Pipeline wall defect detection method and system based on image recognition

Abstract

The invention relates to the field of image processing, in particular to a pipeline wall defect detection method and system based on image recognition, and the method comprises the steps: obtaining a preprocessed gray-scale map of a pipeline wall; calculating a segmentation demand degree of each pixel point in the grey-scale map, dividing the grey-scale map into a plurality of sub-image blocks by using the segmentation demand degrees, and performing histogram equalization on each sub-image block to obtain an enhanced grey-scale map; and inputting the enhanced grey-scale map into a pre-trained neural network model, and outputting a pipeline wall defect detection result. According to the method, the grey-scale map is dynamically divided according to the segmentation demand degree of the pixel points, and the image can be divided into the sub-image blocks which better conform to the defect features, so that the problem of defect feature loss or misjudgment is avoided, and the defect detection accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing. More specifically, the present invention relates to a method and system for detecting pipeline wall defects based on image recognition. Background Art

[0002] As an important infrastructure for national energy transmission, the safety of oil pipelines is directly related to energy supply and environmental protection. During long-term operation, pipelines are extremely prone to structural defects such as wall corrosion and stress cracks due to exposure to complex and variable environmental conditions. If these defects cannot be discovered and repaired in time, it may lead to serious leakage accidents, causing huge economic losses and ecological disasters.

[0003] Traditional pipeline detection mainly relies on manual inspections. This method is not only inefficient but also easily affected by subjective judgments, making it difficult to achieve comprehensive and accurate defect detection. In recent years, intelligent detection methods based on image processing technology have been widely used. By analyzing high-definition images taken outside the pipeline, this method can automatically identify and locate potential defect positions, accurately judge the type of defects, and significantly improve the accuracy and efficiency of detection.

[0004] CLAHE (Contrast Limited Adaptive Histogram Equalization) is an algorithm commonly used in image processing. This algorithm enhances local contrast by dividing the image into multiple small blocks and performing histogram equalization on each small block separately.

[0005] However, oil pipeline images usually have complex backgrounds and diverse defect types. These defects vary in size, shape, gray-scale distribution, etc., making it difficult for the algorithm to accurately identify and enhance all defects. In addition, the unevenness of the lighting conditions may also cause large differences in the lighting intensity of different regions in the image, further increasing the difficulty of defect detection. Summary of the Invention

[0006] To solve the above technical problem of difficultly and accurately identifying oil pipeline wall defects, the present invention provides solutions in the following aspects.

[0007] In a first aspect, a method for detecting pipeline wall defects based on image recognition includes: Obtaining a grayscale image of the pipeline wall after preprocessing; calculating the segmentation requirement degree of each pixel point in the grayscale image, dividing the grayscale image into multiple sub-image blocks using the segmentation requirement degree, performing histogram equalization on each sub-image block to obtain an enhanced grayscale image; inputting the enhanced grayscale image into a pre-trained neural network model, and outputting the pipeline wall defect detection result; The segmentation requirement degree is: ; where is the The segmentation requirement degree of a pixel is the neighborhood detail degree of the th pixel, and the neighborhood rule index of the

[0008] th pixel; the neighborhood rule index is used to measure the degree to which a pixel belongs to a certain defect category.

[0009] Preferably, the neighborhood detail degree of the th pixel satisfies the relational expression: ; in the formula, is the neighborhood detail degree of the th pixel, is the Euclidean distance between the th pixel and the th edge pixel in its neighborhood, represents the total number of edge pixels in the neighborhood of the th pixel, represents the exponential function with the natural number e as the base, is the total number of edge pixels included in the neighborhood of the th edge pixel in the grayscale image, is the Euclidean distance between the th and the th edge pixels in the neighborhood of the th pixel, represents the maximum value, represents normalization.

[0010] By calculating the relationship between a pixel and the edge pixels in its neighborhood, the edge information in the image can be accurately captured. At the same time, the distribution of edge pixels in the neighborhood is also considered. For example, if the edge pixels in a certain area are dense, it indicates that this area may contain rich textures or complex structures, resulting in a relatively high calculated neighborhood detail degree, thus highlighting these local features in subsequent processing and helping to better distinguish different regions.

