An image processing system and method for DR inspection data of circumferential welds

Abstract

This invention relates to the field of computer image processing technology and discloses an image processing system and method for DR inspection data of circumferential welds. By acquiring a multi-frame DR inspection data image sequence, candidate points of minute gray-level anomalies are extracted. Static feature vectors containing morphological features and gradient features are extracted from each candidate point. Temporal thermal evolution differential analysis is performed in the multi-frame image sequence to calculate the temporal correlation coefficient between the candidate points and the surrounding background area, generating temporal thermal evolution differential features. The static feature vectors and temporal thermal evolution differential features are jointly calculated to obtain a comprehensive confidence score, thereby realizing the differential identification of surface attachments and internal micro-defects.

Description

Technical Field

[0001] This invention relates to the field of computer image processing technology, and more specifically, to an image processing system and method for DR detection data of circumferential welds. Background Technology

[0002] In the field of pipeline engineering, circumferential welds are critical joints connecting pipe sections, and their quality directly affects the safe operation of the pipeline system. Digital radiographic testing (DR) technology uses X-rays to penetrate the weld area and image it, acquiring DR inspection data images that reflect the internal structure of the weld, making it an important means of non-destructive testing.

[0003] In DR inspection data images, there are two types of targets that require attention: one is the tiny spatter generated during the welding process, which adheres to the weld surface and is considered surface deposits; the other is the microcrack initiation points that may exist inside the weld or in stress concentration areas such as weld toes and weld roots, which are considered internal microdefects. Both types of targets appear as gray-scale anomaly areas of a few pixels in DR inspection data images.

[0004] Existing image processing methods for DR inspection data of circumferential welds mainly rely on the grayscale or shape features of a single frame image for target identification. However, when spatter is small and irregular in shape, its morphological features (circularity, area) and the gradient concentration features of microcrack initiation points highly overlap. Processing methods based on single-frame static image features struggle to effectively distinguish between these two types of targets, leading to false alarms caused by spatter being misclassified as microcracks, or missed detections caused by microcracks being filtered out as spatter. Therefore, a DR inspection data image processing method capable of effective differential identification based on differences in physical properties is needed. Summary of the Invention

[0005] This invention provides an image processing system and method for DR inspection data of circumferential welds, solving the technical problem in related technologies where similar features in a single frame of static image lead to confusion between surface attachments and internal micro-defects.

[0006] This invention provides an image processing method for DR inspection data of circumferential welds, including:

[0007] A multi-frame image sequence of DR inspection data of circumferential weld is obtained, and candidate points are extracted from each frame to generate a set of candidate points for small gray-scale anomalies.

[0008] Extract static feature vectors for each candidate point, wherein the static feature vectors include morphological features and gradient features;

[0009] Track the position of each candidate point in a multi-frame image sequence, extract the gray value temporal change curve of each candidate point, calculate the temporal correlation coefficient between each candidate point and the surrounding background area, and generate temporal thermal evolution differential features.

[0010] The static feature vector is combined with the time-series thermal evolution differential features to calculate the comprehensive confidence of each candidate point in belonging to the surface attachment class and the internal micro-defect class, and the classification results are output.

[0011] The calculation method of the temporal correlation coefficient is as follows: calculate the deviation sequence of the gray value of the candidate point and its temporal mean, and the deviation sequence of the gray value of the surrounding background area and its temporal mean, respectively. Multiply the corresponding positions of the two deviation sequences and sum them, and then divide by the square root of the product of the sum of the squares of the two deviation sequences for normalization.

[0012] Furthermore, the process of extracting candidate points for each frame of image includes: extracting candidate bright spot regions using morphological top-hat transformation, calculating multi-scale gradient magnitudes in the stress concentration regions of weld toe and weld root, merging the position coordinates of the candidate bright spot regions with the position coordinates of pixels whose gradient magnitudes exceed a preset gradient threshold to remove duplicates, and generating a set of candidate points.

