CN110738642A - Mask R-CNN-based reinforced concrete crack identification and measurement method and storage medium - Google Patents
Mask R-CNN-based reinforced concrete crack identification and measurement method and storage medium Download PDFInfo
- Publication number
- CN110738642A CN110738642A CN201910949294.5A CN201910949294A CN110738642A CN 110738642 A CN110738642 A CN 110738642A CN 201910949294 A CN201910949294 A CN 201910949294A CN 110738642 A CN110738642 A CN 110738642A
- Authority
- CN
- China
- Prior art keywords
- crack
- mask
- edge
- operator
- picture
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/13—Edge detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/60—Analysis of geometric attributes
- G06T7/62—Analysis of geometric attributes of area, perimeter, diameter or volume
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Quality & Reliability (AREA)
- Geometry (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to an reinforced concrete crack identification and measurement method based on Mask R-CNN, which comprises the following steps of obtaining a picture to be detected, searching and positioning a crack region on the picture to be detected based on the Mask R-CNN, segmenting a crack target on the picture to be detected, performing edge detection on the crack target through combination of a Sobel edge filter operator, a Laplacian filter operator and a gaussian smoothing filter to generate a crack Mask, and calculating the length, the width and the crack of the crack according to the pole coordinate of the generated crack Mask and the minimum circumscribed rectangle of the Mask.
Description
Technical Field
The invention relates to the technical field of concrete crack identification, in particular to reinforced concrete crack identification and measurement methods based on Mask R-CNN and a storage medium.
Background
Crack detection is therefore critical in ensuring the safety of reinforced concrete structures, and in view of this, automatic image-based crack detection has recently attracted -wide research interest as a technique to overcome the safety inspection scheme of reinforced concrete structures.
genetic algorithms based on Genetic Programming (GP) and a seepage model are proposed by Zhong Qu in 2019, the method comprises the three steps of firstly extracting cracks through an image processing model of the GP in advance, secondly calculating crack tips after a crack skeleton is extracted, accurately detecting small-width cracks by using high-speed and high-precision seepage with end points as anchor points, scanning a fracture unit region for connection, finally communicating the pre-extracted cracks with the cracks detected by seepage to remove a quality interference region to obtain real cracks on the surface of concrete, deep Convolutional Neural Networks (CNN) called DeepackCr in 2019 are proposed by Yahui Liu et al and used for predicting crack segmentation in an end-to-end method in a pixel mode, deep learning frames based on a Convolutional Neural Network (CNN) and a fusion scheme are proposed by Chen et al in 2018, 3982 deep learning frames based on a Convolutional Neural Network (CNN) and a plain data fusion scheme are proposed by Cha in 2017, methods based on a convolutional neural network (Ns) are used for detecting deep crack structures and the crack defects of concrete crack detection without calculating the depth of the crack detection, 2016 and 2016 (2016) for automatically detecting defects of the crack detection.
Although many studies have proposed various crack detection techniques and algorithms, the efficiency and accuracy of the algorithms need to be improved, and there are fewer studies of the algorithms involving crack detection in combination with dimensional measurement.
Disclosure of Invention
Therefore, reinforced concrete crack identification and measurement methods and storage media based on Mask R-CNN are needed to be provided, and the problems that the efficiency and the detection precision are low and the algorithm research combining crack detection and size measurement is less in the existing crack detection technology are solved.
In order to achieve the purpose, the inventor provides reinforced concrete crack identification and measurement methods based on Mask R-CNN, which comprise the following steps:
acquiring a picture to be detected, wherein the picture to be detected is a picture with a concrete crack;
searching and positioning a crack region on a picture to be detected based on Mask R-CNN, and segmenting a crack target on the picture to be detected;
performing edge detection on the crack target by combining a Sobel edge filtering operator, a Laplacian filtering operator and a gaussian smoothing filter to generate a crack mask;
and calculating the length, the width and the crack of the crack according to the pole coordinate of the generated crack mask and the minimum circumscribed rectangle of the mask.
, optimizing, wherein the 'performing edge detection on the crack target by combining a Sobel edge filter operator, a Laplacian filter operator and a gaussian smoothing filter to generate a crack mask' specifically comprises the following steps:
defining a convolution template of the Sobel edge filter operator in four directions, a convolution template of the laplacian edge filter operator and a convolution template of the gaussian smoothing filter, and respectively converting the convolution templates into convolution kernels of 3 x 3;
splicing the Sobel edge filtering operator and the laplacian edge filtering operator into edge filters;
performing Gaussian smoothing on the crack target through a gaussian smoothing filter, using the crack target as the input of an edge filter, and performing convolution through a Sobel edge filtering operator and a laplacian edge filtering operator to predict an edge;
performing Gaussian smoothing on the crack target through a Gaussian smoothing filter, using the smooth as the input of the GroudTruth, and performing convolution through a Sobel edge filter operator and a laplacian edge filter operator to obtain the GroudTruth;
the mean square error between the predicted edge and the GroudTruth is used to determine the mask edge.
