Pulse noise elimination method and system based on adaptive mean filtering

Abstract

The present invention discloses an impulse noise elimination method and system based on adaptive mean filtering, which belongs to the technical field of signal denoising, and specifically comprises: firstly, by collecting noise-labeled image data sets, and adopting local contrast normalization preprocessing, combining a CNN model with traditional low-level features, a multi-scale feature fusion map is formed; then, the segmentation capability of CNN is used to predict the noise area, and the boundary is dynamically adjusted; secondly, based on the neighborhood pixel characteristics of the noise area, a dynamic weight function and a filtering window are set, and a weighted mean calculation is performed on each pixel to identify and replace noise points; finally, the filtering effect is optimized through deep residual learning, the image quality is enhanced, and evaluation and visualization are performed, so that the present invention effectively improves the accuracy and efficiency of image denoising.

Description

Technical Field

[0001] The invention belongs to the technical field of signal denoising, and in particular to an impulse noise elimination method and system based on adaptive mean filtering. Background Art

[0002] In building risk detectors, impulse noise elimination methods refer to technical means for identifying, analyzing and eliminating sudden, short-term and high-intensity noise interference generated by building environment, equipment operation or environmental factors. This noise interference is usually characterized by randomness, discontinuity and high energy density, which seriously affects the accuracy and reliability of building risk detectors. In the early days of building risk detection technology, the detectors at that time mainly focused on static or continuous risk factors, and simple shielding or filtering methods were often used to deal with impulse noise. With the development of signal processing technology, researchers began to pay attention to the impact of impulse noise on detector performance, and proposed a series of methods for eliminating impulse noise, including pulse detection based on statistical characteristics and noise separation based on time-frequency analysis. Among them, pulse detection based on statistical characteristics has large computational complexity, is sensitive to noise, and has limited generalization ability, while noise separation based on time-frequency analysis has high complexity, high requirements for signal stability, and poor noise resistance.

[0003] For example, the patent with the authorization announcement number CN105590301B discloses an impulse noise elimination method of adaptive positive oblique double cross window mean filtering, including: (1) using the sub-block sorting difference maximum method and voting method to obtain the upper and lower boundaries of the impulse noise, and using the upper and lower boundaries to detect the impulse noise points; (2) firstly using a 3×3 vertical and horizontal cross (positive cross) window to recursively trim the mean filter 3 times for the noise image to be processed, and then using a diagonal cross (oblique cross) window to recursively trim the mean filter 3 times, and using the trimmed mean filter result to replace the value of the noise point, if the noise point has been processed, the mean filter is terminated, otherwise the window is increased to continue similar double cross recursive trimming mean filtering until the 7×7 window is completed; (3) if the noise is still not processed, repeat step (2) to form iterative filtering. This technical solution has the advantages of accurate detection of impulse noise points, good impulse noise elimination effect and fast denoising speed.

[0004] The above existing technologies all have the following problems: 1) lack of deep learning and optimization capabilities, lack of adaptive learning capabilities; 2) low computational efficiency of iterative filtering in noisy or complex images. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes an impulse noise elimination method based on adaptive mean filtering, which collects noise-labeled image data sets, adopts local contrast normalization preprocessing, combines the CNN model with traditional low-level features, and forms a multi-scale feature fusion map; uses the segmentation capability of CNN to predict the noise area and dynamically adjust the boundary; based on the neighborhood pixel characteristics of the noise area, sets a dynamic weight function and a filtering window, and performs weighted mean calculation on each pixel to identify and replace noise points; optimizes the filtering effect through deep residual learning, enhances the image quality, and performs evaluation and visualization, so that the present invention effectively improves the accuracy and efficiency of image denoising.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The impulse noise elimination method based on adaptive mean filtering includes:

[0008] Step S1: collect the image data set after noise labeling, perform preprocessing, extract features from the preprocessed image data set, and combine traditional low-level features to form a multi-scale feature fusion map;