[0013] Furthermore, the static feature vector includes: gray-level concavity depth, concavity area, circularity, gradient energy concentration, and gradient direction consistency index within a small neighborhood.

[0014] Furthermore, the gray-level concavity depth is the difference between the gray-level value of the candidate point and the mean value of the surrounding background; the gradient energy concentration is the ratio of the sum of squares of gradient magnitudes in the neighborhood of the candidate point to the area of ​​the neighborhood; and the consistency index is the ratio between the magnitude of the gradient vector in the neighborhood of the candidate point and the sum of gradient magnitudes.

[0015] Furthermore, the surrounding background area is defined as an annular area centered on the candidate point. Due to the difference in thermal inertia between the surface attachments and the weld body, the grayscale changes of the surface attachments show a low correlation with the surrounding background area. Due to the consistency of thermal conduction with the surrounding material, the grayscale changes of the internal micro-defects show a high correlation with the surrounding background area.

[0016] Furthermore, the calculation of the overall confidence score includes: normalizing the static feature vector, and based on the normalized static feature vector and the time-series thermal evolution differential features, calculating the confidence score of each candidate point belonging to the surface attachment class and the confidence score of belonging to the internal micro-defect class.

[0017] Furthermore, when the roundness of a candidate point is higher than the roundness threshold and the temporal correlation coefficient is lower than the low correlation threshold, the inverse transformation values ​​of roundness and temporal correlation coefficient are weighted and summed to obtain the confidence score of the surface attachment class; when the consistency index of a candidate point is higher than the consistency threshold and the temporal correlation coefficient is higher than the high correlation threshold, the forward transformation values ​​of consistency index and temporal correlation coefficient are weighted and summed to obtain the confidence score of the internal micro-defect class.

[0018] Furthermore, the output classification processing results include: setting an annular buffer band for the high-confidence area of ​​surface attachments and using median filtering for noise suppression processing; sorting internal micro-defects by risk level according to confidence level; and outputting a list of surface attachment locations and a list of internal micro-defect locations, along with their corresponding confidence levels and risk levels.

[0019] Furthermore, the risk level is divided into three levels: high risk, medium risk, and low risk, based on the confidence level of the internal micro-defects.

[0020] This invention provides an image processing system for DR inspection data of circumferential welds, comprising:

[0021] The image acquisition module is used to acquire multi-frame image sequences of DR inspection data for circumferential welds;

[0022] The candidate point extraction module is used to extract candidate points from each frame of the image and generate a set of candidate points with minor gray-level anomalies.

[0023] The static feature extraction module is used to extract static feature vectors containing morphological features and gradient features from each candidate point;

[0024] The temporal analysis module is used to track the positions of candidate points in a multi-frame image sequence and calculate the temporal correlation coefficient to generate temporal thermal evolution differential features.

[0025] The classification processing module is used to jointly calculate the comprehensive confidence score by combining the static feature vector with the time-series thermal evolution differential features and output the classification processing result.

[0026] The beneficial effects of this invention are as follows:

[0027] This invention employs a combination of morphological top-hat transformation and multi-scale gradient analysis to extract candidate points from DR inspection data images, enabling the simultaneous detection of both bright spot-type and gradient-concentrated micro-grayscale anomaly regions. By introducing temporal thermal evolution differential analysis to process DR inspection data, it leverages the physical property of the difference in thermal inertia between surface attachments (splashes) and the weld body. The temperature response of surface attachments under continuous X-ray irradiation is asynchronous with the weld substrate, resulting in low correlation between their temporal grayscale changes and the surrounding background area. In contrast, internal micro-defects (micro-crack initiation points) exhibit high correlation with the surrounding background area due to their consistent thermal conduction with the surrounding material, thus physically distinguishing surface attachments from internal micro-defects. By combining static feature vectors with temporal thermal evolution differential features to calculate the comprehensive confidence level, this invention overcomes the technical problem of confusion between splashes and micro-crack initiation points caused by similar features in single-frame static images, achieving reliable differential identification and classification of surface attachments and micro-defects in the DR inspection data image processing of circumferential welds. Attached Figure Description

