A method for visual detection of thermal shock cracks in an industrial robot protective cover

Abstract

This invention relates to the field of industrial visual inspection technology, specifically a visual detection method for thermal shock cracks in industrial robot protective covers. The method includes: acquiring reference and detection images from multiple fixed perspectives using a high-resolution industrial camera before and after the thermal shock cyclic test of the protective cover; achieving pixel-level alignment through grayscale normalization and geometric registration; dividing the image sequence into blocks for differential analysis; calculating the cumulative difference value of local intensity to mark potential abnormal regions; extracting corresponding pixel blocks and calculating a histogram of texture direction distribution; eliminating pseudo-abnormal regions with low texture distribution consistency through similarity measurement to obtain a set of candidate crack regions; and then verifying the edge continuity of suspected crack regions. This method can effectively reduce interference from ambient light, thermal deformation, etc., filter out non-crack-like false anomalies, and improve the positioning accuracy and recognition accuracy of thermal shock crack detection in protective covers.

Description

Technical Field

[0001] This invention relates to the field of industrial visual inspection technology, and in particular to a visual inspection method for thermal shock cracks in the protective cover of an industrial robot. Background Technology

[0002] Current visual inspection of thermal shock cracks in industrial robot protective covers mostly relies on comparing the overall difference between images before and after the thermal shock test and judging the difference in grayscale values ​​of individual pixels to identify surface anomalies. Some detection methods combine this with simple texture features such as texture entropy and grayscale variance to perform preliminary screening of identified abnormal areas. These detection methods directly rely on the grayscale differences of the global image to complete anomaly localization, without combining the local change characteristics of the thermal shock cracks in the protective cover to construct a targeted anomaly judgment logic.

[0003] Overall image differencing is susceptible to interference from factors such as changes in ambient lighting and thermal expansion deformation of the protective shield. Single-pixel intensity difference determination generates a large number of discrete noise points, thereby marking many non-crack-like false anomaly areas. Simple texture features can only reflect the roughness of local textures and cannot depict the texture direction changes that lead to crack formation. It is difficult to effectively distinguish between thermal shock texture distortion and real cracks, resulting in poor false anomaly removal and insufficient accuracy in crack area localization.

[0004] Existing detection methods cannot achieve quantitative analysis of local intensity accumulation differences in image block dimensions, nor can they filter out false anomalies through texture direction distribution features. They are also difficult to accurately extract crack candidate regions under complex interference, and cannot meet the needs of high-precision visual detection of thermal shock cracks in protective covers. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a visual inspection method for thermal shock cracks in industrial robot protective covers.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a visual inspection method for thermal shock cracks in industrial robot protective covers, comprising: Before and after the industrial robot protective cover undergoes a preset thermal shock cycle test, reference images and test images of the protective cover surface are acquired from multiple fixed angles using a high-resolution industrial camera. The acquired reference image and detection image are subjected to grayscale normalization and geometric position registration to form a pixel-level aligned reference image sequence and detection image sequence; A block-based difference operation is performed on the reference image sequence and the detection image sequence to calculate the local intensity cumulative difference value of each image block. Image blocks whose local intensity cumulative difference value exceeds a preset threshold are marked as potential abnormal regions. Extract the corresponding pixel blocks of the potential abnormal region in the reference image sequence and the detection image sequence, and calculate the texture direction distribution histogram of the corresponding pixel blocks; The similarity of the texture direction distribution histogram is measured to identify pixel blocks whose texture direction distribution consistency is lower than a preset standard, and these blocks are removed from the potential abnormal regions to obtain a set of crack candidate regions. Based on the set of candidate crack regions, suspected crack regions are obtained, and edge continuity verification is performed on the suspected crack regions.

[0007] As a further aspect of the present invention, the acquired reference image and detection image are subjected to grayscale normalization and geometric position registration to form a pixel-aligned reference image sequence and detection image sequence, including: For each of the reference images and detection images, pre-set circular reference markers are identified at fixed positions at the four corners and edges of the images, and the center pixel coordinates of the circular reference markers are extracted. Using the coordinates of the marker points in the reference image acquired before the thermal shock cycle test as a reference, a perspective transformation model is applied to the detection image acquired after the thermal shock cycle test to align the coordinates of the marker points in the detection image with the coordinates of the marker points in the reference image. On the aligned reference image and detection image, the average gray value of the entire image is calculated respectively. Using the global average gray value as the target, linear stretching transformation is performed on the two images respectively to complete the gray value normalization. The reference image and the detection image, which have undergone grayscale normalization and geometric position registration, and are from the same viewpoint, are stored in the reference image sequence and the detection image sequence respectively in the order of acquisition.

[0008] As a further aspect of the present invention, block-based difference operations are performed on the reference image sequence and the detection image sequence to calculate the cumulative local intensity difference value of each image block, including: Each pair of registered images in the reference image sequence and the detection image sequence is divided into square image blocks of equal size; For each pair of square image blocks with the same position, the absolute difference between the pixel gray value in the detection image sequence and the pixel gray value in the reference image sequence is calculated for each pixel. The absolute differences of all pixels within a square image block are summed to obtain the original difference sum of the image blocks. Based on the average gray value of the corresponding image block in the reference image sequence, the original difference is normalized and compensated to obtain the final local intensity cumulative difference value.

[0009] As a further aspect of the present invention, the corresponding pixel blocks of the potential anomaly region in the reference image sequence and the detection image sequence are extracted, and the texture direction distribution histogram of the corresponding pixel blocks is calculated, including: For each potential abnormal region, determine the range of pixel coordinates it occupies in the reference image sequence, and extract the image content within the range of pixel coordinates as a reference pixel block; Image content within the same pixel coordinate range is extracted from the detected image sequence and used as a detection pixel block; The reference pixel block and the detection pixel block are respectively convolved by a multi-directional gradient filter bank to obtain the response amplitude of each pixel in multiple gradient directions; For each pixel, the gradient direction with the largest response amplitude is selected as the main texture direction of the pixel; The distribution of the main texture direction of all pixels in the reference pixel block and the detection pixel block in different directional intervals is statistically analyzed to form a texture direction distribution histogram that reflects the concentration of texture direction.

