A method and system for visually identifying physical damage on the surface of an MBR flat membrane

Abstract

This invention belongs to the field of image processing technology, specifically relating to a method and system for visual recognition of physical damage on the surface of an MBR flat sheet membrane. The method includes: low-pass filtering a microscopic grayscale image to melt the micropore background; calculating a texture variation index based on the outlier degree of grayscale values ​​of neighboring pixels and the dynamic contrast gain of the neighboring grayscale; determining damage clustering weights using spatial attenuation accumulation values ​​to remove isolated noise points; fusing the texture variation index and damage clustering weights through an adaptive gating function to generate a damage confidence score; and finally, marking connected damage regions based on statistical threshold binarization and morphological processing. This invention effectively suppresses interference from dense micropores and improves the accuracy of physical damage recognition.

Description

Technical Field

[0001] This invention relates to the field of image processing technology. More specifically, this invention relates to a method and system for visually recognizing physical damage on the surface of an MBR flat sheet membrane. Background Technology

[0002] The core component of membrane bioreactor (MBR) technology is the flat sheet membrane, whose surface is densely covered with micron-sized pores to achieve efficient filtration. During the manufacturing process, the membrane surface is highly susceptible to physical damage such as mechanical scratches and pressure from hard objects, leading to perforations or fiber layer peeling. Once damaged, it not only significantly reduces filtration efficiency but also allows contaminants to penetrate directly, seriously affecting the quality of the effluent. Therefore, designing a visual inspection method that can automatically identify physical damage to the surface of MBR flat sheet membranes is of great significance for ensuring product quality and stable system operation.

[0003] Currently, in the task of defect detection on the surface of MBR flat sheet membrane, the suppression of background texture is the core challenge. Wavelet Transform, due to its excellent multi-resolution analysis capability, is often used for defect extraction in complex backgrounds. By using wavelet decomposition to separate the image into low-frequency approximate components and high-frequency detail components, and discarding high-frequency information containing noise, it can effectively remove unstructured background interference and restore the main outline of the damage.

[0004] However, when processing microscopic images of MBR flat sheet membranes, the micropore structure on the membrane surface is extremely dense and unevenly distributed. In the frequency domain, it is not simply high-frequency noise, but rather permeates into the mid-to-low frequency band. The reconstructed background area still retains the wavy gray-scale fluctuations caused by the difference in micropore density, which makes the calculated global threshold too low. This leads to a large number of darker normal micropore areas being misjudged as damage, thus affecting the accuracy of physical damage identification on the flat sheet membrane surface. Summary of the Invention

[0005] To address the technical problem of high false alarm rate and low accuracy in MBR membrane surface defect detection caused by the inability of existing technologies to accurately separate damage in a dense microporous background, the present invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides a method for visually recognizing physical damage on the surface of an MBR flat sheet membrane, comprising: acquiring a microscopic grayscale image of the surface of the MBR flat sheet membrane, and performing low-pass filtering on the image to obtain a filtered image; for any pixel in the filtered image, determining a texture variation index based on the outlier degree of the grayscale value of the pixel relative to the mean grayscale value of its neighboring pixels, and the dynamic contrast gain of the grayscale range of its neighboring pixels; acquiring all pixels within a clustering window of the pixel, and performing a weighted summation of the texture variation indices of each pixel based on the distance from each pixel within the clustering window to the pixel to obtain a spatial attenuation cumulative value; Damage clustering weights are determined based on the mean of the texture variation indices of all pixels within the clustering window and the cumulative spatial decay value. Damage confidence is calculated, which is determined based on the product of the texture variation index, damage clustering weight, and an adaptive gating function for that pixel. The adaptive gating function responds to the degree to which the texture variation index of that pixel exceeds the mean of all texture variation indices. A binarization threshold is calculated based on the damage confidence of all pixels, and binarization is performed. Morphological processing is applied to the binarized image to obtain the number of pixels in each connected region. If the number of pixels in a connected region is greater than the damage threshold, the connected region is marked as a damaged region.

