:

[0040] The fusion result is obtained using the following calculation formula:

[0041]

[0042]

[0043] Where r represents the fusion weight of two different super-resolution reconstruction results, used to identify whether it is a high uncertainty region; if this pixel value is used, it is set to 1, otherwise it is set to 0; This represents the super-resolution reconstruction result of the first image; This represents the result of super-resolution reconstruction of the second image; This represents the final super-resolution image.

[0044] Secondly, embodiments of the present invention also provide an image super-resolution reconstruction system based on data uncertainty, comprising:

[0045] The acquisition module is used to acquire a low-resolution image that needs to be reconstructed;

[0046] The first generation module is used to input the low-resolution image into the uncertainty optimization model to generate an uncertainty map and a corresponding first image super-resolution reconstruction result; the uncertainty optimization model is optimized using uncertainty loss.

[0047] The loss weight function calculation module is used to calculate the loss weight function of the uncertainty optimization model based on the uncertainty graph.

[0048] The uncertainty-driven loss calculation module is used to calculate the mask uncertainty-driven loss according to the loss weight function, and is used to optimize the basic model.

[0049] The second generation module is used to input the low-resolution image into the optimized base model, and after processing by the convolution and residual modules, obtain the corresponding second image super-resolution reconstruction result.

[0050] The fusion module is used to fuse the super-resolution reconstruction results of the first image and the super-resolution reconstruction results of the second image, and obtain the corresponding fusion weights according to the loss weight function, so as to finally obtain the super-resolution image.

[0051] As can be seen from the above technical solution, compared with the prior art, this invention discloses an image super-resolution reconstruction method based on data uncertainty. This method can better utilize uncertainty to improve the expressive effect of super-resolution reconstruction tasks and provide uncertainty labels for super-resolution reconstruction results in real-world scenarios. Overall, a two-step network framework is designed. The first step uses a super-resolution reconstruction network to learn uncertainty, thereby effectively obtaining uncertain results and reconstructed high-resolution images. In the second step, the uncertainty results from the first step are used to help the reconstruction network learn about high-uncertainty regions. In the final fusion result, the results from the second step are used in high-uncertainty regions, while the results from the first step are used in low-uncertainty regions, thus comprehensively improving the reconstruction effects for both types of regions. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0053] Figure 1 A flowchart of the image super-resolution reconstruction method based on data uncertainty provided by the present invention;

[0054] Figure 2A schematic diagram of the image super-resolution reconstruction method based on data uncertainty provided by this invention;

[0055] Figure 3 Generate results based on the basic model and based on A comparison chart of the model generation results;

[0056] Figure 4 This is a comparison chart of the visualization results of the method of the present invention and other methods in natural scenes and comic works;

[0057] Figure 5 A block diagram of the image super-resolution reconstruction system based on data uncertainty provided by the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Reference Figure 1 As shown, this embodiment of the invention discloses an image super-resolution reconstruction method based on data uncertainty, including the following steps: S1. Obtain a low-resolution image that needs to be reconstructed;

[0060] S2. Input the low-resolution image into the uncertainty optimization model to generate an uncertainty map and the corresponding first image super-resolution reconstruction result. The uncertainty optimization model is optimized using uncertainty loss.

[0061] S3. Calculate the loss weight function of the uncertainty optimization model based on the uncertainty diagram.

[0062] S4. Calculate the mask uncertainty-driven loss according to the loss weight function to optimize the base model. The base model can use any other super-resolution reconstruction network, such as the EDSR model.

[0063] S5. Input the low-resolution image into the optimized base model, and after processing by the convolution and residual modules, obtain the corresponding second image super-resolution reconstruction result.

[0064] S6. The first image super-resolution reconstruction result and the second image super-resolution reconstruction result are fused together, and the corresponding fusion weights are obtained according to the loss weight function to finally obtain the super-resolution image.

[0065] This method can better utilize uncertainty to improve the expressive performance of super-resolution reconstruction tasks and provide uncertainty labels for super-resolution reconstruction results in real-world scenarios. It is based on a data-driven loss for mask uncertainty and applied to super-resolution reconstruction of natural images. By effectively separating regions of high and low uncertainty, it effectively addresses the contradiction between these two types of regions during the learning process.

[0066] Overall, a two-step network framework was designed. The first step utilizes a super-resolution reconstruction network to learn uncertainty, effectively obtaining both uncertain results and reconstructed high-resolution images. In the second step, the uncertainty results from the first step are used to help the reconstruction network learn about high-uncertainty regions. Finally, in the fusion result, the results from the second step are used in high-uncertainty regions, while the results from the first step are used in low-uncertainty regions, thus comprehensively improving the reconstruction performance for both types of regions.