[0009] Step S2: Performing in-depth analysis on the multi-scale feature fusion image, predicting the noise area, and dynamically adjusting the upper and lower boundaries of the noise area using a dynamic boundary adaptive adjustment method;

[0010] Step S3: according to the determined noise area, a dynamic weight function and a dynamic noise adjustment threshold are set, and according to the statistical characteristics of the pixels in the neighborhood of the noise area, a filtering window is set for each noise point, and each pixel in the noise image is traversed using the filtering window. In each filtering window, different weights are assigned to each pixel using a dynamic weight function according to the distance between the pixel and the noise center point and the gray value difference, and a weighted mean is calculated;

[0011] Step S4: According to the weighted mean, determine whether the current pixel is a noise point. If it is a noise point, replace the current pixel with the mean of the pixels in the filter window, and determine again whether there is a noise point. If so, increase the filter window and calculate the noise change rate. If the noise change rate is less than the noise adjustment threshold, stop traversing the noise image and save the process data.

[0012] Specifically, the specific steps of step S1 include:

[0013] S1.1: Collect a dataset of images labeled with noise And the data of X is enhanced by rotation and translation, where represents image data labeled with building noise, represents image data labeled with traffic noise, represents the image data marked with construction noise, n represents the number of building noise image data, m represents the number of traffic noise image data, and a represents the number of construction noise image data;

[0014] S1.2: Adjust the size of the image dataset X and preprocess it, use the CNN model API to instantiate the model, and traverse the preprocessed image dataset. For each image in the dataset, use the convolutional basis of the CNN model to perform forward propagation to obtain feature maps, and perform feature extraction on the feature maps of different layers;

[0015] S1.3: Perform global average pooling on the extracted features and concatenate and fuse them with color, shape, and texture features to obtain fused features. Based on the fused features, a multi-scale feature fusion graph is formed through a multi-scale feature fusion algorithm.

[0016] Specifically, the specific steps of step S2 include:

[0017] S2.1: The obtained multi-scale feature fusion map is input into the trained CNN model. In the fully connected layer and output layer of the CNN model, the CNN model generates a probability map of the same size as the input image based on the fusion features;

[0018] S2.2: Set threshold h1, if p i >h1, then the pixel on the probability map belongs to the noise area, and a binary noise area prediction map is obtained, where p i Represents the i-th pixel on the probability map, and i represents the index position of the pixel on the probability map.

[0019] Specifically, the specific steps of step S2 also include:

[0020] S2.3: Based on the noise region prediction map, set the initial noise region boundary [b, c], analyze the initial boundary neighborhood pixel features, and evaluate whether the corresponding pixel belongs to the noise region based on the pixel features;

[0021] If A bian ~B noise , then the boundary neighborhood pixels are added to the noise area, where A bian represents the boundary neighborhood pixel features, B noise represents the pixel features inside the noise region, and ~ represents similarity;

[0022] If A bian ~D gro , then the boundary neighborhood pixels are excluded from the noise area, where D gro Represents the pixel features of the background area;

[0023] S2.4: Set the boundary threshold of the noise area to [d, e], and repeat S2.2-S2.3 until the boundary of the noise area changes. satisfy:

[0024] Specifically, the specific formula of the dynamic weight function in step S3 is:

[0025]

[0026] Among them, w(j, k) represents the dynamic weight, β, represents the balance factor, represents the distance between the pixel (j, k) in the noise image and the noise center point, g(j, k) represents the gray value of the pixel (j, k) in the noise image, and g center represents the gray value of the noise center point, var(j, k) represents the local variance in the neighborhood of pixel (j, k) in the noise image, σ g , σ var Represent the standard deviation of distance, gray value difference and local variance respectively.

[0027] Specifically, the formula for the weighted mean in step S3 is:

[0028]

[0029] Among them, E(y, z) represents the gray value of the image at the position (y, z) after filtering, W represents the current filtering window, and F(y+q y (j, k), z+q z (j, k)) represents the noise image at position (y+q y (j, k), z+q z The gray value on (j, k)), q y (j, k), q z (j, k) represents the offset function.