[0011] Preferably, the The neighborhood rule index of a pixel satisfies the following relationship: ; where is the neighborhood rule index of the -th pixel, is the -th direction vector on the -th neighborhood edge of the -th pixel, is the -th direction vector on the -th neighborhood edge of the -th pixel, is the function for taking the vector modulus length, is the arccosine function, is the -th total number of direction vectors on the -th neighborhood edge of the is the minimum value function, represents the exponential function with the natural number e as the base.

[0012] By calculating the angles between the direction vectors on the neighborhood edge of the pixel (calculated by the arccosine function) and taking the minimum value of these angles, and finally obtaining a neighborhood rule index through the exponential function, this calculation method can reflect the geometric regularity of the pixel neighborhood. If the angles between the direction vectors on the neighborhood edge are small, it indicates that the shape of the neighborhood is relatively regular; conversely, if the angles are large, it indicates that the shape of the neighborhood is relatively complex. By analyzing the neighborhood rule index of each pixel, the local structure of the image can be better understood, thereby improving the accuracy and efficiency of image processing.

[0013] Preferably, for a single pixel, if there is any pixel on the edge within the neighborhood of the pixel, then this edge is defined as the neighborhood edge of the pixel; the pixels on the neighborhood edge are sorted by the boundary tracing method, the spatial vectors of two adjacent pixels in ordinal number are obtained, and marked as direction vectors.

[0014] In the edge-based image segmentation method, the boundaries of different regions can be determined according to the changes in the direction vectors. For example, when the direction vectors change drastically, it may mean the junction of different regions. By analyzing and processing the direction vectors, the image can be divided into different regions, and each region has similar characteristics.

[0015] Preferably, the dividing the grayscale image into multiple sub-blocks by using the segmentation requirement degree includes: The grayscale image is segmented into multiple sub - image blocks using a quadtree. Among them, the average segmentation requirement degree of all pixel points in each sub - image block is calculated. If the average segmentation requirement degree of the sub - image block is higher than a preset threshold, and the number of segmentation times of this sub - image block is less than or equal to the preset segmentation threshold, then this sub - image block is further divided into four equal parts.

[0016] By calculating the average segmentation requirement degree of each sub - image block, this algorithm can adaptively segment according to the complexity of the image content. For complex regions (high segmentation requirement degree), further subdivision is carried out to capture more details; for simple regions (low segmentation requirement degree), larger blocks are maintained to avoid over - segmentation.

[0017] Preferably, the histogram equalization adopts contrast - limited adaptive histogram equalization.

[0018] Preferably, the pipeline wall defect detection results include the position, size and type of the defect.

[0019] Preferably, for the neighborhood detail degree of the th pixel point, the relational expression is: where is the neighborhood detail degree of the th pixel point, is the Euclidean distance between the th pixel point and the th edge pixel point in its neighborhood, represents the total number of edge pixel points in the neighborhood of the th pixel point, is the exponential function with the natural number e as the base, is the gradient magnitude of the th edge pixel point, is the gradient direction of the th pixel point, is the gradient direction of the th pixel point, is the average distance between the th pixel point and all edge pixel points in its neighborhood,

[0020] By comprehensively considering the gradient magnitude of the th edge pixel point in the neighborhood, the difference in gradient direction from the current pixel point, and the average distance, the weighted contribution of the th edge pixel point to the neighborhood detail degree of the current pixel point is calculated, so that the neighborhood detail degree can better reflect the detail information around the pixel point.

[0021] Preferably, for the For a pixel, starting from any pixel in its preset neighborhood, traverse the other pixels in the neighborhood in a set direction in sequence, and assign a direction code to each traversed pixel. Compare the differences between adjacent direction codes, count the total number of direction changes, obtain the total number of direction changes, and use the ratio of the total number of direction changes to the number of direction codes as the neighborhood rule index of the pixel.

[0022] In a second aspect, a pipeline wall defect detection system based on image recognition includes: a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, any one of the pipeline wall defect detection methods based on image recognition described above is implemented.