[0007] This invention achieves physical signal-to-noise separation by blurring the microporous background through low-pass filtering; it calculates the texture variation index and combines the grayscale outlier degree with the neighborhood dynamic contrast to capture high-contrast features; it introduces spatial attenuation accumulation value to determine damage clustering weights and uses the spatial continuity of physical damage to distinguish isolated noise; it calculates damage confidence based on an adaptive gating function to effectively suppress weak background signals; and finally, it combines statistical thresholding and morphological processing to restore discontinuous damage to complete patches, solving the problem of high false alarm rate in dense microporous backgrounds and improving recognition accuracy.

[0008] Preferably, the step of acquiring a microscopic grayscale image of the MBR flat sheet membrane surface and performing low-pass filtering on the image to obtain a filtered image includes: the convolution kernel size of the low-pass filter is... The standard deviation is 1.3. A low-pass filter is used to perform convolution operation on the micro grayscale image to obtain the filtered image.

[0009] This invention employs a 9×9 convolution kernel, whose size is slightly larger than twice the diameter of the micropores. By utilizing the filtering and blurring properties, it smoothly integrates the small-scale dense micropore features into the background while preserving the gray-scale depressions or protrusions of large-scale physical damage. This achieves physical-level signal-to-noise separation before feature extraction, reducing the difficulty of subsequent processing.

[0010] Preferably, the method for obtaining the neighboring pixels is as follows: taking the pixel as the center point, select a range of pixels of size... Neighborhood window, and For odd numbers greater than 20, all pixels within the neighborhood window are marked as neighboring pixels of that pixel; the size of the aggregation window is... ,and Less than .

[0011] Preferably, the texture variation index satisfies the expression: In the formula, For the first Texture variation index per pixel; For the first The grayscale value of each pixel; For the first The mean and standard deviation of the gray values ​​of all neighboring pixels of a pixel; For hyperparameters; For the first The range of gray values ​​among all neighboring pixels of a pixel; It is a natural exponential function.

[0012] This invention combines grayscale outlier degree and local dynamic contrast gain. Grayscale outlier degree assesses the degree to which the grayscale value of a pixel deviates from the neighborhood mean, while local dynamic contrast gain uses the range to amplify the signal in high-contrast areas. Combining the two ensures that damaged points in drastically changing areas obtain an extremely high anomaly index, effectively distinguishing normal backgrounds from damaged areas under uneven lighting.

[0013] Preferably, the spatial attenuation cumulative value satisfies the expression: In the formula, For the first Cumulative spatial attenuation value of each pixel; For the first Within the clustering window of the nth pixel Texture variation index per pixel; This represents the index value and total number of pixels within the aggregation window; Represented as the first The pixel and its clustering window Euclidean distance of pixels; It is the natural logarithm function; It is the maximum-minimum normalization function.

[0014] This invention assesses the spatial clustering of outliers in the neighborhood by using an inverse square distance weighting. If highly anomalous points are closely clustered around the center, the cumulative value increases exponentially; if they are scattered or isolated, they are suppressed. This mechanism utilizes the characteristic of physical damage appearing in patches to effectively eliminate random isolated noise points generated by electronic imaging and enhance the connectivity characteristics of the damage.

[0015] Preferably, the damage aggregation weight is equal to the product of the mean of the texture variation index of all pixels within the aggregation window of the pixel and the cumulative spatial decay value of the pixel.

[0016] Preferably, the damage confidence level satisfies the expression: In the formula, For the first Damage confidence of each pixel; For the first Texture variation index per pixel; For the first Damage aggregation weights for each pixel; This is the mean of the texture variation index of all pixels in the filtered image; It is a natural exponential function; This represents the background suppression coefficient.

[0017] This invention utilizes an adaptive gating function to construct a nonlinear mapping mechanism, which exponentially suppresses weak signals below the background noise benchmark to prevent false positives; it uses aggregation weights to nonlinearly amplify damage signals that are significantly higher than the background noise; and it minimizes misjudgments caused by normal micropores by using a background suppression coefficient, thereby improving detection robustness.