[0028] Figure 1 This is a flowchart of the image processing method for DR detection data of circumferential welds according to the present invention;

[0029] Figure 2 This is a dual Y-axis broken line graph showing the temporal grayscale variation of candidate points in this invention;

[0030] Figure 3 This is a static feature comparison grouped bar chart of the present invention;

[0031] Figure 4 This is a scatter plot of the time-series correlation coefficient and grayscale fluctuation amplitude of the present invention;

[0032] Figure 5 This is a scatter plot showing the spatial distribution of candidate points in this invention;

[0033] Figure 6 This is a heat map showing the confidence distribution of the splashes according to the present invention. Detailed Implementation

[0034] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0035] At least one embodiment of the present invention discloses an image processing method for DR detection data of circumferential welds, such as... Figure 1 As shown, it includes the following steps:

[0036] Step 1: Obtain a multi-frame image sequence of DR inspection data for circumferential welds, extract candidate points from each frame image, and generate a set of candidate points for minor grayscale anomalies.

[0037] Specifically, a sequence of continuously acquired multi-frame DR detection data images is obtained from the DR detection system. , ,..., },in This is the first frame image. This is the second frame image. For the first Frame image, This refers to the number of image frames. For each image frame... Morphological top-hat transformation is used to extract candidate bright spot regions. Simultaneously, multi-scale gradient amplitudes are calculated in high-stress-concentration areas such as weld toes and weld roots, and a comprehensive set of candidate points for minute gray-level anomalies, P={ , ,..., },in As the first candidate point, As the second candidate point, For the first 1 candidate point This represents the total number of candidate points.

[0038] Furthermore, the specific method for generating the candidate point set is as follows: the position coordinates of the bright spot candidate region extracted by the morphological top-hat transformation are merged and deduplicated with the position coordinates of the pixels whose multi-scale gradient magnitude exceeds the preset gradient threshold to form a unified candidate point set, wherein the preset gradient threshold is determined according to the statistical characteristics of the overall gradient distribution of the image.

[0039] It should be noted that the aforementioned morphological top-hat transform involves performing an opening operation on the image using a structuring element of a preset size, followed by subtraction with the original image. This is used to extract bright areas smaller than the structuring element. Furthermore, the input to this morphological top-hat transform is a single-frame DR detection data image. The output is a set containing the coordinates of the candidate bright spot regions.

[0040] Furthermore, the size range of the aforementioned pre-defined structural elements is from 3×3 pixels to 7×7 pixels, determined according to the typical size of the target to be detected, to ensure that small grayscale anomaly areas can be effectively extracted.

[0041] The multi-scale gradient magnitude calculation described above involves calculating the gradient magnitude of each pixel using Sobel operators at different scales and then taking the maximum value. Furthermore, the input for this multi-scale gradient magnitude calculation is a single-frame DR detection data image. The location information of stress concentration high-incidence areas such as weld toe and weld root is output as the gradient magnitude value of each pixel.

[0042] Furthermore, the scale range of the Sobel operators mentioned above is from 3×3 to 9×9. Multi-scale computation can simultaneously detect gradient concentration features at different scales.

[0043] Step 2: Extract static feature vectors from each candidate point to generate static feature vectors that include morphological features and gradient features.

[0044] Specifically, for the candidate point set Candidate points in Extract the following static features: grayscale depression depth Area of ​​the depression Circularity Gradient energy concentration And the consistency index of gradient direction in a small neighborhood Generate static feature vectors. .

[0045] Among them, grayscale depression depth The difference between the gray value of the candidate point and the mean of the surrounding background; gradient energy concentration. The ratio of the sum of squared gradient magnitudes in the neighborhood of a candidate point to the area of ​​that neighborhood; directional consistency index. It is the ratio of the magnitude of the gradient direction vector in the neighborhood of the candidate point to the sum of the gradient magnitudes. The closer its value is to 1, the more consistent the gradient directions are.