[0010] As a further aspect of the present invention, a similarity measurement is performed on the texture direction distribution histogram to identify pixel blocks whose texture direction distribution consistency is lower than a preset standard, and these blocks are removed from the potential abnormal regions, including: The texture direction distribution histograms of the reference pixel block and the detection pixel block are smoothed to eliminate statistical fluctuations caused by noise. Calculate the numerical difference between the two texture direction distribution histograms after smoothing in each direction interval, and take the sum of the squares of the numerical differences in all direction intervals as the distribution difference degree; The distribution difference is compared with a texture consistency threshold obtained in advance from training samples; When the distribution difference is greater than the texture consistency threshold, it is determined that the potential abnormal region corresponding to the pixel block is caused by surface stains or changes in lighting, rather than cracks, and the abnormal region is removed from the set of potential abnormal regions. All potential abnormal regions whose distribution difference is less than or equal to the texture consistency threshold are retained to form the crack candidate region set.

[0011] As a further aspect of the present invention, obtaining suspected crack regions based on the set of crack candidate regions includes: Calculate the aspect ratio of the minimum bounding rectangle for each candidate region in the set of crack candidate regions; Set an aspect ratio threshold, filter out candidate regions with aspect ratios greater than the threshold, and form a set of slender candidate regions; For each region in the set of elongated candidate regions, grayscale morphological closing operations are performed on the corresponding original image region in the detection image sequence using multiple linear structuring elements of different sizes. Compare the average grayscale change rate of the image region after performing a closing operation on linear structuring elements of different sizes; Candidate regions whose average grayscale change rate exceeds a preset change rate threshold are identified as suspected crack regions. Record the minimum bounding rectangle parameters of the suspected crack region and its position coordinates in the detection image sequence.

[0012] As a further aspect of the present invention, for each region in the elongated candidate region set, grayscale morphological closing operations are performed on the corresponding original image region in the detected image sequence using multiple linear structuring elements of different sizes, including: Extract the original grayscale image block corresponding to the elongated candidate region from the detected image sequence; Prepare a set of linear structural elements with increasing length but fixed width, wherein the length direction of the linear structural elements is consistent with their preset traversal direction in the image block; Each of the linear structuring elements is used sequentially to perform a grayscale morphological closing operation on the original grayscale image block. The grayscale morphological closing operation includes performing a dilation operation followed by an erosion operation. After each closing operation is completed, the average gray value of the entire image block is calculated; Plot the relationship curve between the length of the linear structuring element used and the average gray value of the image block obtained after the operation.

[0013] As a further aspect of the present invention, edge continuity verification is performed on the suspected crack region, including: For each suspected crack region, an adaptive threshold edge detection operator is applied at the original image position of its corresponding detection image sequence to extract a set of clear edge pixels. The extracted set of edge pixels is then tracked and connected based on orientation consistency. Calculate the total length, cumulative curvature change, and straight-line distance between endpoints of each edge curve formed after connection; The continuity confidence score of the edge curve is calculated based on the ratio of the total length to the straight-line distance between the endpoints and the cumulative value of the curvature change. Suspected crack areas with a continuity confidence score lower than the preset continuity standard are excluded, and areas that meet the continuity requirements are retained to form a final list of confirmed crack areas.

[0014] As a further aspect of the present invention, the step of performing edge pixel tracking and connection based on orientation consistency on the extracted set of edge pixels includes: Arbitrarily select an unvisited edge pixel from the set of edge pixels as the current seed point; Check if there are any neighboring pixels belonging to the set of edge pixels within the eight-neighborhood of the current seed point; If there is a neighboring pixel that satisfies the orientation consistency constraint, then the neighboring pixel is connected to the current edge chain and updated as the new current seed point. The orientation consistency constraint requires that the difference between the gradient direction of the newly added pixel and the average direction of the current edge chain is less than the angle tolerance. Repeat the checking and connection process until the current edge chain can no longer be extended; Select a new seed point from the remaining set of edge pixels and start the tracing and connection of the next edge chain until all edge pixels have been visited, forming several edge curves.

[0015] As a further aspect of the present invention, the method further includes the step of quantifying and annotating the finally confirmed list of crack regions: For each crack region in the final confirmed crack region list, its pixel range is located in the detection image sequence; Calculate the pixel area, principal axis length, principal axis direction, and contrast between the average gray level inside the region and the average gray level of the surrounding background for each crack region. The calculated pixel area, principal axis length, principal axis direction, and contrast are stored as the quantized feature vector of the crack region. On the original sequence of detected images, the boundaries of each crack region are drawn as highlighted outlines; The quantized feature vector is associated with and stored with the corresponding crack region boundary image to generate a detection report file containing complete crack location and morphology information.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: Block-based differential operations are performed on the pixel-aligned reference image sequence and the detection image sequence to calculate the local intensity cumulative difference value of each image block. Potential abnormal areas are marked by comparing the local intensity cumulative difference value with a preset threshold. This avoids the influence of overall grayscale shift caused by factors such as ambient light fluctuations and thermal deformation of the protective cover on the overall image differential method. It also reduces the interference of discrete noise points generated by single-pixel-level differential judgment on the identification of abnormal areas, so that the abnormal marking results focus on areas with significant local grayscale changes in the image, narrowing the processing scope of abnormality analysis. This makes the initially located abnormal areas more consistent with the grayscale change characteristics corresponding to cracks on the surface of the protective cover caused by thermal shock, reducing the occurrence of irrelevant areas being misjudged as abnormal areas.

