Filter element filterability test method and system based on visual inspection

Abstract

The invention discloses a filter element filterability test method and system based on visual inspection, and belongs to the technical field of filter element tests.The test method comprises the steps that multiple cameras and area light sources are arranged, calibration is executed, and a calibration parameter set is manufactured; the method comprises the following steps: immersing a filter element into a detection liquid, synchronously acquiring images, correcting and splicing to generate a panoramic image, dividing the panoramic image into a plurality of detection subareas, applying first gas pressure to the interior of the filter element, and extracting a first total bubble area value in each detection subarea through image processing to judge a broken hole defect; second gas pressure is applied, a deep learning bubble segmentation model is adopted to extract a second total bubble area value in each detection subarea to judge the blockage defect or the defect of insufficient aperture ratio, the invention further discloses a system, automatic and quantitative detection of the filter element defects is achieved, and precision, efficiency and safety are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of filter test. More particularly, the present application relates to a filter filtration test method and system based on visual detection. BACKGROUND

[0002] As the core component of industrial filtration system, the performance reliability and consistency of filter directly relate to the quality of terminal product and even the long-term operation life of major equipment. The traditional filter inspection method is to immerse the filter completely into a transparent container filled with specific detection liquid, to pass compressed gas into the interior of the filter, and to rely on the naked eye observation of the detection personnel to observe whether bubbles are generated on the outer surface of the filter, as well as to observe the shape, size and distribution of the bubbles, and then to judge whether the filter is qualified based on experience.

[0003] However, the immersion gas method widely used in the current filter integrity detection has three main limitations: firstly, this method relies on the naked eye observation of the bubble shape and distribution by the operator, and the detection efficiency is low, which is difficult to adapt to the large-scale continuous production rhythm, and since the judgment standard is highly subjective, the judgment standard of "qualified" and "defect" by different detection personnel often differs, and there is a lack of unified and quantitative basis, secondly, the detection liquid often contains toxic volatile components, and long-term close exposure of the operator to such environment may cause health risks, and there is an occupational health and safety hazard, thirdly, the whole process of filter detection lacks an automatic data recording mechanism, and it is impossible to objectively quantify the bubble information, and it is also impossible to generate a traceable detection report, which is difficult to meet the requirements of modern quality control system on data integrity and process traceability.

[0004] Therefore, it is urgent to propose a filter filtration test method and system based on visual detection, to realize the automatic detection of filter bubbles, to replace manual operation, to shorten the detection period, and to improve the accuracy and consistency of bubble identification. SUMMARY

[0005] An object of the present application is to provide a filter filtration test method and system based on visual detection, to realize the automatic detection of filter bubbles, to replace manual operation, to shorten the detection period, and to improve the accuracy and consistency of bubble identification.

[0006] According to one aspect of the present application, the present application provides a filter filtration test method based on visual detection, comprising the following steps: S1, disposing a surface light source and a plurality of cameras on both sides of a transparent detection cylinder, arranging the plurality of cameras side by side along the preset filter length direction, performing camera calibration to obtain a set of calibration parameters, and the set of calibration parameters includes the internal parameters of each camera, the lens distortion coefficient, and the homography matrix between adjacent camera images; S2, completely immersing a filter to be tested into the transparent detection liquid of the transparent detection cylinder; S3, control all cameras to synchronously collect raw images, perform distortion correction on the raw images based on the set of calibration parameters, and perform image stitching using the homography matrix to generate a panoramic image of the bubble generation area on the circumference of the filter element to be tested, and vertically divide the panoramic image into a plurality of independent detection zones; S4, apply a first gas pressure lower than the nominal bubble point pressure of the filter element to the inside of the filter element to be tested, execute step S3 to obtain a first reference image, perform image processing on the image corresponding to each detection zone in the first reference image, count the single-bubble regions in each detection zone whose pixel area exceeds a single-bubble area threshold, and aggregate the pixel areas of the single-bubble regions that meet the condition to form a first total bubble area value. If the first total bubble area value is higher than a first area threshold, it is determined that the filter element to be tested has a broken hole defect in the detection zone. S5, apply a second gas pressure not lower than the nominal bubble point pressure of the filter element to the inside of the filter element to be tested, execute step S3 to obtain a second reference image, perform semantic segmentation on the image corresponding to each detection zone in the second reference image using a bubble segmentation model trained based on deep learning, extract the total pixel area of the pixels identified as bubbles in each detection zone as a second total bubble area value, and if the second total bubble area value in any detection zone is lower than a second area threshold, it is determined that the filter element to be tested has a clogging defect or insufficient opening rate in the detection zone.

[0007] Preferably, the step of performing camera calibration to obtain a set of calibration parameters comprises: A1, using a calibration board, perform pose image acquisition in multiple poses in the field of view of each camera, and calculate the internal parameter matrix and lens distortion coefficient of each camera based on all the pose images; A2, place a transparent calibration substrate covering the fields of view of all cameras between the area light source and the plurality of cameras, the calibration substrate being provided with a plurality of groups of graphic symbols for providing matching features, place the calibration board in front of the calibration substrate, and each camera field of view contains only one calibration board and has an overlapping area with the adjacent camera field of view containing the same graphic symbol; A3, control all cameras to synchronously collect raw images of the calibration substrate, perform distortion correction on the raw images based on the internal parameter matrix and lens distortion coefficient of each camera to obtain calibration images, use a feature point extraction algorithm to obtain feature points in each calibration image, perform preliminary matching of the feature points of adjacent calibration images, use a false match screening algorithm to screen out correct matching feature point pairs, and calculate the homography matrix between each two adjacent calibration images based on the screened matching feature point pairs; A4, store the internal parameters and distortion coefficients of all cameras obtained in step A1 and the homography matrices between all adjacent cameras obtained in step A3 as a set of calibration parameters.

[0008] Preferably, the feature points in each calibration image are obtained by the following steps: B1, calculating the gray level change function of the local window in the calibration image when moving in each direction E(u,v) wherein, (x, y) is the coordinate of the pixel point in the calibration image, (u, v) represents the small displacement amount of the window in the horizontal and vertical directions, W(i, j) is the window function, I(i, j) represents the gray level value at the point (i, j) w is the window radius, and the value range is 3-7 pixels; B2, constructing the Harris matrix based on the gray level change function M wherein, I x (i, j) and I y (i, j) are the gradient values in the x and y directions at the point (i, j) B3, calculating the corner response function according to the matrix M R , wherein, det(M) is the determinant of the matrix M , is the trace of the matrix trace(M) , is an empirical constant, and the point with a value exceeding a set threshold is determined as a feature point. M k R

[0009] Preferably, the homography matrix between the two adjacent calibration images is realized by the random sample consensus algorithm, including the following steps: C1, randomly extracting four pairs of matched points from the preliminary matched feature point pairs as a minimum sample set for calculating the initial homography matrix model parameters; C2, calculating the error of all feature point pairs with the initial homography matrix model, counting the number of inliers with an error less than a set tolerance, repeating the above random sampling, model fitting and consistency checking process, and retaining the model with the largest number of inliers as the optimal model; C3, terminating the calculation when the maximum iteration number N is reached, wherein the maximum iteration number wherein, p is the expected success rate, ω is the inlier proportion estimate value,​​​​​​​​n The minimum sample number required for fitting the model, and finally re-fitting with all inliers to obtain the homography matrix for calculating the transformation relationship between adjacent camera images.