[0018] Preferably, calculating the binarization threshold and performing binarization includes: calculating the mean of the damage confidence of all pixels in the filtered image. and standard deviation Binarization threshold ; will be greater than The grayscale value of the pixel is set to 255, otherwise it is set to 0.

[0019] Preferably, the step of performing morphological processing on the binarized image to obtain the number of pixels in each connected component includes: using... Large-scale structuring elements are used to perform a closing operation on a binary image, and the resulting image is then processed using... The small-scale structuring elements are opened to obtain the processed binary image, and the connected components are labeled to obtain the number of pixels in each connected component.

[0020] Secondly, the present invention provides a visual recognition system for physical damage on the surface of an MBR flat sheet membrane, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned visual recognition method for physical damage on the surface of an MBR flat sheet membrane is implemented.

[0021] By adopting the above technical solution, a computer program is generated from the above-mentioned method for visual recognition of physical damage on the surface of MBR flat sheet membrane, and stored in the memory so that it can be loaded and executed by the processor. In this way, a terminal device can be made based on the memory and the processor for convenient use.

[0022] The beneficial effects of this invention are as follows:

[0023] (1) The present invention constructs a three-level suppression system of frequency domain filtering-spatial aggregation-statistical gating, from physical-scale micro-aperture smoothing to spatial-scale isolated noise removal and statistical-scale background signal suppression, filtering dense micro-aperture interference and significantly reducing false alarm rate.

[0024] (2) In view of the problem that damage is distributed in a discontinuous filamentous manner from a microscopic perspective, the present invention adopts a combination of large-scale closing operation and small-scale opening operation to forcibly bind the broken damage points into a whole patch and filter out small noise, thus successfully restoring the true physical range of the damage and improving the accuracy of damage area identification. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating a method for visually recognizing physical damage on the surface of an MBR flat sheet membrane according to the present invention.

[0026] Figure 2 This is a schematic representation of a microscopic grayscale image of the surface of an MBR flat sheet membrane;

[0027] Figure 3 This is a schematic diagram showing the effect of binarized masking after traditional wavelet transform processing;

[0028] Figure 4 This is a schematic diagram illustrating the effect of the binarized mask after processing according to the present invention. Detailed Implementation

[0029] 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, not all, of the embodiments of the present invention. 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.

[0030] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0031] This invention discloses a method for visually recognizing physical damage on the surface of an MBR flat sheet membrane, referring to... Figure 1 This includes steps S1-S5:

[0032] S1. Obtain a microscopic grayscale image of the MBR flat sheet membrane surface, and perform low-pass filtering on the image to obtain a filtered image.

[0033] It should be noted that the most prominent feature of the MBR flat sheet membrane surface from a microscopic perspective is the densely packed filter micropores, which are typically around 3-5 pixels in diameter and exhibit extremely strong high-frequency noise in the frequency domain. In contrast, physical damage is much larger than the micropores in terms of physical scale, with a width typically greater than 20 pixels. If the algorithm directly detects the damage while preserving the details of the micropores, it will be difficult to distinguish between the edges of the micropores and the edges of the damage. Therefore, before feature extraction, the concept of scale space is introduced, and low-pass filtering technology is used to melt the small-scale micropores into the background while preserving the outlines of the large-scale damage, thereby achieving signal-to-noise separation at the physical level.

[0034] Specifically, a scanning electron microscope (SEM) was used to acquire microscopic images of the MBR flat sheet membrane surface to obtain the original grayscale image. .

[0035] Design a convolutional kernel with a size of... In this embodiment, a low-pass filter is selected. That is, slightly larger than twice the diameter of the micropore, standard deviation The original image is convolved using a low-pass filter to obtain the filtered image. , ,in, This is the kernel function.

[0036] Because the size of the micropores is smaller than the radius of the convolution kernel, they are weighted and averaged by the surrounding gray background pixels, thus affecting the filtered image. The grayscale becomes blurred or even disappears, while physical damage, due to its larger size, retains its overall grayscale depression or protrusion features even though the edges become blurred.