[0017] The algorithm extracts the corresponding pixel blocks of potential abnormal regions in the baseline image sequence and the detection image sequence, calculates the texture direction distribution histogram of the corresponding pixel blocks, and performs similarity measurement on the histogram. Pixel blocks with texture direction distribution consistency lower than the preset standard are removed. Based on the distribution characteristics of texture direction, it can distinguish the texture distortion caused by thermal shock on the surface of the protective cover from the texture structure changes formed by cracks. It filters out non-crack-like abnormal regions formed by factors such as shooting angle deviation, slight surface wear, and thermal expansion and contraction deformation, and retains abnormal regions that match crack features to form a set of crack candidate regions. This reduces the processing pressure of the subsequent edge continuity verification stage, reduces the interference of false abnormal regions on crack identification results, makes the division of crack candidate regions more in line with the texture change rules of real cracks, improves the accuracy of crack region screening, and makes the finally locked suspected crack regions closer to the actual thermal shock crack regions. Attached Figure Description

[0018] Figure 1 This is a flowchart of the visual inspection method for thermal shock cracks in industrial robot protective covers according to the present invention. Figure 2 A flowchart for calculating the cumulative difference in local intensity between block differential and local intensity values; Figure 3 The flowchart for calculating the histogram of texture direction distribution. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] See Figure 1 This invention provides a visual detection method for thermal shock cracks in industrial robot protective covers. The specific implementation is as follows: Before and after a preset thermal shock cycle test, reference images and test images of the protective cover surface are acquired from multiple fixed perspectives using a high-resolution industrial camera. The acquired reference and test images are normalized in grayscale and geometrically registered to form pixel-aligned reference and test image sequences. Block-based difference operations are performed on the reference and test image sequences to calculate the cumulative local intensity difference value of each image block. Image blocks with cumulative local intensity difference values ​​exceeding a preset threshold are marked as potential anomalous regions. The corresponding pixel blocks of the potential anomalous regions in the reference and test image sequences are extracted, and the texture direction distribution histogram of the corresponding pixel blocks is calculated. A similarity measurement is performed on the texture direction distribution histograms to identify pixel blocks with texture direction distribution consistency lower than a preset standard, and these are removed from the potential anomalous regions, resulting in a crack candidate region set. Suspected crack regions are obtained based on the crack candidate region set, and edge continuity verification is performed on the suspected crack regions.

[0022] In one embodiment of the present invention, see [reference] Figure 2For each reference image and test image, pre-defined circular reference markers are identified at fixed positions at the four corners and edges of the images, and the center pixel coordinates of the circular reference markers are extracted. Using the marker coordinates of the reference image acquired before the thermal shock cycle test as a reference, a perspective transformation model is applied to the test image acquired after the thermal shock cycle test to align the marker coordinates in the test image with those in the reference image. On the aligned reference and test images, the average grayscale value of the entire image is calculated, and a linear stretching transformation is performed on both images with the global average grayscale value as the target to complete grayscale normalization. The reference and test images from the same viewpoint, after grayscale normalization and geometric registration, are stored in the reference image sequence and test image sequence respectively, according to the acquisition order. Each pair of registered images in the reference and test image sequences is divided into equally sized square image blocks. For each pair of square image blocks with the same position, the absolute difference between the pixel grayscale value in the test image sequence and the pixel grayscale value in the reference image sequence is calculated pixel by pixel. The original difference sum of the image blocks is obtained by summing the absolute differences of all pixels within a square image block. Then, based on the average grayscale value of the corresponding image block in the reference image sequence, a normalization compensation calculation is performed on the original difference sum to obtain the final local intensity cumulative difference value.

[0023] In a specific implementation, a visual detection method for thermal shock cracks in an industrial robot protective cover involves acquiring and processing images before and after a thermal shock test. This embodiment details the specific processes of image registration, grayscale normalization, and block-based differential operations. In an example scenario, the protective cover is made of composite ceramic material, with dimensions of 500 mm × 300 mm. The high-resolution industrial camera has a resolution of 4096 × 2160 pixels, and the fixed viewing angles include four directions: front, left side, right side, and top. The preset thermal shock cycle test conditions are 100 cycles between -20 degrees Celsius and 120 degrees Celsius, with each cycle including 20 minutes of high-temperature holding and 20 minutes of low-temperature holding.

[0024] In specific implementation, before the industrial robot protective cover undergoes thermal shock cycle testing, reference images of the cover surface are acquired from four fixed perspectives. After the test, the industrial robot and high-resolution industrial camera remain in the same position, and test images are acquired from the same perspective. Black circular reference markers with a diameter of 5 mm are affixed to the four corners and the midpoint of each side of each reference and test image. In some embodiments, these preset circular reference markers are identified using image processing algorithms, and the center pixel coordinates of each circular reference marker are extracted. For example, in the reference image acquired before the test, the center pixel coordinates of the upper left corner marker are (200, 200), while in the test image acquired after the test, the center pixel coordinates of the same marker may shift to (205, 198).

[0025] In practice, the coordinates of the marker points in the reference image acquired before the thermal shock cycling test are used as a reference. A perspective transformation model is applied to the detection image acquired after the thermal shock cycling test for geometric registration. The perspective transformation model is achieved by solving a set of coordinate transformation matrices corresponding to the marker points, ensuring that the image coordinate system of all pixels in the detection image is aligned with that of the reference image after the transformation. Specifically, after the perspective transformation, the marker point in the detection image with original coordinates (205, 198) has its new coordinates corrected to (200, 200) consistent with the reference image, thus achieving pixel-level geometric alignment of the entire image.

[0026] In the specific implementation, the average grayscale value of all pixels in the aligned reference image and detection image is calculated separately. The average grayscale value of the reference image is calculated to be 128, and the average grayscale value of the detection image is calculated to be 122. Using a global average grayscale value of 125 as the target, linear stretching transformation is performed on both the reference image and the detection image to complete grayscale normalization. The linear stretching transformation adjusts the overall brightness level of the two images to a consistent state by adjusting the grayscale histogram distribution. The reference image and detection image, which have undergone grayscale normalization and geometric registration processing and are from the same viewpoint, are stored in the reference image sequence and detection image sequence respectively, according to the acquisition time sequence.

[0027] In practice, each pair of registered images in the reference image sequence and the detection image sequence is divided into equally sized square image blocks, each 32 pixels × 32 pixels. For each pair of square image blocks with the same position, the absolute difference between the pixel grayscale value in the detection image sequence and the pixel grayscale value in the reference image sequence is calculated pixel by pixel. In a specific block, if the grayscale value of a pixel in the detection image sequence is 130 and the grayscale value of the corresponding pixel in the reference image sequence is 125, then the absolute difference is 5. The absolute differences of all pixels within a square image block are summed to obtain the original sum of differences for the image blocks.