[0010] Preferably, step S4 comprises the following steps: S40, applying a first gas pressure lower than the nominal bubble point pressure of the filter element to the inside of the filter element to be tested, and performing step S3 to obtain a first reference image; S41, performing smoothing processing on the image corresponding to each detection partition in the first reference image using a mean filtering algorithm to obtain a first partition image; S42, processing the first partition image using a dynamic threshold segmentation algorithm, calculating an adaptive threshold value for each pixel point based on a local neighborhood, and identifying and segmenting each single bubble region through a connected component analysis algorithm; S43, calculating the pixel area of each single bubble region, and counting the single bubble regions with a pixel area greater than a single bubble area threshold value in each first partition image, and then adding the pixel areas of the single bubble regions that meet the condition to form a first total bubble area value, and if the first total bubble area value is higher than a first area threshold value, it is determined that the filter element to be tested has a broken hole defect in the corresponding inspection partition, and the single bubble area threshold value and the first area threshold value are both pixel area values that are pre-set.

[0011] Preferably, the single bubble area threshold value is set according to the critical defect hole diameter determined according to the product quality standard of the filter element to be tested. D 0 , and calculating the theoretical value of the bubble volume wherein V B is the theoretical value of the bubble volume, c 1 is a dimensionless coefficient related to factors such as the shape of the gas-liquid interface, σ is the surface tension of the detection liquid, is the density of the detection liquid, is the bubble density, g is the acceleration of gravity, and the corresponding theoretical projection area of the bubble , based on the pixel size and optical magnification obtained through camera calibration, the theoretical projection area of the bubble is converted into the corresponding pixel area, and the pixel area is used as the single bubble area threshold value.

[0012] Preferably, step S5 comprises the following steps: S50, applying a second gas pressure not lower than the nominal bubble point pressure of the filter element to the inside of the filter element to be tested, and performing step S3 to obtain a second reference image; S51, using a plurality of qualified filter elements to collect panoramic images at the second gas pressure and processed by step S3 as training panoramic images, according to the size of the detection zone, the training panoramic images are cropped into a plurality of standard size training images, and the bubble area in the training images is manually labeled to generate corresponding binary mask labels; S52, using a U-Net convolutional neural network with an encoder-decoder architecture as a semantic segmentation model, using the training images as input and the corresponding binary mask labels as training targets, using a combined loss function to optimize network parameters, and training the bubble segmentation model through a back propagation algorithm; S53, pre-processing the image corresponding to each detection zone in the second reference image to obtain a second partition image, inputting the second partition image into the trained bubble segmentation model, and obtaining a pixel-level probability map; S54, binaryzation and segmentation of the pixel-level probability map to obtain a final bubble region segmentation mask, counting the total number of pixels identified as bubbles in the segmentation mask corresponding to each second partition image to obtain a second total bubble area value of the detection zone, and if the second total bubble area value in any detection zone is lower than the second area threshold, it is determined that the filter element to be tested has a clogging defect or insufficient opening rate in the detection zone.

[0013] Preferably, the combined loss function L is composed of a cross-entropy loss function L ce and a Dice loss function L Dice weighted and summed, and the calculation formula is wherein α、β is a weight coefficient, and α+β = 1, using an Adam optimizer to perform a back propagation process, and according to the partial derivative calculated by the combined loss function L from the network prediction result and the true label, iteratively optimizing the weight parameters of the semantic segmentation model until the semantic segmentation model converges.

[0014] Preferably, the first gas pressure is 0.95 times or less than the nominal bubble point pressure of the filter element, the second gas pressure is set to 1.05 times or more than the nominal bubble point pressure of the filter element, and the nominal bubble point pressure of the filter element wherein D 1 is the nominal accuracy of the filter element, θ is the contact angle ,η is the hole shape correction coefficient.

[0015] According to another aspect of the present application, there is also provided a filter element filtration test system based on visual detection, which is used to implement the above method, comprising: A transparent cylinder is used to hold a transparent detection liquid and accommodate a filter core to be tested; A backlight source is arranged on one side of the transparent cylinder to provide uniform backlight illumination for the surface of the filter core; A gas pressure control unit is connected to the filter core to be tested, and is used to introduce gas into the filter core to be tested and accurately control the first gas pressure and the second gas pressure; An image acquisition module includes a plurality of industrial cameras arranged side by side; An image processing module is connected to the image acquisition module and includes: An image preprocessing unit is used to correct the distortion of the image collected by the camera based on a set of calibration parameters of the system, perform image stitching, generate a panoramic image of the bubble generation area, divide the panoramic image into a plurality of independent detection zones, and output an image set of each detection zone; A bubble point detection unit is used to identify a single bubble area with a pixel area greater than a single bubble area threshold after receiving a first reference image, and output a line chart containing a first total bubble area value in each detection zone; A uniformity detection unit is used to identify a second total bubble area value of each detection zone after receiving a second reference image, and output a line chart containing a second total bubble area value in each detection zone.

[0016] The present application at least includes the following beneficial effects: First, the present application adopts a plurality of industrial cameras arranged in parallel along the axis of the filter core, cooperates with a high-brightness area light source to form a uniform illumination system, constructs a panoramic image of the bubble generation area on the surface of the filter core through precise optical calibration and image stitching algorithm, realizes the full automation of the detection process, and converts the original long-time visual observation process into machine automatic completion, significantly improves the detection speed, and can meet the requirements of detection efficiency for large-scale production. At the same time, the system quantitatively analyzes the pixel area and number of bubbles in each detection zone, eliminates the subjective judgment difference between different operators, and ensures the consistency and reliability of the detection results.