[0037] Furthermore, the grayscale value of each pixel in the filtered image needs to be obtained. Simultaneously, the position of each pixel also needs to be determined. Specifically, the pixel at the bottom left corner of the filtered image is taken as the origin, and the horizontal direction to the right from the origin is used as... The positive direction of the axis is defined by taking the vertically upward direction from the origin as... The positive direction of the axis is used to construct a Cartesian coordinate system; the position of each pixel in the filtered image is obtained in the Cartesian coordinate system, including... Coordinates in the axial direction and Coordinates along the axis.

[0038] At this point, the grayscale value and position of each pixel in the filtered image have been obtained.

[0039] S2. For any pixel in the filtered image, determine the texture variation index based on the degree of outlier of the pixel's gray value relative to the average gray value of its neighboring pixels, and the dynamic contrast gain of the gray value difference of its neighboring pixels.

[0040] It should be noted that in the filtered image, the normal film background becomes a smooth gray area, while the damaged area appears as patches with abnormal gray levels. However, due to uneven lighting in industrial production environments or differences in the reflectivity of the film material itself, the baseline brightness of different areas of the image varies. Directly using a globally fixed threshold can lead to misjudgment of normal areas in dark areas or missed detection of damage in bright areas. Therefore, by analyzing the degree of deviation of a pixel from its local neighborhood background, the texture variation index of the region where the pixel is located is evaluated.

[0041] Specifically, for any pixel in the filtered image, its neighboring pixels are obtained. This is done by selecting a range of pixels of size [value missing] centered on the pixel. A neighborhood window is used to mark all pixels within the neighborhood window as neighboring pixels of the current pixel. The size of the neighborhood window needs to be larger than the damage scale and be an odd number to ensure that significant statistical characteristic abrupt changes can be captured when the window slides over the damage edge. In this embodiment of the invention, [the following is taken as an example]. It can be adjusted according to the actual situation.

[0042] The texture variation index of a pixel is determined based on the degree of outlier of its gray value relative to the average gray value of its neighboring pixels, and the dynamic contrast gain of the gray value range of its neighboring pixels; the texture variation index satisfies the expression:

[0043]

[0044] In the formula, For the first Texture variation index per pixel; For the first The grayscale value of each pixel; For the first The mean and standard deviation of the gray values ​​of all neighboring pixels of a pixel; For hyperparameters; For the first The range of gray values ​​among all neighboring pixels of a pixel; It is a natural exponential function.

[0045] in, Reflecting the The degree to which a pixel's grayscale value deviates from the average grayscale value within its neighborhood. The larger this value, the more significantly the pixel's grayscale value still deviates from the average level after considering grayscale fluctuations within the neighborhood, meaning that the pixel may be a damage point different from the background. To prevent the denominator from being zero, the value is set to 0.01 in this embodiment. Since physical damage usually manifests as the coexistence of extremely dark cavities and extremely white broken fibers, the range of gray values ​​in the neighborhood is larger, while the range of smooth background areas is smaller. Reflecting the The dynamic contrast gain within the neighborhood of a pixel; the larger this value, the more severe the black-and-white contrast within the neighborhood of that pixel, further proving the possible presence of damage. In summary... The larger the value, the more likely the pixel is not only a statistical outlier but also located in a region of drastic high contrast change. This means that the pixel is highly likely to be physically damaged, i.e., the larger the texture variation index of the pixel.

[0046] At this point, the texture variation index of each pixel has been obtained.

[0047] S3. Obtain all pixels within the aggregation window of the pixel. Based on the distance from each pixel to the pixel within the aggregation window, perform a weighted summation of the texture variation index of each pixel to obtain the spatial decay cumulative value. Determine the damage aggregation weight based on the mean of the texture variation index of all pixels within the aggregation window and the spatial decay cumulative value.

[0048] It should be noted that random noise is inevitably generated during electronic imaging. These noise points may also exhibit high variation index. However, there is a fundamental difference between noise and physical damage: noise is usually a spatially isolated single point, while physical damage has spatial continuity and will inevitably appear in patches or lines. Therefore, it is necessary to introduce spatial clustering to evaluate its damage clustering weight.