[0028] In practice, based on the average grayscale value of the corresponding image blocks in the reference image sequence, the original differences are normalized and compensated to obtain the final local intensity cumulative difference value. The normalization compensation calculation aims to eliminate the influence of the average brightness of the image blocks themselves on the cumulative difference value, making the differences between different bright and dark areas comparable. Optionally, the local intensity cumulative difference value is calculated using the following formula:

[0029] Where D represents the cumulative difference in local intensity, and S represents the sum of the original differences between image blocks. This represents the average gray value of the corresponding image block in the reference image sequence. C is a small positive constant used to prevent the denominator from being zero and to stabilize the calculation result; for example, C is set to 1.0. Assuming the original difference sum S of an image block is 1580, its corresponding average gray value of the reference image block is... If the value is 120, then the calculated local intensity cumulative difference value D is approximately 13.06. It can be understood that the calculated local intensity cumulative difference value is compared with a preset fixed threshold, for example, a preset threshold of 10.0. Image blocks with local intensity cumulative difference values ​​exceeding the preset threshold are marked as potential anomalous regions.

[0030] In one embodiment of the present invention, see [reference] Figure 3 For each potential anomalous region, its pixel coordinate range in the baseline image sequence is determined, and the image content within this range is extracted as the baseline pixel block. Image content within the same pixel coordinate range is extracted from the detection image sequence as the detection pixel block. Multi-directional gradient filter banks are applied to both the baseline and detection pixel blocks for convolution calculation, yielding the response amplitude of each pixel in multiple gradient directions. For each pixel, the gradient direction with the largest response amplitude is selected as the dominant texture direction. The distribution of dominant texture directions for all pixels in the baseline and detection pixel blocks within different directional intervals is statistically analyzed, forming a texture direction distribution histogram reflecting the concentration of texture directions. The texture direction distribution histograms of the baseline and detection pixel blocks are smoothed to eliminate statistical fluctuations caused by noise. The numerical difference between the two smoothed texture direction distribution histograms in each directional interval is calculated, and the sum of the squares of the numerical differences across all directional intervals is taken as the distribution dissimilarity. The distribution dissimilarity is compared with a texture consistency threshold obtained beforehand from training samples. When the distribution difference is greater than the texture consistency threshold, the potential abnormal region corresponding to the pixel block is determined to be caused by surface stains or lighting changes, rather than a crack, and the region is removed from the potential abnormal region set. All potential abnormal regions with a distribution difference less than or equal to the texture consistency threshold are retained to form a crack candidate region set.

[0031] In a specific implementation, a visual detection method for thermal shock cracks in industrial robot protective covers involves texture consistency analysis of marked potential abnormal regions. This embodiment details the calculation of the texture direction distribution histogram and the similarity measurement process. In an example scenario, the protective cover surface material is a carbon fiber composite material with a woven texture. Potential abnormal regions are marked by block difference operations. One region has pixel coordinates ranging from (100, 150) to (164, 214). This region appears as a dark linear mark in the image after the thermal shock test, but this mark may be caused by surface oil rather than a crack.

[0032] In specific implementations, for each potential abnormal region, the pixel coordinate range it occupies in the reference image sequence is determined, and the image content is cropped based on this coordinate range as a reference pixel block. In the detection image sequence, the image content is cropped using the exact same pixel coordinate range as the detection pixel block, ensuring that the two pixel blocks correspond to the same physical location of the protective shield in the image. In some embodiments, both the reference pixel block and the detection pixel block are 64 pixels × 64 pixels in size, and each contains surface texture information of the same area before and after the thermal shock test.

[0033] In the specific implementation, a multi-directional gradient filter bank is applied to both the reference pixel block and the detection pixel block for convolution calculation, obtaining the response amplitude of each pixel in multiple gradient directions. The multi-directional gradient filter bank can include Sobel or Prewitt operators in four directions: 0 degrees, 45 degrees, 90 degrees, and 135 degrees. For a pixel in the reference pixel block, its gradient response amplitudes in the 0-degree, 45-degree, 90-degree, and 135-degree directions are calculated to be 15, 8, 120, and 10, respectively. For each pixel, the gradient direction with the largest response amplitude is selected as the main texture direction of that pixel. Based on the aforementioned data, the main texture direction of this pixel is determined to be the 90-degree direction. The distribution of the main texture directions of all pixels in the reference pixel block and the detection pixel block within different directional intervals is statistically analyzed to form a texture direction distribution histogram reflecting the concentration of texture directions. It is understandable that the directional range can be evenly divided into 8 intervals, each interval being 22.5 degrees wide, and the number of pixels falling into each interval is counted to form a histogram with 8 bars.

[0034] In practice, the texture orientation distribution histograms of the reference pixel block and the detected pixel block are smoothed to eliminate statistical fluctuations caused by image noise or minor local changes. Smoothing can be achieved by convolving the histogram sequence with a moving average filter. The numerical difference between the two smoothed texture orientation distribution histograms in each orientation interval is calculated, and the sum of the squares of the numerical differences across all orientation intervals is taken as the distribution dissimilarity. Optionally, the distribution dissimilarity is calculated using the following formula:

[0035] Where: E represents the distributional disparity, N represents the total number of directional intervals, for example, N=8, H b (k) represents the value of the smoothed baseline pixel block texture direction distribution histogram in the k-th direction interval, H d (k) represents the value of the smoothed pixel block texture direction distribution histogram in the k-th direction interval. Assume the calculated distribution difference E is 450.7.

[0036] In practice, the distribution difference is compared with a texture consistency threshold obtained from pre-trained samples. The texture consistency threshold is calculated by analyzing the distribution difference of texture directions between a large number of known non-crack change regions and real crack regions. When the distribution difference is greater than the texture consistency threshold, the potential abnormal region corresponding to the pixel block is determined to be caused by surface stains or lighting changes, rather than a crack, and the region is removed from the set of potential abnormal regions. For example, if the texture consistency threshold is 300 and the calculated distribution difference E is 450.7, the region is removed. In essence, retaining all potential abnormal regions with distribution differences less than or equal to the texture consistency threshold constitutes a set of crack candidate regions. The texture structure of these regions maintains high consistency before and after thermal shock, conforming to crack characteristics.