[0017] Second, the present application has high automation and is fully suitable for non-contact automatic detection. The filter core is replaced by cooperating with a mechanical arm, a robot and other automatic equipment. The operator only needs to complete the work of filling and obtaining the workpiece after the robot, and then the robot fills the workpiece. The subsequent gas pressure application, image acquisition and processing analysis can be automatically completed by the system, so that the operator is away from the detection liquid volatile gas environment, effectively avoids the health risks that may be caused by long-term contact with industrial detection liquid containing toxic volatile components, and significantly improves the safety level of detection operation.

[0018] Third, a perfect data collection and recording mechanism is established, which can automatically save the panoramic image of bubble generation on the surface of the filter element and the quantitative data of each detection partition, including the first total bubble area value and the second total bubble area value and other key parameters. These data are automatically associated with product number, detection time, pressure parameter and other metadata to generate a structured detection report, realizing comprehensive data recording of the detection process and providing reliable data support for product quality traceability, process optimization and quality system certification.

[0019] Fourth, for the technical difficulties of filter element defect detection, the system innovatively adopts a differentiated image processing strategy. For the broken hole defect, dynamic threshold segmentation and connected component analysis algorithm are adopted to effectively identify and count the number of large bubbles; for the blockage defect, a bubble segmentation model based on deep learning is adopted to realize accurate identification and area calculation of micro bubbles. The multi-camera calibration technology solves the problem of curved surface imaging distortion, and the bubble judgment threshold is scientifically set based on the fluid mechanics model, forming a comprehensive solution that can accurately detect multiple defect types at the same time, improving the scientificity and adaptability of the detection system.

[0020] Other advantages, objects and features of the present application will be apparent from the following description, and will be understood by those skilled in the art upon reading and understanding the present application. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The system schematic diagram of one of the technical solutions of the present application; Figure 2 The camera calibration schematic diagram of one of the technical solutions of the present application, wherein (a) is a calibration image schematic diagram, and (b) is a panoramic image schematic diagram; Figure 3 The image stitching effect diagram in the embodiment of the present application, wherein (a) is an image before stitching, and (b) is a panoramic image after stitching; Figure 4 The detection result diagram of the filter element to be tested under the first gas pressure in the embodiment of the present application; Figure 5 The A part enlarged view in the embodiment of the present application; Figure 6 The detection result diagram of the filter element to be tested under the second gas pressure in the embodiment of the present application; Figure 7 The B part enlarged view in the embodiment of the present application. DETAILED DESCRIPTION

[0022] The present application will be further described in detail below, so that those skilled in the art can implement it with reference to the description.

[0023] It should be understood that the terms such as "have", "contain" and "include" used herein do not exclude the presence or addition of one or more other elements.

[0024] As shown in Figure 1 The application provides a filter core filtration test method based on visual detection, characterized in that it comprises the following steps: S1, a surface light source and a plurality of cameras are arranged on both sides of a transparent detection cylinder, the plurality of cameras are arranged side by side along the preset filter core length direction, camera calibration is performed to obtain a set of calibration parameters, the set of calibration parameters includes the internal parameters of each camera, the lens distortion coefficient, and the homography matrix between adjacent camera images. Due to the inevitable distortion of the camera lens and the fact that the relative positions of the plurality of cameras are not strictly aligned, if these geometric errors are not accurately calibrated and corrected, the subsequent spliced panoramic image will have serious distortion and misplacement, which will affect the accuracy of the defect determination.

[0025] Specifically, first, a high-brightness surface light source is fixed on one side of the transparent detection cylinder, and a plurality of industrial cameras are arranged side by side on the other side of the cylinder, ensuring that the optical axes of all cameras are approximately perpendicular to the axial direction of the filter core, and then camera calibration is performed. Camera calibration can use a multi-view geometry-based calibration method, using a known accurate size checkerboard or circular dot array calibration board, moving and shooting multiple pose images in various poses within the field of view of each camera. By processing these pose images using existing algorithms, the internal parameters and lens distortion coefficients of each camera can be solved. Then, by placing a calibration board or special marker in the common field of view of all cameras, the homography matrix between the images of two adjacent cameras is calculated using appropriate existing algorithms. Finally, all parameters are summarized and stored as a set of calibration parameters for subsequent image processing steps.

[0026] S2, the filter core to be tested is completely immersed in the transparent detection liquid of the transparent detection cylinder. Specifically, the transparent detection liquid is a liquid with stable physical properties such as clear transparency and surface tension. Deionized water or a specific concentration of alcohol aqueous solution can be selected according to the corresponding technical specifications.

[0027] S3, control all cameras to synchronously capture raw images, perform distortion correction on the raw images based on the set of calibration parameters and perform image stitching using the homography matrix, generate a panoramic image of the bubble generation area on the periphery of the filter element to be tested, and vertically divide the panoramic image into multiple independent detection zones. Specifically, trigger all cameras to perform synchronous exposure, capture a complete set of raw images covering the filter element to be tested and the upper region thereof, then call the set of calibration parameters obtained in S1, first perform distortion correction on each raw image to eliminate barrel or pillow distortion of the lens, then use the homography matrix in the set of calibration parameters to project all corrected images to a unified coordinate system, use an image fusion algorithm to smooth the pixels in the overlapping regions, and finally generate a clear and seamless panoramic image of the bubble generation area on the periphery of the filter element. In order to accurately locate the specific section where defects occur and implement a perfect data acquisition and recording mechanism, the panoramic image is vertically divided into multiple independent monitoring zones.

[0028] S4, apply a first gas pressure lower than the nominal bubble point pressure of the filter element to be tested to the inside of the filter element, obtain a first reference image by performing step S3, perform image processing on the image corresponding to each detection zone in the first reference image, count the single-bubble regions in each detection zone whose pixel area exceeds a single-bubble area threshold, and aggregate the pixel areas of the single-bubble regions that meet the condition to form a first total bubble area value. If the first total bubble area value is higher than a first area threshold, it is determined that there is a broken-hole defect in the detection zone of the filter element to be tested. The first gas pressure refers to a test pressure lower than the nominal bubble point pressure of the filter element. At this pressure, the gas at the perfect filter membrane cannot break through the liquid film to form bubbles, while the broken-hole defect of the filter element continuously produces large bubbles due to the large pore size. The single-bubble area threshold is a pre-set pixel area value used to distinguish between large bubbles generated by broken holes and possible noise or small bubbles.