[0046] Furthermore, the aforementioned surrounding background region is defined as a ring-shaped region centered on the candidate point with a radius of 5 to 10 pixels, the specific radius being determined based on the size of the candidate point. The aforementioned neighborhood is defined as a rectangular region of 3×3 to 5×5 pixels centered on the candidate point. The aforementioned connected regions are obtained by performing region growing or connected component analysis on the candidate points, and the 8-neighborhood connectivity criterion is used to group adjacent pixels with gray values ​​lower than the surrounding background into the same connected region.

[0047] Furthermore, the aforementioned grayscale depression depth A positive value indicates that the candidate point exhibits a concave feature with reduced grayscale compared to the surrounding background. The above gradient energy concentration... The value of is a positive real number greater than 0; a larger value indicates a more concentrated gradient energy. The above-mentioned directional consistency index... The value range of is [0,1]. When When the gradient is close to 1, it indicates that the gradient directions are highly consistent. When the value is close to 0, it indicates that the gradient direction is dispersed.

[0048] Furthermore, the aforementioned roundness The value range is (0,1], when the connected region is a perfect circle. The more irregular the shape, the better. The smaller.

[0049] Figure 3 Compare three key static features of five typical candidate points (circularity, orientation consistency index, and normalized gradient energy concentration).

[0050] Step 3: Track the position of each candidate point in the multi-frame DR detection data image sequence, perform temporal thermal evolution differential analysis on each candidate point, and generate temporal thermal evolution differential features.

[0051] Specifically, in a multi-frame image sequence { , ,..., Track each candidate point in} Extract the temporal variation curve of its grayscale value at the corresponding position. ={ , ,..., },in Candidate points The grayscale value in the first frame, The grayscale value of the second frame. For the first The grayscale value of the frame.

[0052] Furthermore, the aforementioned candidate point tracking employs a direct correspondence method based on position invariance. This assumes that the position coordinates of each candidate point remain unchanged across a series of consecutively acquired image frames, and extracts the grayscale value of the corresponding position using the same pixel coordinates in each frame. The surrounding background region is defined in the temporal analysis as a ring-shaped region with a radius of 8 to 15 pixels centered on the candidate point, and its position remains unchanged across all image frames.

[0053] Calculate the temporal correlation coefficient between each candidate point and the surrounding background region. :

[0054]

[0055] in, Candidate points In the The grayscale value of the frame, Its time-series mean; Candidate points The surrounding background area in the first The average gray value of the frame, Its time series mean.

[0056] Furthermore, the above time series mean The calculation formula is , indicating that the candidate point is Average grayscale value in a frame image; temporal mean The calculation formula is This indicates that the surrounding background area is in The average grayscale value in a frame image.

[0057] Due to differences in physical properties, surface deposits (splashes) exhibit low correlation with the surrounding background area in terms of grayscale variation because of the difference in thermal inertia between them and the weld body. Values ​​close to 0 or negative; internal micro-defects (micro-crack initiation points) exhibit a high correlation between their grayscale changes and the surrounding background area because their thermal conductivity is consistent with that of the surrounding material. Approximately 1. Generation time-series thermal evolution difference characteristics .

[0058] Furthermore, the input to the aforementioned time-series thermal evolution differential analysis is a multi-frame DR detection data image sequence { , ,..., } and each candidate point The position coordinates in each frame are output as the temporal thermal evolution difference features of each candidate point. The time-series thermal evolution differential feature is used to characterize the difference in thermal response between candidate points and the surrounding background area under X-ray irradiation.

[0059] In this embodiment of the application, in order to extract more stable temporal features, the gray-level fluctuation amplitude of each candidate point can also be calculated. Then, after normalization, it is added to the time series feature vector.