[0037] In one embodiment of the present invention, the aspect ratio of the minimum bounding rectangle of each candidate region in the crack candidate region set is calculated. An aspect ratio threshold is set, and candidate regions with aspect ratios greater than the threshold are selected to form a set of elongated candidate regions. For each region in the elongated candidate region set, grayscale morphological closing operations are performed on the corresponding original image region in the detection image sequence using multiple linear structuring elements of different sizes. The average grayscale change rate of the image region after closing operations using linear structuring elements of different sizes is compared. Candidate regions with an average grayscale change rate exceeding a preset change rate threshold are identified as suspected crack regions. The minimum bounding rectangle parameters of the suspected crack regions and their position coordinates in the detection image sequence are recorded. The original grayscale image blocks corresponding to the elongated candidate regions are extracted from the detection image sequence. A set of linear structuring elements with increasing length but fixed width is prepared, where the length direction of the linear structuring elements is consistent with their preset traversal direction in the image block. Each linear structuring element is used sequentially to perform grayscale morphological closing operations on the original grayscale image blocks, including dilation followed by erosion. After each closing operation, calculate the average gray value of the entire image patch. Plot the relationship between the length of the linear structuring element used and the average gray value of the resulting image patch.

[0038] In its implementation, a visual detection method for thermal shock cracks in industrial robot protective covers involves further morphological analysis of crack candidate regions after texture screening. This embodiment details the specific process of identifying suspected crack regions based on aspect ratio screening and grayscale morphological closing operations. The crack candidate region set may contain irregular stain areas and real slender cracks. To distinguish between the two, the aspect ratio of the minimum bounding rectangle of each candidate region in the crack candidate region set is calculated. The minimum bounding rectangle is the rectangle with the smallest area that can completely enclose the pixels of the candidate region. In an example scenario, the crack candidate region set contains 15 regions. Calculations show that the aspect ratio of the minimum bounding rectangle of 5 regions is greater than 8.0, while the aspect ratios of the remaining regions are distributed between 1.2 and 3.5.

[0039] In practice, an aspect ratio threshold is set for filtering. With an aspect ratio threshold of 5.0, candidate regions with aspect ratios greater than the threshold are selected, forming a set of elongated candidate regions. In the previous example, the five regions with aspect ratios greater than 8.0 were included in this set. For each region in the elongated candidate region set, grayscale morphological closing operations are performed on the corresponding original image region in the detection image sequence using multiple linear structuring elements of different sizes. The original grayscale image patch corresponding to the elongated candidate region is extracted from the detection image sequence. The size of the original grayscale image patch is slightly larger than the minimum bounding rectangle of the elongated candidate region to ensure that the entire region and its adjacent background are included.

[0040] In practice, a set of linear structuring elements (LSIs) with increasing length but fixed width is prepared. The width of each LSI is fixed at 1 pixel, and its length direction is consistent with its preset traversal direction within the image block. The preset traversal direction is typically set perpendicular to the principal axis of the candidate region. The length values ​​of the LSIs are 5, 10, 15, 20, 25, and 30 pixels, respectively. Each LSI is used sequentially to perform a grayscale morphological closing operation on the original grayscale image block. This operation involves first performing grayscale dilation on the image block, followed by grayscale erosion. After each closing operation, the average grayscale value of all pixels in the entire image block is calculated, and the length of the currently used LSI and the resulting average grayscale value of the image block are recorded.

[0041] In practical implementation, a curve showing the relationship between the length of the linear structuring element and the average grayscale value of the resulting image patch is plotted. This illustrates that for a real crack region, the image appears as a dark line. When the length of the linear structuring element is less than the crack length, the closing operation partially fills the crack, increasing its average grayscale value. However, when the structuring element length exceeds the crack length, the change in average grayscale value tends to level off. The average grayscale change rate of the image region is compared after closing operations with linear structuring elements of different sizes. The average grayscale change rate characterizes the relative speed at which the average grayscale value changes with the increase in the length of the structuring element. The average grayscale change rate is calculated using the following formula:

[0042] Where: R i V represents the average grayscale change rate after performing a closing operation on the i-th linear structuring element. i Indicates the use of length L i The average gray value of the image patch after the structuring element closing operation, V i-1 Indicates the use of length L i-1 The average gray value of the image patch after the structuring element closing operation, L i and L i-1 These are the lengths of the i-th and (i-1)-th linear structuring elements, respectively. Candidate regions with an average grayscale change rate exceeding a preset change rate threshold are identified as suspected crack regions. The preset change rate threshold is, for example, set to 0.5 grayscale values ​​per pixel length.

[0043] In some embodiments, referring to Table 1, the data of the average gray value of the image patch and the calculated average gray change rate are shown after applying a closing operation to a specific elongated candidate region using linear structuring elements of different lengths.

[0044] Table 1: Effect of Closing Operation on Linear Structural Elements

[0045] As shown in Table 1, when the length of the linear structuring element increases from 5 pixels to 10 pixels, the average grayscale change rate reaches 3.20, exceeding the preset change rate threshold of 0.5. However, when the length continues to increase, the average grayscale change rate decreases rapidly. It can be understood that the grayscale distribution in this region changes significantly under the operation of short structuring elements, consistent with crack characteristics, and is therefore identified as a suspected crack region. The minimum bounding rectangle parameters of the suspected crack region and its position coordinates in the detection image sequence are recorded. The minimum bounding rectangle parameters include the center point coordinates, length, width, and rotation angle. The position coordinates are used to locate the region in the original image sequence. In some embodiments, after this step, 3 out of 5 regions in the elongated candidate region set are identified as suspected crack regions, while the other 2 regions are excluded because their average grayscale change rate is consistently below the threshold.

[0046] In one embodiment of the present invention, for each suspected crack region, an adaptive threshold edge detection operator is applied to the original image position of its corresponding detection image sequence to extract a set of clear edge pixels. Edge pixel tracking and connection based on orientation consistency are performed on the extracted edge pixel set. The total length, cumulative curvature change, and straight-line distance between endpoints of each edge curve formed after connection are calculated. The continuity confidence score of the edge curve is calculated based on the ratio of the total length to the straight-line distance between endpoints and the cumulative curvature change. Suspected crack regions with continuity confidence scores lower than a preset continuity standard are excluded, and regions that meet the continuity requirements are retained to form a final confirmed crack region list. An unvisited edge pixel is arbitrarily selected from the edge pixel set as the current seed point. It is checked whether there are neighboring pixels belonging to the edge pixel set within the eight-neighborhood of the current seed point. If there are neighboring pixels that satisfy the orientation consistency constraint, the neighboring pixels are connected to the current edge chain and updated as the new current seed point. The orientation consistency constraint requires that the difference between the gradient direction of the newly added pixel and the average direction of the current edge chain is less than the angle tolerance. Repeat the checking and connection process until the current edge chain cannot be extended further. Select a new seed point from the remaining set of edge pixels and start tracing and connecting the next edge chain until all edge pixels have been visited, forming several edge curves.