[0029] Specifically, the pressure applied to the inside of the filter core to be tested is accurately adjusted to a first gas pressure by the air pressure control unit and kept stable, and then the image acquisition and processing process in S3 is performed once to obtain a first reference image under the current pressure, the first reference image being a spliced and partitioned image, and then the first partition image corresponding to each detection partition is processed, image filtering algorithm can be used to smooth the image to suppress noise, and then image segmentation algorithm is used to separate the bubble foreground from the background in the image, each independent bubble area is identified through connected component analysis, and its pixel area is calculated, all single bubble areas with a pixel area greater than a single bubble area threshold in the detection partition are counted, the total pixel area value of all single bubble areas meeting the condition is counted as a first total bubble area value, and compared with a first area threshold, if the first total bubble area value of a detection partition exceeds the first area threshold, it is determined that the filter core part corresponding to the detection partition has a broken hole defect, and the lower pressure and the focus on detecting large bubbles can very specifically capture the broken hole defect, while effectively avoiding misjudgment of normally generated small bubbles as defects, ensuring high specificity and low false alarm rate of the detection link.

[0030] S5, a second gas pressure not lower than the nominal bubble point pressure of the filter core is applied to the inside of the filter core to be tested, and step S3 is performed to obtain a second reference image, a bubble segmentation model trained based on deep learning is used to perform semantic segmentation on the image corresponding to each detection partition in the second reference image, and the total area of the pixels identified as bubbles in each detection partition is extracted as a second total bubble area value, if the second total bubble area value in any detection partition is lower than a second area threshold, it is determined that the filter core to be tested has a blockage defect or insufficient opening rate in the detection partition, the second gas pressure is a test pressure not lower than the nominal bubble point pressure of the filter core, under which pressure a large number of small bubbles will be formed on the filter membrane of the intact area of the filter core to be tested, the second area threshold is a pre-set pixel area value representing the minimum total bubble area that a detection partition should generate in a normal state, and the physical manifestation of the blockage defect is completely opposite to that of the broken hole defect, which leads to a decrease or complete disappearance of bubble generation in the blocked part of the filter core. Therefore, sufficient bubbles must be generated in the normal area as background contrast under a pressure higher than the bubble point, and since the small bubbles are intertwined with each other at this time and are difficult to distinguish, the traditional image processing algorithm is difficult to process and identify the number, therefore, a deep learning semantic segmentation model is needed to process the image.

[0031] Specifically, the pressure applied to the inside of the filter element is accurately adjusted to the second gas pressure and kept stable by the air pressure control unit. Then the image acquisition and processing process in S3 is performed again to obtain a second reference image, and a bubble segmentation model trained in advance through deep learning technology is called, which is a semantic segmentation model capable of classifying each pixel as "bubble" or "background" at the pixel level. The second partition image corresponding to each detection partition is input into the model to obtain a binary segmentation mask image of the bubble area, the total number of pixels judged as "bubble" in the segmentation mask image is calculated to obtain the second total bubble area value of the partition, and the second total bubble area value is compared with the second area threshold value. If the second total bubble area value of a certain detection partition is lower than the threshold value, it is determined that there is a blockage defect in the filter element part corresponding to the detection partition.

[0032] In the technical solution, the units of the first area threshold value, the second area threshold value, the single bubble area threshold value, the first total bubble area value, and the second total bubble area value are unified. Optionally, when the camera pixel size and the optical magnification are unified, the number of pixels can be used to represent the area.

[0033] The following is a specific field example: The specific embodiment case is as follows: A petrochemical company filter element regeneration workshop uses a transparent acrylic detection cylinder with a length of 2 meters, a width of 0.8 meters, and a height of 0.5 meters, which contains industrial ethanol detection liquid conforming to the industry standard. A 20000 lux brightness LED surface light source is installed on one side of the detection cylinder to provide uniform backlight for the detection area, and 6 industrial cameras with 20 million pixels are installed side by side along the length direction of the other side of the detection cylinder. The camera is fixed by a support to ensure that its optical axis is perpendicular to the axis of the filter element to be tested. The 6 cameras can completely cover the entire circumferential surface of the filter element to be tested. A high-precision air pressure control unit is connected to the interface of the filter element through a pipeline.

[0034] Before the system is formally put into operation, a one-time calibration process is first performed, and the internal parameters of each camera, the lens distortion coefficient, and the homography matrix between adjacent camera images are calculated and stored, which together constitute the calibration parameter set of the system. The calibration parameter set is directly called in subsequent detection without repeated calibration.

[0035] The operator fully immerses a metal powder sintered filter element to be tested, which is 1.5 meters long, has an outer diameter of DN25, and has a precision of 0.5 microns, in the detection liquid of the detection cylinder, and ensures that the interface of the filter element to be tested is reliably connected to the air pressure pipeline.

[0036] Six cameras are simultaneously triggered to acquire six original images covering the filter to be tested. A calibration parameter set is then called to correct distortion in the six images. Finally, a pre-stored homography matrix is ​​used to stitch the images together to generate a seamless panoramic image, as shown below. Figure 4 As shown, the panoramic image was then divided into 80 consecutive detection zones at equal intervals along its length.

[0037] The air pressure control unit applies a gas pressure of 0.05 MPa to the inside of the filter element under test. After the pressure stabilizes, six cameras are simultaneously triggered within microseconds to acquire six raw images covering the surface of the filter element. These images are then stitched together to form the first reference image. A dynamic threshold segmentation and connected component analysis algorithm is used to quickly identify and calculate the pixel area of ​​each independent bubble in each detection zone. The single bubble area threshold is set to 500 pixels; bubbles with an area greater than 500 pixels are classified as large bubbles. The first area threshold for each detection zone is set to 25,000 pixels (50 bubbles). A line graph showing the number of all detection zones is generated. The detection results are as follows: Figure 4 , Figure 5 As shown, where Figure 4 The vertical axis of the line graph represents the number of bubbles exceeding 500 pixels, in units of individual bubbles. As can be seen from the graph, the number of large bubbles with a pixel area exceeding the single bubble area threshold in detection zone number 72 is 57. The first total bubble area value exceeds the first area threshold, indicating that the filter element has a pore defect in detection zone number 72.

[0038] The air pressure control unit applies a gas pressure of 0.2 MPa to the inside of the filter element under test, triggering all cameras again to form a second baseline image. The pre-trained U-Net deep learning semantic segmentation model is then used to process the image of each detection region, generating a segmentation mask for the bubble region. The total area of ​​pixels identified as bubbles within each detection region is calculated. The second area threshold for each detection region is set to 500,000 pixels. The detection results are as follows: Figure 6 , Figure 7 As shown, where Figure 6 The vertical axis of the line graph represents the total area of ​​pixels identified as bubbles, in thousands of pixels. It was found that the second total bubble area value of all detection zones was higher than the second area threshold, indicating that the filter element under test did not have defects such as blockage or insufficient porosity.