[0060] Furthermore, the aforementioned grayscale fluctuation amplitude It is a positive real number, and its value range depends on the degree of gray value fluctuation of the candidate point in multiple frames of the image. The larger the value, the more drastic the gray value changes over time.

[0061] Figure 2 The temporal variation of grayscale values ​​of typical candidate points (p3 microcrack, p7 spatter) and their surrounding background areas is shown in a 12-frame image sequence.

[0062] Step 4: Combine the static feature vector with the time-series thermal evolution differential features, calculate the comprehensive confidence of each candidate point to belong to the spatter class and the microcrack initiation point class, and output the classification results.

[0063] Specifically, the static feature vector Differential characteristics of time-series thermal evolution The features are combined to form a comprehensive feature vector. First, the static feature vector is analyzed... Each feature component in the data is normalized to eliminate dimensional differences between different feature components such as gray-level concavity depth, concavity area, and gradient energy concentration. Temporal thermal evolution differential features. Since it is a correlation coefficient, which is a dimensionless standardized value with a range of [-1, 1], no additional normalization is required. Candidate points are calculated based on the normalized comprehensive feature vector. Confidence level of classifying it as a splash Confidence level of classifying microcrack initiation points .

[0064] Furthermore, the input for the comprehensive confidence calculation is the normalized static feature vector. Time-series thermal evolution differential features The output is the confidence score of each candidate point belonging to the splash class. Confidence level of classifying microcrack initiation points .

[0065] Furthermore, the aforementioned confidence level and The values ​​range from [0,1] and are calculated by comprehensively judging the degree of agreement between each feature component and the typical features of the corresponding category. The closer the value is to 1, the higher the confidence level of belonging to the category.

[0066] Furthermore, the specific calculation method for the aforementioned confidence level is as follows: for the confidence level of splashes... When candidate points simultaneously satisfy circularity The correlation coefficient is higher than the preset threshold and the time-series correlation coefficient is higher than the preset threshold. When the value is below the low correlation threshold, calculate ,

[0067] in and The weighting coefficients and ; Confidence level for microcrack initiation points When candidate points simultaneously satisfy the directional consistency index The correlation coefficient is higher than the preset threshold and the time-series correlation coefficient is higher than the preset threshold. When the value is above the high correlation threshold, calculate ,

[0068] in and The weighting coefficients and Candidate points that do not meet the judgment criteria have their corresponding category confidence level set to 0.

[0069] Furthermore, the aforementioned weighting coefficients , , , The values ​​of all values ​​are in the range [0,1], and the constraints are satisfied. and In this embodiment of the application, the weighting coefficient can be set to... to , to The specific values ​​are determined based on the importance of static and temporal characteristics in practical applications.

[0070] Among them, the criteria for determining the type of splash include: roundness. Higher than a preset threshold, time-series correlation coefficient Below the low relevance threshold; the criteria for classifying microcrack initiation points include: directional consistency index. Higher than a preset threshold, time-series correlation coefficient Above the high correlation threshold.

[0071] Furthermore, the preset threshold range for circularity is 0.6 to 0.8, the preset threshold range for low correlation is -0.2 to 0.3, the preset threshold range for high correlation is 0.6 to 0.9, and the preset threshold range for directional consistency index is 0.7 to 0.9. Specific thresholds are selected within these ranges based on the feature distribution of the actual object being detected and the required detection accuracy.

[0072] A ring-shaped buffer zone is set for high-confidence areas of spatter, and median filtering is used for noise suppression. Microcrack initiation points are sorted by risk level according to confidence level. The output is a list of spatter locations after image processing. List of microcrack initiation points ,in The coordinates of the splash are the position of the debris. For the confidence level of splashes, These are the coordinates of the microcrack initiation point. The confidence level for microcrack initiation points. This corresponds to the risk level.

[0073] Furthermore, the width of the aforementioned annular buffer band is set to 2 to 5 pixels to create a transition area around the splatter area, and then a 3×3 or 5×5 median filter kernel is applied to the transition area for noise suppression.