[0047] In a specific implementation, a visual detection method for thermal shock cracks in industrial robot protective covers involves verifying the edge continuity of suspected crack regions obtained from morphological analysis. This embodiment details the specific process of edge detection, pixel connection, and continuity confidence assessment. For each suspected crack region, an adaptive threshold edge detection operator is applied to the original image position of its corresponding detection image sequence to extract a set of clear edge pixels. The adaptive threshold edge detection operator can be, for example, the Canny operator, whose high and low thresholds are automatically calculated based on the grayscale statistical characteristics of the local image of the suspected crack region. In an example scenario, a suspected crack region located near image coordinates (200, 300) and approximately 50 pixels long is processed by the Canny operator, extracting a total of 68 discrete edge pixels to form an initial set of edge pixels.

[0048] In the specific implementation, edge pixel tracking and connection based on orientation consistency are performed on the extracted set of edge pixels. An unvisited edge pixel is arbitrarily selected from the set as the current seed point, for example, the pixel with coordinates (205, 305). It is then checked whether there are any neighboring pixels belonging to the edge pixel set within the eight-neighborhood of the current seed point. The eight-neighborhood includes eight adjacent pixel positions in the top, bottom, left, right, and four diagonal directions. If a neighboring pixel satisfying the orientation consistency constraint exists, it is connected to the current edge chain, and this neighboring pixel is updated as the new current seed point. The orientation consistency constraint requires that the difference between the gradient direction of the newly added pixel and the average direction of the current edge chain is less than an angle tolerance, which can be set to 15 degrees. This checking and connection process is repeated until the current edge chain cannot be extended further, meaning that there are no more unvisited edge pixels satisfying the orientation consistency constraint within the eight-neighborhood of the current seed point. Select a new seed point from the remaining set of edge pixels and start the tracing and connection of the next edge chain until all edge pixels have been visited, forming several edge curves. In the example above, the 68 edge pixels are finally connected into a main edge curve containing 63 pixels and a short isolated curve containing 5 pixels.

[0049] In practice, the total length, cumulative curvature change, and straight-line distance between endpoints of each edge curve formed after connection are calculated. The total length of the edge curve is obtained by accumulating the Euclidean distance between adjacent pixels in the edge chain. For an 8-connected chain, the distance between adjacent pixels is 1 (horizontal / vertical) or... (Diagonal). The cumulative curvature change is obtained by summing the absolute values ​​of the changes in direction angle at each inflection point on the edge chain, with the direction angle change expressed in degrees. The straight-line distance between endpoints refers to the Euclidean distance between the pixel coordinates of the starting and ending points of the edge curve. The continuity confidence score of the edge curve is calculated based on the ratio of the total length to the straight-line distance between endpoints, and the ratio of the cumulative curvature change to a reference curvature change value. It can be understood that a continuous, smooth crack edge curve should have a large ratio of total length to the straight-line distance between endpoints, and a relatively small cumulative curvature change.

[0050] In practice, the continuity confidence score of the edge curve is calculated using the following formula:

[0051] Where: C s The continuous confidence score is a dimensionless value; L represents the total length of the edge curve in pixels; D represents the straight-line distance between the endpoints of the edge curve in pixels; K represents the cumulative curvature change of the edge curve in degrees. refThis represents a reference curvature change value used to make the ratio dimensionless; its unit is also degrees, such as K. ref The value can be 90 degrees; α and β are positive weighting coefficients, dimensionless values, used to balance the contributions of length ratio and normalized curvature change, for example, α can be set to 10.0 and β to 1.0. Suspected crack areas with continuity confidence scores lower than the preset continuity standard are excluded, and areas that meet the continuity requirements are retained to form the final confirmed crack area list. The preset continuity standard can be set to 5.0.

[0052] In some embodiments, the main edge curve and the short isolated curve obtained by connecting in the foregoing examples are calculated separately (K). ref =90°), see Table 2.

[0053] Table 2: Calculation Table for Continuity Verification of Edge Curves

[0054] As shown in Table 2, the continuity confidence score of the main edge curve is 13.13, higher than the preset continuity standard of 5.0, while the score of the short isolated curve is 7.78, also higher than the standard, but it may be excluded in other steps due to its excessive shortness. It is understandable that, after calculation, the main edge curve, with its larger length proportion and relatively straight shape, received a higher score. In some embodiments, after edge continuity verification, two of the three suspected crack regions were confirmed as actual crack regions. Optionally, the weighting coefficients α and β, and the reference value K... ref The specific values ​​can be obtained by analyzing and training a known crack sample library to optimize the ability to distinguish between real cracks and pseudo cracks.

[0055] In one embodiment of the present invention, quantized feature extraction and annotation are performed on the final confirmed list of crack regions. For each crack region in the final confirmed list, its pixel range is located in the detection image sequence. The pixel area, principal axis length, principal axis direction, and contrast between the average gray level inside the region and the average gray level of the surrounding background are calculated for each crack region. The calculated pixel area, principal axis length, principal axis direction, and contrast are stored as quantized feature vectors of the crack region. The boundary of each crack region is drawn as a highlighted outline on the original detection image sequence. The quantized feature vectors are associated and stored with the corresponding crack region boundary images to generate a detection report file containing complete crack location and morphology information.

[0056] In a specific implementation, a visual detection method for thermal shock cracks in industrial robot protective covers, after confirming the crack regions, performs quantitative feature extraction and annotation on the final confirmed crack region list. The final confirmed crack region list contains several crack region information verified by the aforementioned steps. This embodiment details the specific process of extracting quantitative features for these crack regions, drawing boundaries, and generating an inspection report. In an example scenario, the final confirmed crack region list contains two entries, corresponding to two cracks at different locations on the protective cover surface. The first crack is denoted as crack region A, located near image coordinates (150, 300), and the second crack is denoted as crack region B, located near image coordinates (400, 120).