[0039] The whole detection process is completed within 35 seconds, and a detection report containing the filter cartridge number to be tested, detection time, pressure curve, defect partition position information with panoramic icon annotation, and final determination result is automatically generated and stored in the database. The operator can query and review the historical detection data at any time according to the filter cartridge number, successfully replacing the original manual visual detection post, and the single filter cartridge detection efficiency is improved by more than 10 times. The safety risk of the operator contacting toxic ethanol liquid is completely eliminated, the filter cartridge quality is determined by objective and quantitative image data, the result consistency reaches 100%, and the problem of strong subjectivity of manual judgment is solved. All detection data is electronic, which perfectly meets the needs of the petrochemical company for strict traceability and control of filter cartridge quality.

[0040] In another technical solution, the step of performing camera calibration to obtain a set of calibration parameters comprises: A1, using a calibration board to collect pose images in multiple poses in the field of view of each camera, and calculating the internal parameter matrix and lens distortion coefficient of each camera based on all pose images.

[0041] Specifically, in implementation, an operator holds a high-precision calibration board with a checkerboard pattern, moves the calibration board in multiple different angles and distances within the field of view of each camera, triggers the camera to collect an image every time the pose is changed, ensures that the calibration board appears in different positions and orientations in the image, collects 15-25 pose images of different poses for each camera, and then processes the pose images using existing algorithms based on a plane template, such as Zhang Zhengyou algorithm. The algorithm identifies the corner points of the checkerboard in each pose image, and uses the corresponding relationship between the image coordinates of these corner points and their known world coordinates to finally solve the high-precision internal parameter matrix and lens distortion coefficient of the camera through least squares optimization and other mathematical methods.

[0042] A2, a transparent calibration substrate covering the fields of view of all cameras is placed between the area light source and the plurality of cameras, the calibration substrate is provided with a plurality of groups of graphic symbols for providing matching features, the calibration board is placed in front of the calibration substrate, each camera field of view only contains one calibration board and has an overlapping area with the adjacent camera field of view containing the same graphic symbol, and the calibration substrate is a material with good light transmission and flatness, and the graphic symbol is a pattern printed on the calibration substrate and can provide rich and stable features for feature extraction algorithms. Figure 2 As shown in the figure, specifically, the calibration substrate is placed between the installed area light source and the camera array, ensuring that the field of view of any two adjacent cameras must contain at least one set of identical graphic symbols, and a standard calibration board is placed on the transparent substrate, ensuring that there is only one complete calibration board in the central region of the field of view of each camera.

[0043] A3、control all cameras to synchronously collect the original images of the calibration substrate, perform distortion correction on the original images based on the internal parameter matrix of each camera and the lens distortion coefficient to obtain calibration images, use a feature point extraction algorithm to obtain feature points in each calibration image, perform preliminary matching on the feature points of adjacent calibration images, use a false matching screening algorithm to screen out correct matching feature point pairs, calculate the homography matrix between each two adjacent calibration images based on the screened matching feature point pairs, the feature point is a pixel point in the calibration image that has a significant feature and is easy to be repeatedly positioned in different images, the false matching screening algorithm is a robust algorithm for eliminating false matching pairs from the preliminary matching result, and the homography matrix is a 3x3 matrix of the mapping relationship between image pixel coordinates when the same plane is imaged from two different perspectives.

[0044] Specifically, control all cameras to synchronously capture, collect the original images of the substrate with the calibration board and the graphic mark at one time, call the internal parameter matrix and the lens distortion coefficient of each camera obtained in step A1 to perform distortion correction on each original image to obtain geometrically accurate calibration images, then select a feature point extraction algorithm such as Harris corner detection, SIFT algorithm or ORB algorithm to extract a large number of feature points from each calibration image, use a false matching screening algorithm to purify the preliminary matching result, automatically screen out correct matching feature point pairs with high confidence, and finally use the correct matching feature points to calculate an optimal homography matrix through a direct linear transformation algorithm, which can accurately map the points in one image to the coordinate system of another image.

[0045] A4、store the internal parameters and distortion coefficients of all cameras obtained in step A1 and the homography matrices between all adjacent cameras obtained in step A3 as a calibration parameter set, and pack the internal parameter matrix and the lens distortion coefficient of each camera obtained in step A1 and the homography matrix between each two adjacent images obtained in step A3 into a structured calibration parameter set.

[0046] In another technical solution, the feature points in each calibration image are obtained through the following steps: B1、calculate the gray level change function of the local window in the calibration image when moving in each direction E(u,v) : wherein, (x, y) is the coordinate of the pixel point in the calibration image, (u, v) represents the micro displacement amount of the window in the horizontal and vertical directions, W(i, j) is the window function, I(i, j) represents the point (i, j)grayscale value at that location w The window radius is 3 to 7 pixels. Specifically, it iterates through every pixel in the calibration image, and for each point to be judged... (x,y) A local window is extracted centered on this point, and the grayscale change function is calculated after the window undergoes minute displacements in different directions. E(u, v) To assess whether the point might be a corner point, where W(x,y) Pixels in the center of the window are given higher weights, while pixels further away from the center are given lower weights. This makes the calculation of grayscale changes focus more on the center of the window.

[0047] B2. Construct the Harris matrix based on the grayscale change function. M : in, I x (i, j) and I y (i, j) At points respectively (i, j) place, along x and y The gradient value in the direction.

[0048] B3. According to the matrix M Calculate the corner response function R , ,in det(M) For matrix M The determinant, trace(M) For matrix M traces, k Let be an empirical constant, R Points whose values ​​exceed a set threshold are identified as feature points. Specifically, by calculating the corner response function, corner regions in an image can be effectively distinguished from flat or edge regions. When the corner response function value exceeds the set threshold, the point is identified as a feature point, thereby enabling the extraction of stable feature points in the image. The setting of the preset threshold needs to take into account both the sensitivity and specificity of feature point detection. If the threshold is too high, it may lead to the missed detection of real feature points, while if the threshold is too low, it may introduce too many noise points.

[0049] In another technical solution, the homography matrix between the two adjacent calibration images is achieved through a random sampling consensus algorithm, including the following steps: C1, randomly extract four pairs of matching points from the preliminarily matched feature point pairs as a minimum sample set, which is used to calculate the initial homography matrix model parameters, exemplarily, the homography matrix H is a 3x3 matrix, which describes the projection transformation relationship between two planes, and has 8 degrees of freedom, so at least four pairs of matching feature points are needed to solve the model parameters, and the equation set is solved by direct linear transformation algorithm: wherein and are the coordinates of a pair of matching feature points in two calibration images, respectively, and the random sampling method is used to avoid the interference of false matching points on the model calculation, so as to provide initial parameter estimation for subsequent model optimization.