[0030] The impulse noise elimination system based on adaptive mean filtering includes: a data processing module, a feature fusion module, a mean filtering module, and an effect optimization module;

[0031] The data processing module is used to pre-process the noise-marked image data set;

[0032] The feature fusion module is used to integrate the features extracted by CNN and traditional low-level features to form a multi-scale feature fusion graph;

[0033] The mean filter module is used to predict the noise area and eliminate the noise using an adaptive mean filter method according to the determined noise area;

[0034] The effect optimization module is used to optimize the filtering effect and enhance and evaluate the optimized filtered image.

[0035] Specifically, the mean filter module includes: a region prediction unit, a boundary adjustment unit, and an adaptive filtering unit;

[0036] The region prediction unit is used to predict the noise region using a deep learning method;

[0037] The boundary adjustment unit is used to dynamically adjust the boundary of the predicted noise area;

[0038] The adaptive filtering unit is used to dynamically adjust the weights to perform weighted mean calculation according to the relative positions and gray value difference information between pixels within a set filtering window.

[0039] Specifically, the effect optimization module includes: a residual learning unit and an image enhancement unit;

[0040] The residual learning unit is used to utilize the deep residual learning concept to perform iterative optimization by comparing the residual of the original image and the processed image, and continuously fine-tune the filter parameters until a denoising effect is achieved;

[0041] The image enhancement unit is used to perform contrast enhancement on the optimized filtered image using an enhancement algorithm.

[0042] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of an impulse noise elimination method based on adaptive mean filtering when executing the computer program.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. The present invention proposes an impulse noise elimination method based on adaptive mean filtering. Through the segmentation capability of CNN and combined with a multi-scale feature fusion map, the noise area boundary is dynamically adjusted, and the filtering window is adaptively set, thereby avoiding the problem of low filtering calculation efficiency caused by too many unnecessary factors due to a large filtering window. At the same time, the setting of the adaptive window can more effectively remove noise while retaining the detailed information of the image, and has good flexibility and scalability.

[0045] 2. The present invention proposes an impulse noise elimination method based on adaptive mean filtering. By constructing a residual model, the filtering effect is further optimized, the information loss in the filtering process is reduced, and the filtered image is closer to the original noise-free image. At the same time, after the filtering is completed, the enhanced algorithm is used to enhance the optimized filtered image, thereby improving the overall quality of the image. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Schematic diagram of the impulse noise elimination method based on adaptive mean filtering of the present invention;

[0047] Figure 2 The following is a flowchart of the pulse noise elimination method based on adaptive mean filtering of the present invention;

[0048] Figure 3 It is a flow chart of noise area prediction of the impulse noise elimination method based on adaptive mean filtering of the present invention;

[0049] Figure 4 This is a diagram of the architecture of the impulse noise elimination system based on adaptive mean filtering of the present invention. DETAILED DESCRIPTION

[0050] Example 1

[0051] See also Figure 1-Figure 3 The present invention provides an embodiment: an impulse noise elimination method based on adaptive mean filtering, which is applied in the field of signal denoising technology, and is also applicable to the fields of signal acquisition, signal processing, and image processing. The present application includes the following steps:

[0052] Step S1: collect the image data set after noise labeling, and use the local contrast normalization method to preprocess it, input the preprocessed image data set into the trained CNN model for feature extraction, and combine the traditional low-level features to form a multi-scale feature fusion map, wherein the training process of the CNN model is the prior art content in this field, and is not the inventive solution of the present application, and will not be described in detail here;

[0053] Step S2: Analyze the multi-scale feature fusion map using the segmentation capability of the convolutional neural network, predict the noise area, and dynamically adjust the upper and lower boundaries of the noise area using a dynamic boundary adaptive adjustment method;