[0023] The beneficial effects of the present invention are: Dynamically divide the grayscale image according to the segmentation requirement degree of pixels, and the image can be divided into sub-image blocks that more conform to the defect characteristics. This adaptive block division method avoids the problems of defect feature loss or misjudgment that may be caused by the fixed block division method, thereby improving the accuracy of defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a flowchart of the method of steps S1 - S3 in a pipeline wall defect detection method based on image recognition according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention.

[0026] The application scenario of the present invention is: detecting defects in the pipeline wall of an oil pipeline.

[0027] Referring to Figure 1 , a pipeline wall defect detection method based on image recognition includes steps S1 - S3, specifically as follows: S1: Obtain the grayscale image of the preprocessed pipeline wall.

[0028] Select a suitable industrial camera according to the shooting requirements (such as field of view size, accuracy requirements, shooting speed, etc.), fix the industrial camera in a suitable position to ensure that a complete pipeline wall image can be captured.

[0029] Preprocess the collected image, such as denoising, cropping, etc., and convert the collected pipeline wall image into a grayscale image for analysis.

[0030] S2: Calculate the segmentation requirement degree of each pixel point in the grayscale image, divide the grayscale image into multiple sub-image blocks by using the segmentation requirement degree, and perform histogram equalization on each sub-image block to obtain the enhanced grayscale image.

[0031] Since defects (such as cracks and corrosion areas) on the oil pipeline wall are usually accompanied by obvious grayscale changes, first, the Canny edge detection algorithm is used to extract the edges in the image to obtain a binary image containing the image edges. The Canny algorithm can well suppress non-maximum values, thus accurately extracting the edges.

[0032] In one embodiment, first, obtain the Euclidean distance between the coordinates of any two pixel points in the grayscale image, and take each pixel point as the center to obtain the neighborhood of the corresponding pixel point, such as 3×3, 5×5, 7×7, etc. In the embodiment of the present invention, 5×5 is selected as the neighborhood of the pixel point.

[0033] Then, mark the pixel points belonging to the edges in the grayscale image as edge pixel points. Edges are usually areas in the image where the grayscale values change drastically, and can be identified through edge detection algorithms (such as Canny edge detection).

[0034] Finally, by considering the total number of edge pixel points in the neighborhood of each edge pixel point in the grayscale image and the distance between the edge pixel point and other edge pixel points in the neighborhood, calculate the non-noise degree of each edge pixel point. At the same time, also consider the distance between the edge pixel points in the neighborhood of the pixel points other than the edge pixel points and this pixel point, calculate the distance influence weight of each edge pixel point, multiply the non-noise degree and the distance influence weight of each edge pixel point, then sum up all these products and perform normalization to obtain the neighborhood detail degree of the pixel point.

[0035] The neighborhood detail degree of the above pixel points satisfies the following relationship:

[0036] In the formula, is the neighborhood detail degree of the th pixel point, is the th pixel point and the th edge pixel point in its neighborhood, represents the th pixel point, represents the exponential function with the natural number e as the base, is the total number of edge pixel points contained in the neighborhood of the th edge pixel point in the grayscale image, is the th pixel point, is the The Euclidean distance between adjacent edge pixels denotes the maximum value denotes normalization processing

[0037] If there are more edge pixels in the neighborhood of an edge pixel and its distance from other edge pixels is relatively close, then its degree of non-noise is relatively high; if the distance between the edge pixel and the central pixel is relatively small, then its influence weight is relatively high

[0038] The higher the neighborhood detail degree of a pixel, the more likely it is that the area where it is located contains details that need attention. This is because when there are more non-noise edge pixels in the neighborhood and these edge pixels are relatively closely distributed, this area is more likely to contain valid edge information

[0039] It should be noted that in the image detection of oil pipelines, cracks and corrosion areas on the pipeline wall are two important defects. They are manifested in the image as having a high neighborhood detail degree, that is, the gray values of the pixels in these areas differ greatly from those of the surrounding pixels. However, the processing requirements for cracks and corrosion areas are different