[0050] C2, calculate the error of all feature point pairs with the initial homography matrix model, count the number of inliers with an error less than a set tolerance, repeat the random sampling, model fitting and consistency test process, and keep the model with the most inliers as the optimal model, exemplarily, calculate the projection error of all feature point pairs with the initial homography matrix model, and for the i-th matching point, the projection error calculation formula is: i wherein the project function represents the conversion from homogeneous coordinates to Cartesian coordinates, and the inliers with an error less than a set tolerance τ (1-3 pixels) are counted. Repeat the random sampling, model fitting and consistency test process, and keep the model with the most inliers as the current optimal model.

[0051] C3, terminate the calculation when the maximum number of iterations N is reached, wherein the maximum number of iterations , wherein p is the expected success rate, ω is the inlier ratio estimate, n is the minimum sample size required for the fitted model, and finally, the homography matrix used to calculate the transformation relationship between adjacent camera images is obtained by re-fitting all inliers, exemplarily, wherein p 0.99 is taken, and finally, the final homography matrix is obtained by re-fitting all inliers by least squares method.

[0052] In another technical solution, step S4 comprises the following steps: S40, apply a first gas pressure lower than the nominal bubble point pressure of the filter element to the inside of the filter element to be tested, and perform step S3 to obtain a first reference image.

[0053] ​S41, the image corresponding to each detection partition in the first reference image is smoothed by using a mean filtering algorithm to obtain a first partition image. For example, a sliding window with a size of mxn is used in the mean filtering, and the arithmetic mean of the gray values of all pixels in the sliding window is taken as the new gray value of the center pixel.

[0054] S42, the first partition image is processed by using a dynamic threshold segmentation algorithm, an adaptive threshold value of each pixel point is calculated based on a local neighborhood, and each single bubble region is identified and segmented by using a connected component analysis algorithm. For example, the adaptive threshold value is calculated based on the local neighborhood of each pixel. T(x,y) The calculation formula is as follows: ， Wherein μ (x,y) And σ(x,y) The mean and standard deviation of the gray values of the pixels in the local neighborhood are represented by and respectively, C is a constant adjustment factor. By comparing the gray value of each pixel with the corresponding adaptive threshold value, the image is binarized, and then the connected component analysis algorithm is used to identify and segment each independent single bubble region by using region growing and labeling techniques, so as to accurately separate the bubbles that are in contact or overlap with each other.

[0055] S43, the pixel area of each single bubble region is calculated, the single bubble regions with a pixel area greater than a single bubble area threshold value in each first partition image are counted, the pixel areas of the single bubble regions meeting the condition are summarized to form a first total bubble area value, and if the first total bubble area value is higher than a first area threshold value, it is determined that the filter element to be tested has a broken hole defect in the corresponding inspection partition. The single bubble area threshold value and the first area threshold value are both pixel area values that are set in advance.

[0056] In another technical solution, the single bubble area threshold value is set according to the critical defect hole diameter determined according to the product quality standard of the filter element to be tested. D 0 The theoretical value of the bubble volume is calculated as follows: Wherein V B is the theoretical value of the bubble volume, c 1 is a dimensionless coefficient related to factors such as the shape of the gas-liquid interface, σ is the surface tension of the detection liquid, is the density of the detection liquid, is the bubble density, g is the acceleration of gravity, and the corresponding theoretical projection area of the bubble is , based on the pixel size and optical magnification obtained by camera calibration, the bubble theoretical projection area is converted into the corresponding pixel area, and the pixel area is used as the single bubble area threshold, specifically, the single bubble area threshold is set by using the calculation method based on the physical model, wherein c 1 The empirical value of 0.93 can be obtained. After obtaining the bubble theoretical projection area, the pixel size and optical magnification parameters obtained by camera calibration are used to calculate the actual physical area corresponding to a single pixel, and then the pixel number corresponding to the bubble theoretical projection area, i.e. the single bubble area threshold, is obtained.

[0057] In another technical solution, step S5 includes the following steps: S50, a second gas pressure not lower than the nominal bubble point pressure of the filter element is applied to the inside of the filter element to be tested, and step S3 is performed to obtain a second reference image.

[0058] S51, a plurality of qualified filter elements are used to collect panoramic images under the second gas pressure and process the panoramic images by step S3 to obtain training panoramic images, the training panoramic images are cropped to a plurality of standard size training images according to the size of the detection partition, and the bubble region in the training image is manually labeled to generate a corresponding binary mask label, for example, first, collect not less than 100 groups of qualified filter element samples to collect panoramic images under the second gas pressure and process the panoramic images to obtain training panoramic images, then according to the actual physical size of the detection partition, use bicubic interpolation algorithm to crop all training panoramic images to standard size training images with 512x512 pixels, and use LabelMe to outline the bubble region in each training image during the labeling process to generate a binary mask label file in PNG format, wherein the pixel value of the bubble region is 255 and the pixel value of the background region is 0.

[0059] S52, the U-Net convolutional neural network with an encoder-decoder architecture is used as the semantic segmentation model, the training image is input, the corresponding binary mask label is the training target, the network parameter optimization is performed using a combined loss function, and the bubble segmentation model is trained through a back propagation algorithm. Illustratively, the encoder part uses VGG16 backbone as a feature extractor, contains 13 convolutional layers and 4 maximum pooling layers, gradually extracts image features while reducing the size by half step by step, the decoder part performs upsampling through 4 times of transposed convolution, and each level is connected with the image of the corresponding level of the encoder to fuse low-level detail features and high-level semantic features. The model training is set to a batch size of 16, an initial learning rate of 0.001, and the parameters are optimized using an Adam optimizer. The loss function adopts a linear combination of cross-entropy loss and Dice loss, the training process lasts for 100 rounds, the learning rate is reduced to half of the original every 20 rounds, and the model parameters with the best performance are saved on the validation set.

[0060] S53, the image corresponding to each detection partition in the second reference image is preprocessed to obtain a second partition image, the second partition image is input into the trained bubble segmentation model, and a pixel-level probability map is obtained. Illustratively, the image corresponding to each detection partition in the second reference image to be detected is standardized, including adjusting the size of the second partition image to 512*512 pixels, normalizing the pixel value to the range of [0, 1], inputting the processed second partition image into the trained bubble segmentation model, and the model output is a probability map with the same size as the input image. The value of each pixel point represents the probability value of the point belonging to the bubble category.