[0049] Specifically, for any pixel, define The aggregation window is used to obtain all pixels within the aggregation window of the pixel. The size of the aggregation window must be smaller than the size of the neighboring window to avoid the dilution of features with minor damage due to excessive smoothing. In this embodiment of the invention, 15 is selected, but it can be adjusted according to the actual situation.

[0050] Based on the distances from each pixel within the aggregation window to that pixel, the texture variation indices of each pixel are weighted and summed to obtain the cumulative spatial attenuation value of that pixel; the cumulative spatial attenuation value satisfies the expression:

[0051]

[0052] In the formula, For the first Cumulative spatial attenuation value of each pixel; For the first Within the clustering window of the nth pixel Texture variation index per pixel; This represents the index value and total number of pixels within the aggregation window; Represented as the first The pixel and its clustering window Euclidean distance of pixels; It is the natural logarithm function; It is the maximum-minimum normalization function.

[0053] in, Reflecting the The spatial distance decay of the texture variation index of pixels within the aggregation window of pixels n is cumulative. The larger the texture variation index of a pixel, the greater the distance from the first pixel. When the pixels are closer together The contribution value increases exponentially; conversely, if the pixel with a high texture variation index is far from the first... Individual pixels that are far apart, or those that are close together but have a low texture distortion index. The contribution value will be significantly suppressed; if The larger the value, the more closely the anomalous signals of high texture variation index of pixels within the aggregation window are focused on the first... The distance is around each pixel, rather than scattered around the window edge. Adding 1 is to avoid the denominator being zero when the distance is zero.

[0054] Furthermore, the product of the mean texture variation index of all pixels within the aggregation window and the cumulative spatial attenuation value of that pixel is calculated as the damage aggregation weight; wherein, the mean texture variation index of the pixels within the aggregation window reflects the... Whether the nth pixel is located in a high anomaly region; if the product value is larger, it indicates that the nth pixel is located in a high anomaly region. The nth pixel is not only located in a high-anomaly region, but also at the high-density core of that anomaly region, suggesting that the nth pixel is... The higher the damage aggregation weight of each pixel, the better.

[0055] At this point, the damage aggregation weight for each pixel has been obtained.

[0056] S4. Calculate the damage confidence, which is determined based on the product of the texture variation index of the pixel, the damage clustering weight, and the adaptive gating function, wherein the adaptive gating function responds to the degree to which the texture variation index of the pixel exceeds the mean of all texture variation indices.

[0057] It should be noted that, considering that there are still many weak signals in the actual image background caused by uneven film surface roughness, in order to construct a high-contrast damage probability map, a non-linear gating mechanism is set: for regions where the signal intensity is lower than the average level of the background, its response value should be suppressed to 0; while for regions where the signal intensity is significantly higher than the background, its response should be preserved and enhanced; this adaptive suppression can minimize false positives.

[0058] Specifically, the damage confidence of any pixel is determined based on its texture variation index and damage clustering weight; the damage confidence satisfies the expression:

[0059]

[0060] In the formula, For the first Damage confidence of each pixel; For the first Texture variation index per pixel; For the first Damage aggregation weights for each pixel; This is the mean of the texture variation index of all pixels in the filtered image; It is a natural exponential function; This represents the background suppression coefficient.

[0061] in, Representing feature enhancement terms, through the first The product of the texture variation index of each pixel and the damage aggregation weight is used to nonlinearly amplify the strong texture variation index signal. This represents an adaptive gating function used to measure the signal relative to a full-image background noise benchmark. The significance of the expression, and its range. ,when When the exponent term is close to or less than the background mean, the exponent term approaches 1, and the entire bracketed term approaches 0, thus suppressing the background; when When the value is significantly greater than the background mean, the exponential term approaches 0, and the bracketed term approaches 1, thus enhancing the signal. In summary, if... The larger the value, the more significant the effect. The first pixel not only exhibits extremely high variation and aggregation characteristics, but its signal strength also far exceeds the background noise level of the entire image, meaning that the first pixel... Each pixel is a valid target for damage; it should be added that, since in actual images, about half of the normal micropore background pixels will naturally have a variation index higher than the mean, setting... The suppression threshold used to adjust the adaptive gating function is, in this embodiment, taken as... This reduces the risk of misjudging normal micropores; when When the value is 0, it indicates that there is no damage. At this point, let... Take 1.