[0074] Furthermore, the above risk levels Based on the confidence level of microcrack initiation points It is divided into three levels: high risk, medium risk, and low risk. A value greater than 0.7 is considered high risk. A value between 0.4 and 0.7 is considered medium risk. A value less than 0.4 is considered low risk.

[0075] In this embodiment, the method further includes the following steps: based on step 2, the static feature vector of each candidate point is compared with the feature threshold of the spatter to preliminarily screen high-probability spatter regions; the concentration of gradient energy at a very small scale is calculated for each candidate point to identify high-probability microcrack initiation regions with directional convergence characteristics. The above preliminary screening results can be used to narrow the calculation range of subsequent time series analysis.

[0076] Furthermore, the aforementioned splash feature thresholds include a roundness threshold and a recessed area threshold, wherein the roundness threshold ranges from 0.5 to 0.7, and the recessed area threshold is determined based on the typical size of the splash to be detected.

[0077] Figure 4 Display the temporal correlation coefficient (X-axis) and grayscale fluctuation amplitude (Y-axis) distribution of candidate points, with different colors indicating different classification results.

[0078] Figure 5 The spatial distribution of 35 candidate points in a 2048×512 pixel DR image is shown, with different colors and shapes indicating the classification results.

[0079] Figure 6 The confidence distribution of splashes is shown under different combinations of circularity and temporal correlation coefficients.

[0080] The DR inspection data image processing method of this embodiment uses morphological top-hat transformation and multi-scale gradient analysis to extract candidate points in the DR inspection data image, which can simultaneously detect small gray-scale abnormal areas of bright spot type and gradient concentration type.

[0081] By incorporating time-series thermal evolution differential analysis into the DR inspection data, the physical property of the difference in thermal inertia between surface deposits (splashes) and the weld body is utilized. The temperature response of surface deposits under continuous X-ray irradiation is asynchronous with the weld substrate, resulting in a low correlation between their grayscale temporal changes and the surrounding background area. Conversely, internal micro-defects (micro-crack initiation points) exhibit consistent thermal conduction with the surrounding material, leading to a high correlation between their grayscale temporal changes and the surrounding background area. Therefore, by calculating the temporal correlation coefficient between each candidate point and the surrounding background area, surface deposits and internal micro-defects can be physically distinguished.

[0082] By using a combination of static feature vectors and time-series thermal evolution differential features to calculate the comprehensive confidence level, the factor of confusion between spatter and microcrack initiation points caused by the similarity of features in a single frame static image is overcome. At the same time, by combining targeted noise suppression processing of the spatter area and risk level ranking of microcrack initiation points, reliable differential identification and classification processing of surface attachments and microdefects is achieved in the image processing of DR inspection data of circumferential welds.

[0083] Application Scenario: In a long-distance natural gas pipeline project, DR (radiofrequency) inspection was performed on the circumferential weld of a DN1200×12mm X70 grade spiral submerged arc welded pipeline. A portable X-ray machine was used with a flat panel detector to continuously acquire images of the circumferential weld numbered WJ-2024-1138, located at pipeline mileage marker K328+456. During the inspection, the X-ray source parameters were set to 220kV pipe voltage and 5mA pipe current, with the detector continuously acquiring images at 8-second intervals, resulting in 12 frames of DR inspection data images.

[0084] Example of the implementation process of step 1: The basic parameters of the multi-frame image sequence acquired by the DR detection system are shown in the table below.

[0085] Table 1. Acquisition parameters for multi-frame image sequences:

[0086]

[0087] For the first frame image A morphological top-hat transformation was performed using a 5×5 pixel circular structuring element, extracting 23 candidate bright spot locations in the weld toe region. Simultaneously, Sobel operators at three scales (3×3, 5×5, and 7×7) were used to calculate the gradient magnitude in the stress concentration region of the weld root. Using the 85th percentile of the overall image gradient magnitude distribution as the preset gradient threshold (18.7), 19 candidate gradient concentration locations were extracted. After coordinate merging and deduplication, a final set containing 35 candidate points was generated. .