[0061] S54, the pixel-level probability map is binarized and segmented to obtain a final bubble region segmentation mask, and the total number of pixels identified as bubbles in the segmentation mask corresponding to each second partition image is counted to obtain a second total bubble area value of the detection partition. If the second total bubble area value in any detection partition is lower than the second area threshold, it is determined that the filter element to be tested has a clogging defect or insufficient opening rate in the detection partition. Illustratively, the probability map is converted into a binary segmentation mask, then 8-connected region analysis is used for post-processing of the segmentation result to remove noise points with an area less than 10 pixels and fill possible holes in the bubble region. Finally, the total number of pixels with a value of 1 in the segmentation mask corresponding to each second partition image is counted to obtain the second total bubble area value of the partition.

[0062] In another technical solution, the combined loss function L is composed of a cross-entropy loss function L ce and a Dice loss function L Dice weighted and summed, and the calculation formula is wherein α、β are weight coefficients, and α+β = 1, the back propagation process is performed by using an Adam optimizer, and the weight parameters of the semantic segmentation model are iteratively optimized according to the partial derivatives calculated from the combined loss function L of the network prediction results and the true labels until the model converges.

[0063] In another technical solution, the first gas pressure is 0.95 times or less than the nominal bubble point pressure of the filter element, the second gas pressure is set to be 1.05 times or more than the nominal bubble point pressure of the filter element, and the nominal bubble point pressure of the filter element wherein D 1 is the nominal accuracy of the filter element, θ is the contact angle ,η is the hole shape correction coefficient, η is 0.5-1.0, used to correct the geometric difference between the ideal cylindrical hole and the actual irregular hole, θ can be empirically valued 0.

[0064] In another technical solution, a filter element filtration test system based on visual detection is also provided for implementing the above method, comprising: a transparent cylinder for containing a transparent detection liquid and accommodating a filter element to be tested; a backlight source arranged on one side of the transparent cylinder to provide uniform backlight illumination for the surface of the filter element; a gas pressure control unit connected to the filter element to be tested, for introducing gas into the interior of the filter element to be tested and accurately controlling the first gas pressure and the second gas pressure; an image acquisition module comprising a plurality of industrial cameras arranged side by side; an image processing module connected to the image acquisition module, comprising: an image preprocessing unit for performing distortion correction on the images collected by the cameras based on a set of calibration parameters of the system and performing image stitching to generate a panoramic image of the bubble generation area, and dividing the panoramic image into a plurality of independent detection zones and outputting the image set of each detection zone; a bubble point detection unit for identifying the single bubble area with a pixel area greater than a single bubble area threshold after receiving the first reference image, and outputting a line chart containing the first total bubble area value in each detection zone; a uniformity detection unit for identifying the second reference image after receiving the second reference image, calculating the second total bubble area value of each detection zone, and outputting a line chart containing the second total bubble area value in each detection zone.

[0065] It is to be understood that even though numerous characteristics and embodiments of the application have been set forth in the foregoing disclosure, the details are for the purpose of example and illustration and not of limitation as to the scope of the application, which is to be measured by the claims and equivalents thereof. It is intended that the scope of the application encompass all technical equivalents which perform similar to the same purpose described above.

[0066] While the embodiments of the application have been disclosed in the foregoing disclosure, it is to be understood that the application is not limited to the above-described embodiments and modifications and changes can be made by those skilled in the art without departing from the spirit and scope of the application as defined by the appended claims and their equivalents.

Claims

1. A filter element filtration test method based on visual detection, characterized by, The method comprises the following steps: S1, disposing a surface light source and a plurality of cameras on both sides of a transparent detection cylinder respectively, arranging the plurality of cameras side by side along a preset filter length direction, performing camera calibration to obtain a set of calibration parameters, and the set of calibration parameters comprising internal parameters of each camera, lens distortion coefficients, and homography matrices between adjacent camera images; S2, completely immersing the filter to be tested into the transparent detection liquid in the transparent detection cylinder; S3, controlling all cameras to synchronously acquire original images, performing distortion correction on the original images based on the set of calibration parameters, and performing image stitching using the homography matrices to generate a panoramic image of a bubble generation area on the side of the filter to be tested, and vertically dividing the panoramic image into a plurality of independent detection zones; S4, applying a first gas pressure lower than a nominal bubble point pressure of the filter to the inside of the filter to be tested, obtaining a first reference image by performing step S3, performing image processing on the images corresponding to each detection zone in the first reference image, counting single-bubble regions in each detection zone whose pixel area exceeds a single-bubble area threshold, and adding up the pixel areas of the single-bubble regions that meet the condition to form a first total bubble area value, and if the first total bubble area value is higher than a first area threshold, determining that the filter to be tested has a broken hole defect in the detection zone; S5, applying a second gas pressure not lower than the nominal bubble point pressure of the filter to the inside of the filter to be tested, obtaining a second reference image by performing step S3, performing semantic segmentation on the images corresponding to each detection zone in the second reference image using a bubble segmentation model trained based on deep learning, extracting a total pixel area of the pixels identified as bubbles in each detection zone as a second total bubble area value, and if the second total bubble area value in any detection zone is lower than a second area threshold, determining that the filter to be tested has a blockage defect or insufficient opening rate in the detection zone.

2. The visual inspection-based test method for filterability of a filter cartridge according to claim 1, characterized by, The step of performing camera calibration to obtain a set of calibration parameters comprises: A1, using a calibration board to perform pose image acquisition in a plurality of poses in the field of view of each camera, and calculating the internal parameter matrix and the lens distortion coefficient of each camera based on all the pose images; A2, placing a transparent calibration substrate covering the fields of view of all the cameras between the surface light source and the plurality of cameras, the calibration substrate being provided with a plurality of groups of graphic marks for providing matching features, placing the calibration board in front of the calibration substrate, and each camera field of view containing only one calibration board and having an overlapping area with the adjacent camera field of view containing the same graphic mark; A3, controlling all cameras to synchronously acquire original images of the calibration substrate, performing distortion correction on the original images based on the internal parameter matrix and the lens distortion coefficient of each camera to obtain calibration images, respectively acquiring feature points in each calibration image using a feature point extraction algorithm, and after preliminary matching of the feature points of adjacent calibration images, screening out correct matching feature point pairs using a mismatch screening algorithm, and calculating the homography matrix between each two adjacent calibration images based on the screened matching feature point pairs; A4, storing the internal parameters and the distortion coefficients of all the cameras obtained in step A1 and the homography matrices between all adjacent cameras obtained in step A3 together as the set of calibration parameters.