[0054] Step S3: according to the determined noise area, a dynamic weight function and a dynamic noise adjustment threshold are set, and according to the statistical characteristics of the pixels in the neighborhood of the noise area, a filtering window is set for each noise point, and each pixel in the noise image is traversed using the filtering window. In each filtering window, different weights are assigned to each pixel using a dynamic weight function according to the distance between the pixel and the noise center point and the gray value difference, and a weighted mean is calculated;

[0055] Furthermore, the specific steps of step S3 include:

[0056] S3.1: Design a dynamic weight function based on the mean and standard deviation of the determined noise area, where the dynamic weight function should take into account the distance between the pixel and the noise center point and the gray value difference;

[0057] S3.2: Setting a dynamic noise adjustment threshold to distinguish between noise points and non-noise points. For each detected noise point, a suitable filter window size is set according to the density and size of the noise points of its neighboring pixels. The filter window size selected by the present invention is [5, 5];

[0058] S3.3: Use the set [5, 5] filter window to traverse each pixel in the noise image. In each filter window, use the dynamic weight function to assign a weight to each pixel based on the distance and grayscale value difference between the pixel and the noise center point, and use the calculated weighted mean as the new grayscale value of the pixel after filtering.

[0059] Step S4: judging whether the current pixel is a noise point according to the weighted mean value, if it is a noise point, replacing the current pixel with the mean value of the pixels in the filter window, and judging whether there is a noise point again, if there is, increasing the filter window, calculating the noise change rate, and stopping traversing the noise image if the noise change rate is less than the noise adjustment threshold;

[0060] Furthermore, the formula for calculating the noise change rate is:

[0061]

[0062] Among them, R r represents the noise change rate, r represents the number of iterations, L r Indicates the current noise level, L r-1 represents the noise level of the previous iteration, ε represents a small positive number, and |·| represents the absolute value. Since the noise level is increasing or decreasing, when the noise change rate R r When the noise level is less than a preset noise adjustment threshold h1, the traversal of the noise image is stopped. The threshold h1 is set according to the specific application and noise characteristics.

[0063] Step S5: Utilize the deep residual learning idea to optimize the filtering effect by constructing a residual model between the original image and the filtered image, use an enhancement algorithm to enhance the optimized filtered image, and use image quality assessment indicators to evaluate the optimized filtered image, output and visualize the evaluation results, and at the same time, record and save the parameters and results in the processing process. The deep residual learning idea and enhancement algorithm are prior art contents in this field, and are not the creative solutions of this application, so they will not be elaborated here.

[0064] The specific steps of step S1 include:

[0065] S1.1: Collect a dataset of images labeled with noise And the data of X is enhanced by rotation and translation, where represents image data labeled with building noise, represents image data labeled with traffic noise, represents the image data marked with construction noise, n represents the number of building noise image data, m represents the number of traffic noise image data, and a represents the number of construction noise image data;

[0066] S1.2: Adjust the size of the image dataset X and preprocess it, use the CNN model API to instantiate the model, and traverse the preprocessed image dataset. For each image in the dataset, use the convolutional basis of the CNN model to perform forward propagation to obtain feature maps, and perform feature extraction on the feature maps of different layers;

[0067] S1.3: Perform global average pooling on the extracted features and concatenate and fuse them with color, shape, and texture features to obtain fused features. Based on the fused features, a multi-scale feature fusion graph is formed through a multi-scale feature fusion algorithm.

[0068] The specific steps of step S2 include:

[0069] S2.1: The obtained multi-scale feature fusion map is input into the trained CNN model. In the fully connected layer and output layer of the CNN model, the CNN model generates a probability map of the same size as the input image based on the fusion features, where each pixel value on the probability map represents the probability that the pixel belongs to the noise area;

[0070] S2.2: Set threshold h1, if p i > h1, then the pixel on the probability map belongs to the noise area, and a binary noise area prediction map is obtained, where p i Represents the i-th pixel on the probability map, and i represents the index position of the pixel on the probability map.