[0088] Example of the implementation process of step 2: Extract static feature vectors for typical candidate points in the candidate point set. The feature data of some candidate points are shown in the table below.

[0089] Table 2 Static feature extraction results of typical candidate points:

[0090]

[0091] Among them, candidate points The surrounding background region is defined as a ring-shaped area with a radius of 8 pixels. The average gray level of this region is 178.6, and the gray level of the candidate point center is 146.2. Therefore, the gray level concavity depth is... Candidate points The area of ​​the connected region is 9 pixels and the perimeter is 10.83 pixels. The circularity is calculated. At candidate points Within a 5×5 pixel neighborhood, the sum of squared gradient magnitudes is 3565, therefore the gradient energy concentration is... .

[0092] Example of Step 3 implementation: Track the positions of candidate points in a 12-frame image sequence and extract the temporal grayscale changes of each candidate point. (Using candidate points...) For example, in time series analysis, the surrounding background area is defined as a ring-shaped area with a radius of 10 pixels, and the grayscale value changes of the area in each frame are shown in the table below.

[0093] Table 3 Candidate Points Temporal grayscale values ​​in multi-frame images:

[0094]

[0095] For candidate points Calculate the time series mean Background region time series mean The time-series correlation coefficient was calculated using the formula. This indicates that the grayscale changes of the candidate points are highly synchronized with the background area.

[0096] The temporal correlation coefficients of all candidate points were calculated, and the results are shown in the table below.

[0097] Table 4: Calculation results of temporal correlation coefficients for candidate points:

[0098]

[0099] Among them, candidate points and The time-series correlation coefficients were 0.142 and -0.073, respectively, significantly lower than the high correlation threshold, exhibiting the thermal response characteristics of surface attachments; while the candidate points , and The temporal correlation coefficients all exceeded 0.98, exhibiting synchronous thermal conduction characteristics of internal micro-defects.

[0100] Example of step 4 implementation: Combine the static feature vector with the time-series thermal evolution differential features, and set weighting coefficients. , The preset thresholds for roundness are 0.65, low correlation, and high correlation, and the preset threshold for directional consistency index is 0.75. The comprehensive confidence score for each candidate point is calculated and classified. The final results are shown in the table below.

[0101] Table 5: Final results of classification processing:

[0102]

[0103] Candidate points The circularity is 0.68 and the temporal correlation coefficient is 0.142, which meets the criteria for splash determination. Calculate the confidence level of splash:

[0104] This was determined to be splashing debris. Candidate point. The directional consistency index is 0.85 and the temporal correlation coefficient is 0.984, which meets the criteria for determining microcrack initiation points. The confidence level of the microcrack is calculated as follows:

[0105] ,because It was determined to be a high-risk microcrack initiation point.

[0106] A 3-pixel-wide annular buffer band was set for the two identified splash locations, and noise suppression was performed using a 5×5 median filter kernel.

[0107] Sort the two high-risk microcrack initiation points in descending order of confidence level and output a list of spatter locations. List of microcrack initiation points This provides a basis for classification and processing in subsequent weld quality assessment.