[0067] The specific implementation principle is as follows: Figure 2 As shown, first input a low-resolution image, and then use the image... This means that in step 1, an uncertainty optimization model is used to generate uncertainty, and at the same time, the corresponding image super-resolution reconstruction results are generated. Uncertainty is represented by unc in the graph, and the second step of optimizing the model's loss weight function is completed using the unc graph. The calculation, in the second step, utilizes To achieve mask uncertainty-driven loss The calculations are used to optimize the base model and generate image super-resolution reconstruction results. Simultaneously, by combining these two results, a fusion is generated. The following sections will detail the method of this invention from three aspects: problem analysis, uncertainty generation, and uncertainty utilization optimization.

[0068] 1. Problem Analysis

[0069] In previous work, someone proposed a loss mechanism driven by uncertainty. To improve the model's performance, however, a comparison of experimental results revealed that, based on The inference results of the loss-trained model only outperform the base model in regions of high uncertainty; the results are similar to, for example... Figure 3 As shown, the HR image represents the high-resolution image, the SR image represents the super-resolution result reconstructed by the network, R represents the difference between the reconstructed result and the true high-resolution image, base represents the result obtained using L1 loss, and udl represents the result obtained using... The result obtained from the loss, where mudl represents the result obtained using the mask uncertainty-guided fusion loss proposed in this invention, is obtained using... The last column of the image represents the comparison relationship between pixel intensities; when this relationship is satisfied, the image pixel value is set to 1, and when the relationship is not satisfied, the image pixel value is set to 0. The results in the figure demonstrate that when using... During the learning process, the network considers the reconstruction effect in high uncertainty areas more, but ignores low uncertainty areas, resulting in poor reconstruction effect in low uncertainty areas.

[0070] To enhance the reconstruction of different region types, this invention proposes a novel training strategy based on mask uncertainty. Analysis of the network uncertainty graph shows that low-uncertainty regions are mainly concentrated in smoother areas, while high-uncertainty regions are mainly located in textured areas, with a significant difference in magnitude. Based on this observation, this invention uses the average value of the uncertainty graph as a filtering threshold to ignore regions with low uncertainty levels, thereby minimizing their impact on the modeling process of high-uncertainty regions.

[0071] 2. Generation of Uncertainty

[0072] In the first step, this invention uses a general uncertainty optimization loss to optimize the super-resolution reconstruction network. Assume the low-resolution image input to the network is represented as x, and the high-resolution image as y, where high and low resolution depend on the image size; smaller images are low-resolution images, and larger images are high-resolution images. The reconstructed super-resolution image is... And assume that y follows a Gaussian distribution. ,in The mean of the sample to be reconstructed.

[0073] Let be the variance of the sample to be reconstructed. After m random samplings, the conditional likelihood function... It can be expressed by the following formula:

[0074]

[0075] in, Let be the parameters of the given Gaussian prior network; i represents the index of the random sample. This represents the uncertainty in the model's prediction during the i-th sampling. Indicates the parameters used during the i-th sampling. The optimized network generates the reconstruction results. By maximizing the log-likelihood in the above formula, a general loss function for reducing data uncertainty can be obtained, as shown below:

[0076]

[0077] In the formula, This represents the uncertainty loss used in the uncertainty optimization model, where N represents the number of sample points. This represents the generated uncertainty graph.

[0078] With the help of this loss function, the uncertainty estimate of the generated super-resolution reconstruction result can be effectively provided. During the learning process, when the prediction error of a sample is large, the prediction difficulty of that sample is usually high, and there is a large amount of inherent noise. Therefore, the uncertainty value of the prediction for that sample will also increase with the increase of the prediction error. Conversely, when the prediction error of a sample is small, the prediction difficulty of that sample is usually low, the sample points often have a large overlap with the surrounding area, the noise is low, and the prediction uncertainty is low. It will also decrease as the prediction error decreases.

[0079] 3. Optimization using uncertainty

[0080] After obtaining the uncertainty map corresponding to the reconstructed image, this invention proposes to utilize this uncertainty map to help the reconstruction network achieve better optimization. This invention uses the uncertainty map to provide weights for the L1 loss, making regions with high uncertainty have a higher weight in the model optimization process. To better utilize uncertainty, this invention chooses a non-negative linear function as the weight coefficients of the loss function, the specific expression of which is shown below:

[0081]

[0082]

[0083] in, This represents the uncertainty graph, where s represents the result of the uncertainty transformation function. It is a non-negative linear scaling function. This is the value at each point in the result of the uncertainty transformation function. Then, the average value of the scaling function can be calculated. The specific form is as follows:

[0084]