[0071] S2.3: Based on the noise region prediction map, set the initial noise region boundary [b, c], analyze the initial boundary neighborhood pixel features, and evaluate whether the corresponding pixel belongs to the noise region based on the pixel features;

[0072] If A bin ~B noise , then the boundary neighborhood pixels are added to the noise area, where A bian represents the boundary neighborhood pixel features, B noise represents the pixel features inside the noise region, and ~ represents similarity;

[0073] If A bian ~D gro , then the boundary neighborhood pixels are excluded from the noise area, where D gro Represents the pixel features of the background area;

[0074] S2.4: Set the boundary threshold of the noise area to [d, e], and repeat S2.2-S2.3 until the boundary of the noise area changes. satisfy:

[0075] The specific formula of the dynamic weight function in step S3 is:

[0076]

[0077] Among them, w(j, k) represents the dynamic weight of pixel (j, k) in the noise image, β, represents the balance factor, which is used to adjust the contribution of different factors in the weight calculation, and β∈[0,1], represents the distance between the pixel (j, k) in the noise image and the center point of the noise. The closer the distance, the greater the weight, and the farther the distance, the smaller the weight. g(j, k) represents the gray value of the pixel (j, k) in the noise image. The smaller the gray value difference, the greater the weight, and the greater the gray value difference, the smaller the weight. center represents the gray value of the noise center point, var(j, k) represents the local variance in the neighborhood of pixel (j, k) in the noise image, σ g , σ var They represent the standard deviation of distance, gray value difference and local variance respectively, which are used to measure the intensity of noise and control the decay speed of the weight function.

[0078] It should be noted that, in the present invention, based on the original weight function, a distance variable between the pixel and the noise center point is added. Considering the spatial relationship between the pixel and the center point of the noise helps to adaptively adjust the filter window, maintain the local continuity of the image and reduce the information loss in the filtering process; secondly, the weight distribution will also be affected by the fluctuation of the gray value, so the local variance of the pixel var(j, k) is considered. Finally, according to The distribution and characteristics of var(j, k) are controlled by the balance factor β, Assign and adjust corresponding weight values ​​to achieve better filtering effect.

[0079] The formula for the weighted mean in step S3 is:

[0080]

[0081] Where E(y, z) represents the weighted mean of the filtered image at position (y, z), which is used as the filtered pixel value to replace the noise pixels in the original image, W represents the current filter window, and W = [5, 5], F(y+q y (j, k), z+q z(j, k)) represents the noise image at position (y+q y (j, k), z+q z The gray value on (j, k)), q y (j, k), q z (j, k) represents the offset function.

[0082] It should be noted that when filtering using a filter window, different processing is required due to different reasons such as sensor noise, illumination change, and image degradation. Therefore, an offset is added to each pixel in the present invention to compensate or adjust for these specificities, so that the weighted mean algorithm can be more flexible to adapt to different image contents and processing requirements. In the present invention, by adjusting the offset, local changes in the image can be better adapted, thereby removing noise or performing other types of image processing while maintaining image details, and can also help correct pixel value deviations caused by sensor defects, uneven illumination, or image degradation factors, thereby improving the accuracy of image processing.

[0083] Example 2

[0084] See also Figure 4 , another embodiment provided by the present invention: an impulse noise elimination system based on adaptive mean filtering, comprising: a data processing module, a feature fusion module, a mean filtering module, and an effect optimization module;

[0085] A data processing module, used for preprocessing the image data set after noise marking;

[0086] Feature fusion module, used to integrate the features extracted by CNN and traditional low-level features to form a multi-scale feature fusion graph;

[0087] The mean filter module is used to predict the noise area and eliminate the noise using the adaptive mean filter method according to the determined noise area;

[0088] The effect optimization module is used to optimize the filtering effect and enhance and evaluate the optimized filtered image.

[0089] The data processing module includes: a data acquisition unit, a noise marking unit, and a normalization unit;

[0090] A data acquisition unit, used for collecting or generating a noise-labeled image dataset;

[0091] A noise marking unit, used to identify and mark noise areas in the image data;

[0092] The normalization unit is used to use the local contrast normalization method to standardize the brightness and contrast of the image, enhance the local contrast of the image, and reduce the impact of uneven illumination factors on subsequent processing.