[0057] In specific implementation, for each crack region in the final confirmed crack region list, its pixel range is located in the detected image sequence. The pixel range of the crack region is defined by its minimum bounding rectangle or precise boundary contour point set. The pixel area, principal axis length, principal axis direction, and contrast between the average gray level inside the region and the average gray level of the surrounding background are calculated for each crack region. The pixel area is obtained by counting all pixels belonging to the crack region. The principal axis length and principal axis direction are calculated by performing principal component analysis on the crack region pixels to obtain the eigenvalues ​​and eigenvectors of the first principal component. In some embodiments, crack region A has a pixel area of ​​85 square pixels, a principal axis length of 22 pixels, and a principal axis direction of 65 degrees (relative to the horizontal axis of the image); crack region B has a pixel area of ​​120 square pixels, a principal axis length of 35 pixels, and a principal axis direction of 10 degrees.

[0058] In practical implementation, the contrast ratio between the average gray level within the crack area and the average gray level of the surrounding background is calculated as follows: First, calculate the average gray level of all pixels within the crack area. Then, extend the minimum bounding rectangle of the crack area by 5-10 pixels to form a ring-shaped background area. Calculate the average gray level of all pixels within this ring-shaped background area. The contrast ratio between the average gray level within the crack area and the average gray level of the surrounding background is calculated using the following formula:

[0059] Where: Q represents contrast, G in G represents the average gray level within the crack region. out This represents the average gray level of the background surrounding the crack region. G represents the gray level of crack region A. in The value is 50, G out With a value of 130, the calculated contrast Q is approximately 0.615, and the G of the crack region B... in The value is 60, G out With a value of 125, the contrast ratio Q is approximately 0.520. It can be understood that the contrast ratio Q reflects the degree of difference in grayscale between the crack and the background.

[0060] In practice, the calculated pixel area, principal axis length, principal axis direction, and contrast are stored as quantized feature vectors for the crack region. These quantized feature vectors can be represented as (pixel area, principal axis length, principal axis direction, contrast). For example, the quantized feature vector for crack region A is stored as (85, 22, 65, 0.615), and the quantized feature vector for crack region B is stored as (120, 35, 10, 0.520). On the original detection image sequence, the boundary of each crack region is drawn as a highlighted outline. This highlighted outline can be a solid red line, 2 pixels wide, drawn along the precise edge pixels of the crack region. It can be understood that in a detection image sequence containing multiple viewpoints, if a crack exists in each viewpoint, feature extraction and boundary annotation will be performed independently.

[0061] In specific implementation, the quantized feature vectors are associated and stored with the corresponding crack region boundary images to generate an inspection report file containing complete crack location and morphology information. This association storage can be achieved by creating a data table within the inspection report file. Each row of the data table records the crack region number, its corresponding viewpoint image number, minimum bounding rectangle coordinates, quantized feature vector, and references the path to the corresponding boundary annotation image or the embedded image data itself. In some embodiments, the inspection report file uses a structured JSON or XML format to facilitate reading and parsing by other quality management systems. Optionally, the inspection report file may also include overall information such as inspection time, protective cover component number, and number of thermal shock test cycles. Optionally, for ease of manual review, the inspection report file can be output as a PDF, displaying the annotation images from each viewpoint and their corresponding feature data tables in a visually appealing format. In some embodiments, the final generated inspection report file records complete information on two cracks, providing quantitative data for evaluating the structural integrity of the protective cover after the thermal shock test.

[0062] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A visual inspection method for thermal shock cracks in the protective cover of an industrial robot, characterized in that, include: Before and after the industrial robot protective cover undergoes a preset thermal shock cycle test, reference images and test images of the protective cover surface are acquired from multiple fixed angles using a high-resolution industrial camera. The acquired reference image and detection image are subjected to grayscale normalization and geometric position registration to form a pixel-level aligned reference image sequence and detection image sequence; A block-based difference operation is performed on the reference image sequence and the detection image sequence to calculate the local intensity cumulative difference value of each image block. Image blocks whose local intensity cumulative difference value exceeds a preset threshold are marked as potential abnormal regions. Extract the corresponding pixel blocks of the potential abnormal region in the reference image sequence and the detection image sequence, and calculate the texture direction distribution histogram of the corresponding pixel blocks; The similarity of the texture direction distribution histogram is measured to identify pixel blocks whose texture direction distribution consistency is lower than a preset standard, and these blocks are removed from the potential abnormal regions to obtain a set of crack candidate regions. Based on the set of candidate crack regions, suspected crack regions are obtained, and edge continuity verification is performed on the suspected crack regions.

2. The method for visually detecting thermal shock cracks in an industrial robot protective cover according to claim 1, characterized in that, The acquired reference image and detection image are subjected to grayscale normalization and geometric position registration to form a pixel-aligned reference image sequence and detection image sequence, including: For each of the reference images and detection images, pre-set circular reference markers are identified at fixed positions at the four corners and edges of the images, and the center pixel coordinates of the circular reference markers are extracted. Using the coordinates of the marker points in the reference image acquired before the thermal shock cycle test as a reference, a perspective transformation model is applied to the detection image acquired after the thermal shock cycle test to align the coordinates of the marker points in the detection image with the coordinates of the marker points in the reference image. On the aligned reference image and detection image, the average gray value of the entire image is calculated respectively. Using the global average gray value as the target, linear stretching transformation is performed on the two images respectively to complete the gray value normalization. The reference image and the detection image, which have undergone grayscale normalization and geometric position registration, and are from the same viewpoint, are stored in the reference image sequence and the detection image sequence respectively in the order of acquisition.

3. The method for visually inspecting thermal shock cracks in an industrial robot protective cover according to claim 1, characterized in that, Perform block-based difference operations on the reference image sequence and the detection image sequence to calculate the cumulative local intensity difference value of each image block, including: Each pair of registered images in the reference image sequence and the detection image sequence is divided into square image blocks of equal size; For each pair of square image blocks with the same position, the absolute difference between the pixel gray value in the detection image sequence and the pixel gray value in the reference image sequence is calculated for each pixel. The absolute differences of all pixels within a square image block are summed to obtain the original difference sum of the image blocks. Based on the average gray value of the corresponding image block in the reference image sequence, the original difference is normalized and compensated to obtain the final local intensity cumulative difference value.