[0085] Where h and w represent the height and width of the uncertainty transformation function graph, respectively. With the help of Using a filtering threshold, regions with relatively low uncertainty levels can be effectively selected, and a new set of weighting coefficients can be generated to adjust the original L1 loss function. The main purpose of this method is to mitigate the bias that may arise from the influence of low uncertainty regions when learning high uncertainty regions. New weighting coefficients The specific form is as follows:

[0086]

[0087] in, This represents the pixel position in the corrected weight map. Therefore, the corrected weight map can effectively reduce the influence of low-uncertainty regions. The final uncertainty-driven loss (mudl) used is shown below:

[0088]

[0089] exist With loss optimization, the base model exhibits better optimization performance in high-uncertainty regions. In low-uncertainty regions, this invention proposes using the results generated in the first step to jointly complete the final super-resolution reconstruction output. The specific calculation formula is shown below:

[0090]

[0091]

[0092] Where r is the fusion weight of two different super-resolution reconstruction results, used to identify whether it is a high uncertainty region. If this pixel value is used, it is set to 1, otherwise it is set to 0. This represents the super-resolution reconstruction result of the first image; This represents the result of super-resolution reconstruction of the second image; This is the final super-resolution reconstruction result.

[0093] The image super-resolution reconstruction method based on data uncertainty provided in this invention applies a data-driven loss based on mask uncertainty to natural images for super-resolution reconstruction tasks. In the experimental section, this invention was trained using the DIV2k dataset and tested on the Set5, Set14, BSD100, Urban100, and Manga109 datasets to observe the optimization effect of the uncertainty-based loss.

[0094] The evaluation metric is PSNR; a higher PSNR indicates smaller image differences.

[0095] Table 1. Comparison of the method of the present invention with other methods

[0096]

[0097] As can be seen from the table, the method of this invention achieves better metrics and image reconstruction quality compared to other methods. This indicates that low-uncertainty regions are learned and optimized early in the training phase. However, this optimization may hinder the learning ability of other regions, and the network tends to prioritize ensuring that this type of region is continuously optimized better. In contrast, the loss function proposed in this invention ignores low-uncertainty regions in the model in step 2, focusing instead on optimizing the reconstruction of high-uncertainty regions. Therefore, the method proposed in this invention is more effective in terms of reconstruction results. Figure 4 This is a comparison chart of the visualization results of the method of this invention and other methods in natural scenes and comic works. In the chart, a represents a high-resolution image, b represents the reconstruction result of the base model, c represents the output result of the uncertainty optimization model, and d represents the result of using... The optimized model result, e indicates the use of The optimized model result, where f represents the uncertainty plot. Figure 4 The middle three rows show the super-resolution reconstruction results at 2x, 3x, and 4x magnification, respectively. As the results in the figure show, compared to other methods, the method of this invention effectively improves image quality at different magnifications and exhibits better performance in processing texture edge regions, further demonstrating the effectiveness of this invention.

[0098] The method of this invention can be used to improve low-resolution medical images, such as MRI scans or ultrasound images, and helps improve diagnostic accuracy. Obtaining high-resolution images helps doctors understand more detailed information about a patient's problem, which helps to provide better treatment.

[0099] Furthermore, in the field of remote sensing imagery, satellite images are always low in resolution due to technological limitations. Therefore, the method of this invention can be used to enlarge these low-resolution images to facilitate better analysis and understanding of geographical features, weather patterns, etc.

[0100] Based on the same inventive concept, this invention also provides an image super-resolution reconstruction system based on data uncertainty. Since the principle of the problem solved by this system is similar to the aforementioned image super-resolution reconstruction method based on data uncertainty, the implementation of this system can refer to the implementation of the aforementioned method, and the repeated parts will not be described again.

[0101] Reference Figure 5 As shown, this embodiment of the invention also provides an image super-resolution reconstruction system based on data uncertainty, comprising:

[0102] The acquisition module is used to acquire a low-resolution image that needs to be reconstructed;

[0103] The first generation module is used to input the low-resolution image into the uncertainty optimization model to generate an uncertainty map and a corresponding first image super-resolution reconstruction result; the uncertainty optimization model is optimized using uncertainty loss.

[0104] The loss weight function calculation module is used to calculate the loss weight function of the uncertainty optimization model based on the uncertainty graph.

[0105] The uncertainty-driven loss calculation module is used to calculate the mask uncertainty-driven loss according to the loss weight function, and is used to optimize the basic model.

[0106] The second generation module is used to input the low-resolution image into the optimized base model, and after processing by the convolution and residual modules, obtain the corresponding second image super-resolution reconstruction result.

[0107] The fusion module is used to fuse the super-resolution reconstruction results of the first image and the super-resolution reconstruction results of the second image, and obtain the corresponding fusion weights according to the loss weight function, so as to finally obtain the super-resolution image.