[0108] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. An image processing method for DR inspection data of circumferential welds, characterized in that, Includes the following steps: A multi-frame image sequence of DR inspection data of circumferential weld is obtained, and candidate points are extracted from each frame to generate a set of candidate points for small gray-scale anomalies. Extract static feature vectors for each candidate point, wherein the static feature vectors include morphological features and gradient features; Track the position of each candidate point in a multi-frame image sequence, extract the gray value temporal change curve of each candidate point, calculate the temporal correlation coefficient between each candidate point and the surrounding background area, and generate temporal thermal evolution differential features. The static feature vector is combined with the time-series thermal evolution differential features to calculate the comprehensive confidence of each candidate point in belonging to the surface attachment class and the internal micro-defect class, and the classification results are output. The static feature vector includes: gray-level depression depth, depression area, circularity, gradient energy concentration, and gradient direction consistency index in a small neighborhood. The gray-level indentation depth is the difference between the gray value of the candidate point and the mean value of the surrounding background; the gradient energy concentration is the ratio of the sum of squares of the gradient magnitudes in the neighborhood of the candidate point to the area of ​​the neighborhood. The consistency index is the ratio of the magnitude of the sum of the gradient vectors of all pixels in the neighborhood of a candidate point to the sum of the gradient magnitudes of all pixels in the neighborhood. Confidence level for classifying it as a surface attachment When candidate points simultaneously satisfy circularity The correlation coefficient is higher than the preset threshold and the time-series correlation coefficient is higher than the preset threshold. When the value is below the low correlation threshold, calculate , in and The weighting coefficients and ; The overall confidence level for classifying internal minor defects When candidate points simultaneously satisfy the directional consistency index The correlation coefficient is higher than the preset threshold and the time-series correlation coefficient is higher than the preset threshold. When the value is above the high correlation threshold, calculate , in and The weighting coefficients and ; The calculation method of the temporal correlation coefficient is as follows: calculate the deviation sequence of the gray value of the candidate point and its temporal mean, and the deviation sequence of the gray value of the surrounding background area and its temporal mean, respectively. Multiply the corresponding positions of the two deviation sequences and sum them, and then divide by the square root of the product of the sum of the squares of the two deviation sequences for normalization.

2. The method according to claim 1, characterized in that, The process of extracting candidate points for each frame of image includes: using morphological top-hat transformation to extract candidate bright spot regions, while calculating multi-scale gradient magnitudes in the stress concentration regions of weld toe and weld root, merging the position coordinates of the candidate bright spot regions with the position coordinates of pixels whose gradient magnitudes exceed a preset gradient threshold to remove duplicates, and generating a set of candidate points.

3. The method according to claim 1, characterized in that, The surrounding background area is defined as an annular area centered on the candidate point. Due to the difference in thermal inertia between the surface attachments and the weld body, the grayscale changes of the surface attachments show a low correlation with the surrounding background area. Due to the consistency of thermal conduction with the surrounding material, the grayscale changes of the internal micro-defects show a high correlation with the surrounding background area.

4. The method according to claim 1, characterized in that, The calculation of the overall confidence score includes: normalizing the static feature vector, and based on the normalized static feature vector and the time-series thermal evolution differential features, calculating the confidence score of each candidate point belonging to the surface attachment class and the confidence score of belonging to the internal micro-defect class.

5. The method according to claim 1, characterized in that, The output classification processing results include: setting an annular buffer band for high-confidence areas of surface attachments and using median filtering for noise suppression; sorting internal micro-defects by risk level according to confidence level; and outputting a list of surface attachment locations and a list of internal micro-defect locations, along with their corresponding confidence levels and risk levels.

6. The method according to claim 5, characterized in that, The risk levels are divided into three categories: high risk, medium risk, and low risk, based on the confidence level of internal micro-defects.

7. An image processing system for DR inspection data of circumferential welds, used to perform the method according to any one of claims 1-6, characterized in that, include: The image acquisition module is used to acquire multi-frame image sequences of DR inspection data for circumferential welds; The candidate point extraction module is used to extract candidate points from each frame of the image and generate a set of candidate points with minor gray-level anomalies. The static feature extraction module is used to extract static feature vectors containing morphological features and gradient features from each candidate point; The temporal analysis module is used to track the positions of candidate points in a multi-frame image sequence and calculate the temporal correlation coefficient to generate temporal thermal evolution differential features. The classification processing module is used to jointly calculate the comprehensive confidence score by combining the static feature vector with the time-series thermal evolution differential features and output the classification processing result.

Images