[0062] At this point, the damage confidence level for each pixel has been obtained.

[0063] S5. Calculate the binarization threshold based on the damage confidence of all pixels and perform binarization. Perform morphological processing on the binarized image to obtain the number of pixels in each connected region. If the number of pixels in a connected region is greater than the damage threshold, mark the connected region as a damaged region.

[0064] It should be noted that, from a microscopic perspective, damaged fibers are distributed in a filamentous pattern, interspersed with undamaged gaps, rather than a complete connected domain. If binarization is performed directly and the area is filtered out, these tiny damage points will be filtered out as noise due to their small size, resulting in missed detection. Therefore, morphological closing operations are used to connect adjacent fragments into a whole, thereby restoring the true extent of the damage.

[0065] Specifically, according to the normal distribution theory, values ​​exceeding 2.5 standard deviations from the mean are considered extremely low-probability anomalies, which closely matches the sparsity characteristics of physical damage; the mean damage confidence score of all pixels in the filtered image is calculated. and standard deviation Binarization threshold The damage confidence score is binarized using the threshold to obtain a binarized image; wherein, all values ​​greater than the threshold are... Pixels marked as candidate damage points are set to a grayscale value of 255; less than or equal to The pixels marked as normal points have their grayscale values ​​set to 0.

[0066] Furthermore, morphological processing is performed on the binarized image, including: using... Large-scale structuring elements are used to perform a closing operation on a binary image, and the resulting image is then processed using... The small-scale structuring elements are opened to obtain the processed binary image; connected component labeling is performed on the processed binary image to obtain all connected components and the number of pixels in each connected component. If the number of pixels in a connected component is greater than the damage threshold, the connected component is marked as a damaged region.

[0067] Since the feature width of physical damage is typically greater than 20 pixels, a damage threshold of 300 is selected in this embodiment of the invention through geometric mapping at the physical scale to ensure that the damaged area can be effectively detected; Large-scale structural elements are closed to ensure they are larger than the width of the physical damage, effectively binding closely spaced fracture points into a connected whole patch. However, after binarization and closing, randomly distributed isolated noise points may still remain in the background. This can be addressed by taking... Opening operations on small-scale structuring elements can eliminate random noise points with a diameter of less than 5 pixels, thus achieving a noise reduction effect of preserving large elements and eliminating small ones. Implementers can adjust the damage threshold and the size of the structuring elements according to the actual situation.

[0068] For example, Figure 2 This is a microscopic grayscale image of the surface of an MBR flat sheet membrane. The background of the image is densely covered with a large number of normal filter micropores at the micrometer level, while there are also several obvious perforations and traces of fiber peeling. Figure 3 The image shows the binarized masking effect after traditional wavelet transform processing. Because traditional wavelet transform cannot completely suppress the interference of dense micropore texture, the processing result is filled with a large number of salt and pepper black and white noise. Although the physical damage area is partially extracted, it appears as broken, discontinuous discrete spots, and a large number of normal background micropores are incorrectly marked as damage areas, resulting in severe oversegmentation and failing to meet the detection accuracy requirements. Figure 4 The image shows the binarized mask effect after processing according to the present invention. After the micropore suppression and connectivity enhancement processing of the present invention, the background area appears as a pure black, and most of the normal micropore noise has been effectively removed. At the same time, the damage features that were originally discontinuously distributed under the microscopic level have been successfully repaired into complete, dense, and connected white patches, which significantly enhances the recognition accuracy of physical damage.