[0108] This image super-resolution reconstruction system based on data uncertainty can be applied to super-resolution reconstruction tasks of natural images. By effectively separating regions of high and low uncertainty, it effectively solves the problem of contradictions between the two types of regions during the learning process. It achieves effective application in practical scenarios such as remote sensing image processing and medical image processing, and produces high-quality reconstruction results.

[0109] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0110] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An image super-resolution reconstruction method based on data uncertainty, characterized in that, Includes the following steps: S1. Obtain a low-resolution image that needs to be reconstructed; S2. Input the low-resolution image into the uncertainty optimization model to generate an uncertainty map and the corresponding first image super-resolution reconstruction result; the uncertainty optimization model is optimized using uncertainty loss. S3. Calculate the loss weight function of the uncertainty optimization model based on the uncertainty diagram. S4. Calculate the mask uncertainty-driven loss based on the loss weight function, which is used to optimize the basic model; S5. Input the low-resolution image into the optimized base model, and after processing by the convolution and residual modules, obtain the corresponding second image super-resolution reconstruction result. S6. The first image super-resolution reconstruction result and the second image super-resolution reconstruction result are fused together, and the corresponding fusion weight is obtained according to the loss weight function to finally obtain the super-resolution image. Specifically, step S2 includes: Construct a super-resolution reconstruction network, assuming the low-resolution image input to the network is represented as x, the high-resolution image as y, and the reconstructed super-resolution image as... And assume that y follows a Gaussian distribution. ,in The mean of the sample to be reconstructed. The variance of the sample to be reconstructed; After m random samplings, the conditional likelihood function Expressed using the following formula: in, Let be the parameters of the given Gaussian prior network; i represents the index of the random sample. This represents the uncertainty in the model's prediction during the i-th sampling. Indicates the parameters used during the i-th sampling. Reconstruction results generated by the optimized network; By maximizing the log-likelihood in the above formula, we obtain the loss function used to reduce data uncertainty, as shown below: In the formula, This represents the uncertainty loss used in the uncertainty optimization model, where N represents the number of sample points. This represents the uncertainty graph generated by the model during the i-th sampling. Step S3 specifically includes: The uncertainty plot is used to select non-negative linear functions as weighting coefficients for the loss function, as shown in the following form: in, This represents the uncertainty graph, where s represents the result of the uncertainty transformation function. This represents a non-negative linear scaling function. This represents the value at each point in the result of the uncertainty transformation function; Calculate the average value of the non-negative linear scaling function. The specific form is as follows: Where h and w represent the height and width of the uncertainty transformation function graph, respectively. The region with relatively low uncertainty level is selected as a filtering threshold, and a new set of weight coefficients is generated to adjust the original L1 loss function. New weighting coefficients The specific form is as follows: in, This indicates the pixel position of the corrected weight map.

2. The image super-resolution reconstruction method based on data uncertainty according to claim 1, characterized in that, Step S4 specifically includes: The mask uncertainty-driven loss is calculated based on the loss weight function, and the uncertainty-driven loss mudl is as follows: In the formula, This indicates that uncertainty drives the loss.

3. The image super-resolution reconstruction method based on data uncertainty according to claim 2, characterized in that, Step S6 specifically includes: The fusion result is obtained using the following calculation formula: Where r represents the fusion weight of two different super-resolution reconstruction results, used to identify whether it is a high uncertainty region; if it is a high uncertainty region, it is set to 1 and the second image super-resolution reconstruction result is used; if it is a low uncertainty region, it is set to 0 and the first image super-resolution reconstruction result is used. This represents the super-resolution reconstruction result of the first image; This represents the result of super-resolution reconstruction of the second image; This represents the final super-resolution image.

4. An image super-resolution reconstruction system based on data uncertainty, characterized in that, The system, employing the image super-resolution reconstruction method based on data uncertainty as described in any one of claims 1-3, comprises: The acquisition module is used to acquire a low-resolution image that needs to be reconstructed; The first generation module is used to input the low-resolution image into the uncertainty optimization model to generate an uncertainty map and a corresponding first image super-resolution reconstruction result; the uncertainty optimization model is optimized using uncertainty loss. The loss weight function calculation module is used to calculate the loss weight function of the uncertainty optimization model based on the uncertainty graph. The uncertainty-driven loss calculation module is used to calculate the mask uncertainty-driven loss according to the loss weight function, and is used to optimize the basic model. The second generation module is used to input the low-resolution image into the optimized base model, and after processing by the convolution and residual modules, obtain the corresponding second image super-resolution reconstruction result. The fusion module is used to fuse the super-resolution reconstruction results of the first image and the super-resolution reconstruction results of the second image, and obtain the corresponding fusion weights according to the loss weight function, so as to finally obtain the super-resolution image.

Images