[0040] Cracks require more refined segmentation to ensure that the gray distribution of the cracks will not be diluted by the gray distribution of the background. In other words, the details of the cracks need to be clearly retained; the corrosion areas need to avoid over-segmentation to prevent inconsistent local enhancement, thereby generating inter-block artifacts (i.e., unnatural blocky areas appear in the image)

[0041] In order to identify which type of defect (crack or corrosion) a pixel specifically belongs to, it is necessary to obtain the neighborhood rule index of each pixel

[0042] In one embodiment, first, for a single pixel, if there is any pixel on the edge within the neighborhood of this pixel, then this edge is defined as the "neighborhood edge" of this pixel

[0043] Then, the pixels on these neighborhood edges are sorted by the boundary tracking method, the spatial vectors of two ordinally adjacent pixels are obtained, and are marked as "direction vectors". For example, if a pixel has 5 neighborhood edges, then it has 5 direction vector sequences

[0044] For each direction vector sequence, the included angle between each pair of adjacent direction vectors is calculated through the inverse cosine function. When the sum of the included angles of all ordinally adjacent two direction vectors on the neighborhood edge of a pixel is relatively small, it indicates that the possibility of a smooth edge existing near this pixel is relatively high; further calculate the neighborhood rule index of the pixel, that is, the relational expression is satisfied as:

[0045] In the formula, is the neighborhood rule index of the th pixel point, is the th pixel point's th direction vector on the th neighborhood edge, is the th pixel point's th direction vector on the th neighborhood edge, is the function to take the vector modulus length, is the arccosine function, is the th pixel point's total number of direction vectors on the th neighborhood edge, represents the exponential function with the natural number e as the base.

[0046] Among them, summing the included angles of all adjacent direction vectors of a certain neighborhood edge can reflect the tortuosity of the edge. Traverse all neighborhood edges of the current pixel, find the edge with the smallest sum of included angles (the smoothest edge), and map the sum of included angles to the range (0, 1]. The smaller the sum of included angles (the smoother the edge), the closer it is to 1; the larger the sum of included angles (the more tortuous the edge), the closer it is to 0.

[0047] The neighborhood rule index can effectively quantify the irregularity of cracks and the smoothness of corrosion by capturing the continuity of direction vectors and the cumulative effect of angle changes. Cracks have a low neighborhood rule index due to sudden direction changes, while corrosion has a high neighborhood rule index due to gradual direction changes, thus realizing the distinction between the two.

[0048] To sum up, in image processing, the fineness of segmentation is determined according to the features around pixel points (neighborhood detail degree and neighborhood rule index). For places with edges and possible cracks, more refined segmentation is required; for large areas of disordered defects, the segmentation should be moderate to avoid artifacts.

[0049] Therefore, calculate the segmentation requirement degree of pixel points according to the neighborhood detail degree and neighborhood rule index of the above pixel points, that is, the satisfaction relationship is:

[0050] In the formula, is the segmentation requirement degree of the th pixel point, is the neighborhood detail degree of the th pixel point, is the The neighborhood rule index of a pixel point.

[0051] Furthermore, according to the segmentation requirement degree of the th pixel point, the segmentation requirement degrees of all pixel points in the grayscale image can be calculated.

[0052] In another embodiment, the neighborhood detail degree of the above-mentioned th pixel point further satisfies the relational expression:

[0053] In the formula, is the neighborhood detail degree of the th pixel point, is the Euclidean distance between the th pixel point and the th edge pixel point in its neighborhood, represents the total number of edge pixel points in the neighborhood of the th pixel point, represents the exponential function with the natural number e as the base, is the gradient magnitude of the th edge pixel point, is the gradient direction of the th pixel point, is the gradient direction of the th pixel point, is the average distance between the th pixel point and all edge pixel points in its neighborhood, represents normalization processing.

[0054] Among them, by using the gradient magnitude and gradient direction information of edge pixel points, the strength and direction of the edge are comprehensively considered. A large gradient magnitude indicates a strong edge and may contribute more to the neighborhood detail degree; while the cosine term considers the difference between the gradient directions of the current pixel point and edge pixel points, and edges with similar directions have a greater impact on the detail degree of the current pixel point, which helps to more accurately describe the detail features and direction relationships of the edge. In addition, using for normalization processing avoids the excessive influence of the distance factor on the result.