4. The method for visually detecting thermal shock cracks in an industrial robot protective cover according to claim 1, characterized in that, Extracting the corresponding pixel blocks of the potential anomaly region in the reference image sequence and the detection image sequence, and calculating the texture orientation distribution histogram of the corresponding pixel blocks, including: For each potential abnormal region, determine the range of pixel coordinates it occupies in the reference image sequence, and extract the image content within the range of pixel coordinates as a reference pixel block; Image content within the same pixel coordinate range is extracted from the detected image sequence and used as a detection pixel block; The reference pixel block and the detection pixel block are respectively convolved by a multi-directional gradient filter bank to obtain the response amplitude of each pixel in multiple gradient directions; For each pixel, the gradient direction with the largest response amplitude is selected as the main texture direction of the pixel; The distribution of the main texture direction of all pixels in the reference pixel block and the detection pixel block in different directional intervals is statistically analyzed to form a texture direction distribution histogram that reflects the concentration of texture direction.

5. The method for visually detecting thermal shock cracks in an industrial robot protective cover according to claim 4, characterized in that, The similarity measurement of the texture direction distribution histogram is performed to identify pixel blocks whose texture direction distribution consistency is lower than a preset standard, and these blocks are removed from the potential abnormal regions, including: The texture direction distribution histograms of the reference pixel block and the detection pixel block are smoothed to eliminate statistical fluctuations caused by noise. Calculate the numerical difference between the two texture direction distribution histograms after smoothing in each direction interval, and take the sum of the squares of the numerical differences in all direction intervals as the distribution difference degree; The distribution difference is compared with a texture consistency threshold obtained in advance from training samples; When the distribution difference is greater than the texture consistency threshold, it is determined that the potential abnormal region corresponding to the pixel block is caused by surface stains or changes in lighting, rather than cracks, and the abnormal region is removed from the set of potential abnormal regions. All potential abnormal regions whose distribution difference is less than or equal to the texture consistency threshold are retained to form the crack candidate region set.

6. The method for visually inspecting thermal shock cracks in an industrial robot protective cover according to claim 1, characterized in that, Based on the set of candidate crack regions, suspected crack regions are obtained, including: Calculate the aspect ratio of the minimum bounding rectangle for each candidate region in the set of crack candidate regions; Set an aspect ratio threshold, filter out candidate regions with aspect ratios greater than the threshold, and form a set of slender candidate regions; For each region in the set of elongated candidate regions, grayscale morphological closing operations are performed on the corresponding original image region in the detection image sequence using multiple linear structuring elements of different sizes. Compare the average grayscale change rate of the image region after performing a closing operation on linear structuring elements of different sizes; Candidate regions whose average grayscale change rate exceeds a preset change rate threshold are identified as suspected crack regions. Record the minimum bounding rectangle parameters of the suspected crack region and its position coordinates in the detection image sequence.

7. A visual inspection method for thermal shock cracks in an industrial robot protective cover according to claim 6, characterized in that, For each region in the set of elongated candidate regions, grayscale morphological closing operations are performed on the corresponding original image region in the detected image sequence using multiple linear structuring elements of different sizes, including: Extract the original grayscale image block corresponding to the elongated candidate region from the detected image sequence; Prepare a set of linear structural elements with increasing length but fixed width, wherein the length direction of the linear structural elements is consistent with their preset traversal direction in the image block; Each of the linear structuring elements is used sequentially to perform a grayscale morphological closing operation on the original grayscale image block. The grayscale morphological closing operation includes performing a dilation operation followed by an erosion operation. After each closing operation is completed, the average gray value of the entire image block is calculated; Plot the relationship curve between the length of the linear structuring element used and the average gray value of the image block obtained after the operation.

8. A visual inspection method for thermal shock cracks in an industrial robot protective cover according to claim 6, characterized in that, The edge continuity verification of the suspected crack area includes: For each suspected crack region, an adaptive threshold edge detection operator is applied at the original image position of its corresponding detection image sequence to extract a set of clear edge pixels. The extracted set of edge pixels is then tracked and connected based on orientation consistency. Calculate the total length, cumulative curvature change, and straight-line distance between endpoints of each edge curve formed after connection; The continuity confidence score of the edge curve is calculated based on the ratio of the total length to the straight-line distance between the endpoints and the cumulative value of the curvature change. Suspected crack areas with a continuity confidence score lower than the preset continuity standard are excluded, and areas that meet the continuity requirements are retained to form a final list of confirmed crack areas.

9. A visual inspection method for thermal shock cracks in an industrial robot protective cover according to claim 8, characterized in that, The step of performing edge pixel tracking and connection based on orientation consistency on the extracted set of edge pixels includes: Arbitrarily select an unvisited edge pixel from the set of edge pixels as the current seed point; Check if there are any neighboring pixels belonging to the set of edge pixels within the eight-neighborhood of the current seed point; If there is a neighboring pixel that satisfies the orientation consistency constraint, then the neighboring pixel is connected to the current edge chain and updated as the new current seed point. The orientation consistency constraint requires that the difference between the gradient direction of the newly added pixel and the average direction of the current edge chain is less than the angle tolerance. Repeat the checking and connection process until the current edge chain can no longer be extended; Select a new seed point from the remaining set of edge pixels and start the tracing and connection of the next edge chain until all edge pixels have been visited, forming several edge curves.

10. A visual inspection method for thermal shock cracks in an industrial robot protective cover according to claim 8, characterized in that, The method further includes the step of quantifying and annotating the finally confirmed list of crack regions: For each crack region in the final confirmed crack region list, its pixel range is located in the detection image sequence; Calculate the pixel area, principal axis length, principal axis direction, and contrast between the average gray level inside the region and the average gray level of the surrounding background for each crack region. The calculated pixel area, principal axis length, principal axis direction, and contrast are stored as the quantized feature vector of the crack region. On the original sequence of detected images, the boundaries of each crack region are drawn as highlighted outlines; The quantized feature vector is associated with and stored with the corresponding crack region boundary image to generate a detection report file containing complete crack location and morphology information.

Images