[0069] This invention also discloses a visual recognition system for physical damage on the surface of an MBR flat sheet membrane, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a visual recognition method for physical damage on the surface of an MBR flat sheet membrane according to the present invention.

[0070] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

Claims

1. A method for visually identifying physical damage on the surface of an MBR flat membrane, characterized in that, include: A microscopic grayscale image of the MBR flat sheet membrane surface is acquired, and the image is low-pass filtered to obtain a filtered image. For any pixel in the filtered image, the texture variation index is determined based on the degree of outlier of the pixel's gray value relative to the average gray value of its neighboring pixels, and the dynamic contrast gain of the gray value difference of its neighboring pixels. Obtain all pixels within the aggregation window of the given pixel. Based on the distance from each pixel to the given pixel, perform a weighted summation of the texture variation indices of each pixel to obtain the cumulative spatial decay value. Determine the damage aggregation weight based on the mean of the texture variation indices of all pixels within the aggregation window and the cumulative spatial decay value. The damage confidence is calculated based on the product of the texture variation index of the pixel, the damage clustering weight, and the adaptive gating function, which responds to the degree to which the texture variation index of the pixel exceeds the mean of all texture variation indices. The binarization threshold is calculated based on the damage confidence of all pixels and then binarized. Morphological processing is performed on the binarized image to obtain the number of pixels in each connected region. If the number of pixels in a connected region is greater than the damage threshold, the connected region is marked as a damaged region. The texture variation index satisfies the expression: In the formula, For the first Texture variation index per pixel; For the first The grayscale value of each pixel; For the first The mean and standard deviation of the gray values ​​of all neighboring pixels of a pixel; For hyperparameters; For the first The range of gray values ​​among all neighboring pixels of a pixel; It is a natural exponential function; The cumulative spatial attenuation value satisfies the expression: In the formula, For the first Cumulative spatial attenuation value of each pixel; For the first Within the clustering window of the nth pixel Texture variation index per pixel; This represents the index value and total number of pixels within the aggregation window; Represented as the first The pixel and its clustering window Euclidean distance of pixels; It is the natural logarithm function; It is the maximum-minimum normalization function; The damage clustering weight is equal to the product of the mean of the texture variation index of all pixels within the clustering window of that pixel and the cumulative spatial decay value of that pixel; Damage confidence satisfies the expression: In the formula, For the first Damage confidence of each pixel; For the first Damage aggregation weights for each pixel; This is the mean of the texture variation index of all pixels in the filtered image; This represents the background suppression coefficient.

2. The method according to claim 1, wherein the method is characterized by, The step of acquiring a microscopic grayscale image of the MBR flat sheet membrane surface and performing low-pass filtering on the image to obtain a filtered image includes: The convolution kernel size of the low-pass filter is The standard deviation is 1.3, and the low-pass filter is used for convolution operation on the microscopic gray scale image to obtain a filtered image.

3. The method according to claim 1, wherein the method is characterized by, The method for obtaining the neighboring pixels is as follows: taking the pixel as the center point, select a range of sizes... Neighborhood window, and For odd numbers greater than 20, all pixels within the neighborhood window are marked as neighboring pixels of that pixel; the size of the aggregation window is... ,and Less than .

4. The method for visually recognizing physical damage on the surface of an MBR flat sheet membrane according to claim 1, characterized in that, The calculation of the binarization threshold and the binarization process include: Calculate the mean of the damage confidence scores for all pixels in the filtered image. and standard deviation Binarization threshold ; will be greater than The grayscale value of the pixel is set to 255, otherwise it is set to 0.

5. The method according to claim 1, wherein the method is characterized by, The morphological processing of the binarized image to obtain the number of pixels in each connected component includes: use Large-scale structuring elements are used to perform a closing operation on a binary image, and the resulting image is then processed using... The small-scale structuring elements are opened to obtain the processed binary image, and the connected components are labeled to obtain the number of pixels in each connected component.

6. A visual recognition system for physical damage on the surface of MBR flat membranes, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement a visual recognition method for physical damage on the surface of an MBR flat sheet membrane according to any one of claims 1-5.

Images