3. The visual detection-based filter element filtration test method according to claim 2, characterized by, The feature points in each calibration image are obtained, including the following steps: B1. Computing a function of the gray level variation of a local window in the calibration image when moving in various directions E(u, v) : wherein, (x, y) is the coordinate of a pixel point in the calibration image, (u, v) represents a small displacement amount of the window in the horizontal and vertical directions, W(i, j) is a window function, I(i, j) represents a gray value at the point (i, j) , w is a window radius, and the value range is 3-7 pixels; B2. constructing a Harris matrix based on the gray scale variation function M : wherein I x (i, j) and I y (i, j) gradient values in the x and y directions, respectively, at points (i, j) , B3. According to the matrix M Calculate the corner response function R , ,in det(M) For matrix M The determinant, trace(M) For matrix M traces, k Let be an empirical constant, R Points whose values ​​exceed a set threshold are identified as feature points.

4. The visual inspection-based test method for filterability of a filter cartridge according to claim 2, characterized by, The homography matrix between the two adjacent calibration images is realized by a random sample consensus algorithm, including the following steps: C1, randomly selecting four pairs of matched points from the preliminary matched feature point pairs as a minimum sample set for calculating the initial homography matrix model parameters; C2, calculating the error of all feature point pairs with the initial homography matrix model, counting the number of inliers with an error less than a set tolerance, repeating the random sampling, model fitting and consistency testing process, and retaining the model with the most inliers as the optimal model; C3. terminate the computation when the maximum number of iterations N is reached, said maximum number of iterations wherein p is the expected success rate, ω is the inlier ratio estimate, n is the minimum number of samples required for the fitting model, and finally re-fitting with all inliers yields the homography matrix used for computing the transformation between the adjacent camera images.

5. The visual inspection-based test method for filterability of a filter cartridge according to claim 1, characterized by, Step S4 includes the following steps: S40, apply a first gas pressure lower than the nominal bubble point pressure of the filter element to the inside of the filter element to be tested, and execute step S3 to obtain a first reference image; S41, perform mean filtering algorithm on the image corresponding to each detection partition in the first reference image to obtain a first partition image; S42, process the first partition image using a dynamic threshold segmentation algorithm, calculate the adaptive threshold value of each pixel point based on the local neighborhood, and identify and segment each single bubble region through a connected component analysis algorithm; S43, calculate the pixel area of each single bubble region, count the single bubble regions with a pixel area greater than a single bubble area threshold in each first partition image, and aggregate the pixel areas of the single bubble regions meeting the condition to form a first total bubble area value, if the first total bubble area value is higher than a first area threshold, it is determined that the filter element to be tested has a broken hole defect in the corresponding detection partition, and the single bubble area threshold and the first area threshold are both pixel area values preset.

6. The visual inspection based filter cartridge filtration integrity test method as claimed in claim 1, wherein, The method for setting the single-bubble area threshold value is to determine a critical defect aperture according to a product quality standard to be tested D 0 , calculate a theoretical value of the bubble volume wherein V B is the theoretical value of the bubble volume, c 1 is a dimensionless coefficient related to factors such as the gas-liquid interface morphology, σ is the surface tension of the detection liquid, is the density of the detection liquid, is the bubble density, g is the acceleration of gravity, and the corresponding theoretical projection area of the bubble , the theoretical projection area of the bubble is converted into a corresponding pixel area as the single-bubble area threshold value based on the pixel size and the optical magnification obtained through camera calibration.

7. The visual inspection-based test method for filterability of a filter cartridge according to claim 1, characterized by, Step S5 includes the following steps: S50, apply a second gas pressure not lower than the nominal bubble point pressure of the filter element to the inside of the filter element to be tested, and execute step S3 to obtain a second reference image; S51, use panoramic images obtained by a plurality of qualified filter elements under the second gas pressure and processed by step S3 as training panoramic images, crop the training panoramic images into a plurality of training images of a standard size according to the size of the detection partition, and manually label the bubble regions in the training images to generate corresponding binary mask labels; S52, use a U-Net convolutional neural network with an encoder-decoder architecture as a semantic segmentation model, input the training images, and use the corresponding binary mask labels as training targets, use a combined loss function to optimize network parameters, and train the bubble segmentation model through a back propagation algorithm; S53, pre-process the images corresponding to each detection partition in the second reference image to obtain a second partition image, input the second partition image into the trained bubble segmentation model, and obtain a pixel-level probability map; S54, binarize the pixel-level probability map to obtain a final bubble region segmentation mask, and count the total number of pixels identified as bubbles in the segmentation mask corresponding to each second partition image to obtain a second total bubble area value of the detection partition. If the second total bubble area value in any detection partition is lower than the second area threshold, it is determined that the filter element to be tested has a clogging defect or insufficient opening rate in the detection partition.

8. The visual detection-based test method for filterability of a filter cartridge according to claim 7, characterized by, The combination loss function L is composed of a cross-entropy loss function L ce and a Dice loss function L Dice and is weighted and summed, and the calculation formula is wherein α、β is a weight coefficient, and α+β = 1, and an Adam optimizer is used to perform a back propagation process, and the weight parameters of the semantic segmentation model are iteratively optimized according to the partial derivatives calculated by the combination loss function L from the network prediction results and the real labels, until the semantic segmentation model converges.

9. The visual inspection-based test method for filterability of a filter cartridge according to claim 1, characterized by, The first gas pressure is 0.95 times or less of a nominal bubble point pressure of the filter element, the second gas pressure is set to 1.05 times or more of the nominal bubble point pressure of the filter element, and the nominal bubble point pressure of the filter element wherein D 1 is a nominal accuracy of the filter element, θ is a contact angle , η is a hole shape correction coefficient.

10. A visual inspection based filter element filtration test system for implementing the method of any one of claims 1 to 9, characterized by comprise: a transparent cylinder for containing a transparent detection liquid and accommodating the filter element to be tested; a backlight source arranged on one side of the transparent cylinder to provide uniform backlight illumination for the surface of the filter element; a gas pressure control unit connected to the filter element to be tested for introducing gas into the filter element to be tested and accurately controlling the first and second gas pressures; an image acquisition module comprising a plurality of industrial cameras arranged side by side; an image processing module connected to the image acquisition module, comprising: an image preprocessing unit for correcting the distortion of the images captured by the cameras based on a set of calibration parameters of the system and performing image stitching to generate a panoramic image of the bubble generation area, and dividing the panoramic image into a plurality of independent detection partitions to output a set of images of each detection partition; a bubble point detection unit for identifying single bubble regions with a pixel area greater than a single bubble area threshold after receiving the first reference image, and outputting a line chart containing the first total bubble area value in each detection partition; a uniformity detection unit for identifying after receiving the second reference image, calculating the second total bubble area value of each detection partition, and outputting a line chart containing the second total bubble area value in each detection partition.

Images