[0055] In another embodiment, the calculation process of the neighborhood rule index of the th pixel point can also be: Set the neighborhood of the th pixel point as an eight-neighborhood, select any pixel point in its neighborhood as the starting point, then traverse the other pixel points in the neighborhood in a clockwise or counterclockwise direction, and assign a direction code to each pixel point. The value range of the direction code is from 0 to 7. Record the direction codes of each pixel point encountered during the traversal to form a chain code sequence. The chain code length is the number of codewords in the chain code.

[0056] Further, traverse the chain code direction codes within the neighborhood of the th pixel. Compare the differences between adjacent direction codes. If adjacent direction codes are different, record it as a direction change. Count the total number of direction changes within this neighborhood to obtain the total number of direction changes.

[0057] Further, use the ratio of the total number of the above-mentioned direction changes to the total length (the number of direction codes) of the chain code within the neighborhood of the th pixel as the neighborhood rule index of the th pixel.

[0058] The direction codes of the chain code reflect the direction changes from one pixel to another. By comparing the differences between adjacent direction codes, the number of direction changes within the neighborhood can be counted, thereby quantifying the regularity of the neighborhood. This quantification method can help distinguish regular structures (such as straight lines, rectangles) and complex structures (such as curves, irregular shapes).

[0059] Quadtree is a method of recursive partitioning used to gradually divide a large area (such as an image) into smaller sub-areas.

[0060] In one embodiment, a grayscale image is used as the target to be segmented. The grayscale image is divided into four sub-image blocks of equal size. Among them, if the number of rows or columns of the image is odd, there will be one extra row or column on one side during segmentation to ensure that there are still four sub-blocks after segmentation. After the initial segmentation, the segmentation times of each sub-image block are set to 1.

[0061] For each sub-image block, calculate the average segmentation demand degree of all pixels in each sub-image block. If the average segmentation demand degree of the sub-image block is higher than the threshold (set to 0.1 according to experience), and the current segmentation times is less than or equal to the segmentation threshold (exemplarily, set the segmentation threshold to 10 times), then this sub-image block will be further divided into four equal parts. After each segmentation, the segmentation times of the sub-image block will be incremented by 1. When the segmentation times of the sub-image block reach or exceed 10 times, it will no longer be segmented, and it will also no longer be segmented when the average segmentation demand degree of the sub-image block is lower than the threshold.

[0062] Through the above recursive segmentation process, a series of sub-image blocks of different sizes are finally obtained.

[0063] Further, perform histogram equalization with limited contrast on each sub-block respectively, and recombine all the processed sub-image blocks into an enhanced grayscale image.

[0064] S3: Input the enhanced grayscale image into a pre-trained neural network model to output the detection result of the pipeline wall defect.

[0065] In one embodiment, the enhanced grayscale image obtained in S2 above is input into a pre-trained defect detection and classification model (this model is trained with a large amount of labeled data (including images of defective and non-defective pipe walls)). After analyzing the input image, the model will output the results of defect detection, such as whether there are defects, the location, size, and type of the defects, etc.

[0066] If a defect is detected, the maintenance personnel can locate the defect position according to the results and take corresponding repair measures. If no defect is detected, continuous monitoring or regular inspections can be carried out.

[0067] The system includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the method for detecting pipe wall defects based on image recognition according to the first aspect of the present invention.

[0068] The system also includes other components well-known to those skilled in the art, such as a communication bus and a communication interface. Their settings and functions are known in the art, so they will not be elaborated here.

[0069] It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent shall be subject to the appended claims.