[0093] The feature fusion module includes: a feature extraction unit and a feature fusion unit;

[0094] Feature extraction unit, used to extract different levels of features from the input image, including features extracted by the CNN model and traditional low-level features such as edges and textures;

[0095] The feature fusion unit is used to fuse the features extracted by CNN with traditional low-level features to form a multi-scale feature fusion graph.

[0096] The mean filter module includes: a region prediction unit, a boundary adjustment unit, a window setting unit, and an adaptive filtering unit;

[0097] A region prediction unit, which uses deep learning methods to predict noisy regions to more accurately locate pixels that need to be filtered and avoid misprocessing important details of the image;

[0098] A boundary adjustment unit is used to dynamically adjust the boundary of the predicted noise area to ensure that the filtering range is neither too large to affect the image quality nor too small to cause the noise to not be completely eliminated;

[0099] A window setting unit is used to set a filtering window for each noise point according to the predicted noise area characteristics. The window size varies with the local image content to adapt to structures of different scales.

[0100] The adaptive filtering unit is used to dynamically adjust the weights to perform weighted mean calculation within the set filtering window according to the relative positions and gray value difference information between pixels, so as to effectively eliminate impulse noise while protecting image details.

[0101] The effect optimization module includes: residual learning unit, image enhancement unit, quality assessment unit, and result recording unit;

[0102] The residual learning unit is used to use the deep residual learning idea to perform iterative optimization by comparing the residuals of the original image and the processed image, and continuously fine-tune the filter parameters until the denoising effect is achieved;

[0103] An image enhancement unit, used for performing contrast enhancement on the optimized filtered image using an enhancement algorithm to restore and improve the overall visual quality of the image;

[0104] The quality assessment unit is used to quantify the denoising effect using objective evaluation indicators to ensure that the processed image is both noise-free and retains detail information;

[0105] The result recording unit is used to record and save the parameters, intermediate results and final denoised image during the processing.

[0106] Example 3

[0107] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of an impulse noise elimination method based on adaptive mean filtering when executing the computer program.

[0108] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in the field can also change, modify, replace and modify the above-mentioned embodiments without departing from the purpose and scope of protection of the present invention, which are all within the protection of the present invention.

Claims

1. The impulse noise elimination method based on adaptive mean filtering is characterized by: include: Step S1: collect the image data set after noise labeling, perform preprocessing, extract features from the preprocessed image data set, and combine traditional low-level features to form a multi-scale feature fusion map; Step S2: Performing in-depth analysis on the multi-scale feature fusion image, predicting the noise area, and dynamically adjusting the upper and lower boundaries of the noise area using a dynamic boundary adaptive adjustment method; Step S3: according to the determined noise area, a dynamic weight function and a dynamic noise adjustment threshold are set, and according to the statistical characteristics of the pixels in the neighborhood of the noise area, a filtering window is set for each noise point, and each pixel in the noise image is traversed using the filtering window. In each filtering window, different weights are assigned to each pixel using a dynamic weight function according to the distance between the pixel and the noise center point and the gray value difference, and a weighted mean is calculated; Step S4: According to the weighted mean, determine whether the current pixel is a noise point. If it is a noise point, replace the current pixel with the mean value of the pixels in the filter window, and determine again whether there is a noise point. If so, increase the filter window and calculate the noise change rate. If the noise change rate is less than the noise adjustment threshold, stop traversing the noise image and save the process data. The specific steps of step S2 include: S2.1: The obtained multi-scale feature fusion map is input into the trained CNN model. In the fully connected layer and output layer of the CNN model, the CNN model generates a probability map of the same size as the input image based on the fusion features; S2.2: Setting thresholds ,like , then the pixel on the probability map Belonging to the noise area, obtain the binary noise area prediction map, where, represents the i-th pixel on the probability map, i represents the index position of the pixel on the probability map; S2.3: Set the initial noise area boundary based on the noise area prediction map , analyze the initial boundary neighborhood pixel features, and evaluate whether the corresponding pixel belongs to the noise area based on the pixel features; like , then the boundary neighborhood pixels are added to the noise area, where represents the boundary neighborhood pixel features, represents the pixel features inside the noise area, To express similarity; like , then the boundary neighborhood pixels are excluded from the noise area, where Represents the pixel features of the background area; S2.4: Set the boundary threshold of the noise area to , repeat S2.2-S2.3 until the boundary of the noise area changes satisfy: .