Claims

1. A pipeline wall defect detection method based on image recognition, characterized in that, Including: Obtain the grayscale image of the preprocessed pipeline wall; Calculate the segmentation requirement degree of each pixel point in the grayscale image, divide the grayscale image into multiple sub-image blocks by using the segmentation requirement degree, perform histogram equalization on each sub-image block to obtain the enhanced grayscale image; input the enhanced grayscale image into the pre-trained neural network model, and output the pipeline wall defect detection result; The segmentation requirement degree is: ; wherein, is the segmentation requirement degree of the th pixel point, is the neighborhood detail degree of the th pixel point, is the neighborhood rule index of the th pixel point; the neighborhood rule index is used to measure the degree to which a pixel point belongs to a certain defect category.

2. The method for detecting pipeline wall defects based on image recognition according to claim 1, wherein, The neighborhood detail degree of the pixel point satisfies the relational expression as follows: ; where is the neighborhood detail degree of the th pixel point, is the Euclidean distance between the th pixel point and the th edge pixel point in its neighborhood, represents the total number of edge pixel points in the neighborhood of the th pixel point, represents the exponential function with the natural number e as the base, is the total number of edge pixel points contained in the neighborhood of the th edge pixel point in the grayscale image, is the Euclidean distance between the th and the th edge pixel points in the neighborhood of the th pixel point, represents the maximum value, represents normalization.

3. The method for detecting pipeline wall defects based on image recognition according to claim 2, wherein, The neighborhood rule index of the pixel satisfies the relational expression as follows: ; where, is the neighborhood rule index of the th pixel point, is the th th direction vector on the th neighborhood edge of the th pixel point, is the th th direction vector on the is the function for taking the vector modulus length, is the arccosine function, is the th total number of direction vectors on the th neighborhood edge of the represents the exponential function with the natural number e as the base.

4. The method for detecting pipeline wall defects based on image recognition according to claim 3, characterized in that, For a single pixel point, if there is any pixel point on the edge within the neighborhood of this pixel point, then this edge is defined as the neighborhood edge of this pixel point; sort the pixel points on the neighborhood edge by the boundary tracing method, obtain the spatial vectors of two pixel points with adjacent ordinals, and mark them as direction vectors.

5. The method for detecting pipeline wall defects based on image recognition according to claim 3, characterized in that The dividing the grayscale image into multiple sub-blocks by using the segmentation requirement degree includes: Use a quadtree to divide the grayscale image into blocks to obtain multiple sub-image blocks; among them, calculate the average segmentation requirement degree of all pixel points in each sub-image block, if the average segmentation requirement degree of the sub-image block is higher than the preset threshold, and the segmentation times of this sub-image block are less than or equal to the preset segmentation threshold, then this sub-image block is further divided into four equal parts.

6. The pipeline wall defect detection method based on image recognition according to claim 5, characterized in that The histogram equalization adopts contrast limited adaptive histogram equalization.

7. The method for detecting pipeline wall defects based on image recognition according to claim 6, wherein, The pipeline wall defect detection result includes the position, size and type of the defect.

8. A pipeline wall defect detection method based on image recognition according to claim 1, characterized in that The neighborhood detail degree of the ; In the formula, is the neighborhood detail degree of the -th pixel point, is the Euclidean distance between the -th pixel point and the -th edge pixel point in its neighborhood, represents the total number of edge pixel points in the neighborhood of the -th pixel point, represents the exponential function with the natural number e as the base, is the gradient amplitude of the -th edge pixel point, is the gradient direction of the -th pixel point, is the gradient direction of the -th pixel point, is the average distance between the -th pixel point and all edge pixel points in its neighborhood, represents normalization processing.

9. A pipeline wall defect detection method based on image recognition according to claim 1, characterized in that, The process of obtaining the neighborhood rule index of the th pixel is as follows: For the th pixel, starting from any pixel in its preset neighborhood, traverse the other pixels in the neighborhood in sequence according to the set direction, and assign a direction code to each traversed pixel; Compare the differences between adjacent direction codes, count the number of all direction changes to obtain the total number of direction changes, and use the ratio of the total number of direction changes to the number of direction codes as the neighborhood rule index of the pixel point.

10. A pipeline wall defect detection system based on image recognition, characterized in that, Including: A processor and a memory, the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the image recognition-based pipeline wall defect detection method according to any one of claims 1-9 is implemented.

Images