2. The impulse noise elimination method based on adaptive mean filtering according to claim 1, characterized in that: The specific steps of step S1 include: S1.1: Collect a dataset of images labeled with noise , and perform data enhancement on X by rotation and translation, where represents image data labeled with building noise, represents image data labeled with traffic noise, represents the image data marked with construction noise, n represents the number of building noise image data, m represents the number of traffic noise image data, and a represents the number of construction noise image data; S1.2: Adjust the size of the image dataset X and preprocess it, use the CNN model API to instantiate the model, and traverse the preprocessed image dataset. For each image in the dataset, use the convolutional basis of the CNN model to perform forward propagation to obtain feature maps, and perform feature extraction on the feature maps of different layers; S1.3: Perform global average pooling on the extracted features and concatenate and fuse them with color, shape, and texture features to obtain fused features. Based on the fused features, a multi-scale feature fusion graph is formed through a multi-scale feature fusion algorithm.

3. The impulse noise elimination method based on adaptive mean filtering according to claim 2, characterized in that: The specific formula of the dynamic weight function in step S3 is: ； in, represents the dynamic weight, , represents the balance factor, Represents the pixels in the noisy image The distance from the center of the noise, Represents the pixels in the noisy image The gray value of Represents the gray value of the center point of the noise, Represents the pixels in the noisy image The local variance within the neighborhood, , , Represent the standard deviation of distance, gray value difference and local variance respectively.

4. The impulse noise elimination method based on adaptive mean filtering according to claim 3, characterized in that: The formula of the weighted mean in step S3 is: ； in, Indicates that the image after filtering is at position Gray value on, W represents the current filter window, Represents the noise image at position The gray value on , Represents the offset function.

5. An impulse noise elimination system based on adaptive mean filtering, which is used to implement the impulse noise elimination method based on adaptive mean filtering according to any one of claims 1 to 4, characterized in that: include: Data processing module, feature fusion module, mean filtering module, effect optimization module; The data processing module is used to pre-process the noise-marked image data set; The feature fusion module is used to integrate the features extracted by CNN and traditional low-level features to form a multi-scale feature fusion graph; The mean filter module is used to predict the noise area and eliminate the noise using an adaptive mean filter method according to the determined noise area; The effect optimization module is used to optimize the filtering effect and enhance and evaluate the optimized filtered image.

6. The impulse noise elimination system based on adaptive mean filtering as claimed in claim 5, characterized in that: The mean filter module includes: a region prediction unit, a boundary adjustment unit, and an adaptive filter unit; The region prediction unit is used to predict the noise region using a deep learning method; The boundary adjustment unit is used to dynamically adjust the boundary of the predicted noise area; The adaptive filtering unit is used to dynamically adjust the weights to perform weighted mean calculation according to the relative positions and gray value difference information between pixels within a set filtering window.

7. The impulse noise elimination system based on adaptive mean filtering as claimed in claim 6, characterized in that: The effect optimization module includes: a residual learning unit and an image enhancement unit; The residual learning unit is used to utilize the deep residual learning concept to perform iterative optimization by comparing the residual of the original image and the processed image, and continuously fine-tune the filter parameters until a denoising effect is achieved; The image enhancement unit is used to perform contrast enhancement on the optimized filtered image using an enhancement algorithm.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the impulse noise elimination method based on adaptive mean filtering according to any one of claims 1 to 4 are implemented.

Images