And , optimizing, wherein the Sobel edge filter operator is an eight-direction Sobel edge filter operator.
, optimizing, wherein the step of searching and positioning the crack region on the picture to be detected based on Mask R-CNN specifically comprises the following steps:
independently extracting frame information of a crack target of a picture to be detected;
combining pyramid characteristics, inputting the length and width of a picture, and performing pyramid ROI processing;
and aligning by ROI to obtain a crack target.
And , optimizing, wherein the basic backbone network of the Mask R-CNN is ResNet152 or MobileNet.
The inventor also provides another technical solutions, wherein the storage media store computer programs, and when the computer programs are executed by a processor, the computer programs execute the following steps:
acquiring a picture to be detected, wherein the picture to be detected is a picture with a concrete crack;
searching and positioning a crack region on a picture to be detected based on Mask R-CNN, and segmenting a crack target on the picture to be detected;
performing edge detection on the crack target by combining a Sobel edge filtering operator, a Laplacian filtering operator and a gaussian smoothing filter to generate a crack mask;
and calculating the length, the width and the crack of the crack according to the pole coordinate of the generated crack mask and the minimum circumscribed rectangle of the mask.
, optimizing, wherein when the computer program is run by the processor to execute the step of performing edge detection on the crack target through combination of a Sobel edge filter operator, a Laplacian filter operator and a gaussian smoothing filter to generate a crack mask, the following steps are specifically executed:
defining a convolution template of the Sobel edge filter operator in four directions, a convolution template of the laplacian edge filter operator and a convolution template of the gaussian smoothing filter, and respectively converting the convolution templates into convolution kernels of 3 x 3;
splicing the Sobel edge filtering operator and the laplacian edge filtering operator into edge filters;
performing Gaussian smoothing on the crack target through a gaussian smoothing filter, using the crack target as the input of an edge filter, and performing convolution through a Sobel edge filtering operator and a laplacian edge filtering operator to predict an edge;
performing Gaussian smoothing on the crack target through a Gaussian smoothing filter, using the smooth as the input of the GroudTruth, and performing convolution through a Sobel edge filter operator and a laplacian edge filter operator to obtain the GroudTruth;
the mean square error between the predicted edge and the GroudTruth is used to determine the mask edge.
And , optimizing, wherein the Sobel edge filter operator is an eight-direction Sobel edge filter operator.
, optimizing, wherein when the computer program is executed by the processor to execute the step of searching and positioning the crack region on the picture to be detected based on Mask R-CNN, the following steps are executed:
independently extracting frame information of a crack target of a picture to be detected;
combining pyramid characteristics, inputting the length and width of a picture, and performing pyramid ROI processing;
and aligning by ROI to obtain a crack target.
And , optimizing, wherein the basic backbone network of the Mask R-CNN is ResNet152 or MobileNet.
Different from the prior art, according to the technical scheme, a crack region in an image to be detected is searched and positioned based on Mask R-CNN to obtain a crack target, edge detection is performed on the crack target through combination of a Sobel edge filtering operator, a Laplacian filtering operator and a gaussian smoothing filter to generate a crack Mask, detection of the crack is achieved, and meanwhile the width, the length and the area of the crack are calculated. Compared with other technologies, the method has relatively high efficiency in measuring the crack, more accurate extraction of crack pixels, and realization of crack detection and identification and calculation of the length, width and area of the crack.
Drawings
FIG. 1 is a schematic view of kinds of processes of a method for identifying and measuring a reinforced concrete crack based on Mask R-CNN according to an embodiment;
FIG. 2 is a diagram illustrating structures of Mask R-CNN model according to an embodiment;
FIG. 3 is a schematic diagram of structures of an FPN structure according to an embodiment;
FIG. 4a is a schematic representation of configurations of longitudinal slits according to an embodiment;
FIG. 4b is another schematic diagrams of longitudinal slits according to embodiments;
FIG. 5 is a schematic representation of structures of the reticulated seam of an embodiment;
FIG. 6 is a diagram illustrating types of structures of a block crack according to an embodiment.
Detailed Description
To explain technical contents, structural features, and objects and effects of the technical solutions in detail, the following detailed description is given with reference to the accompanying drawings in conjunction with the embodiments.
Referring to fig. 1, the embodiment provides reinforced concrete crack identification and measurement methods based on Mask R-CNN, including the following steps:
step S110: acquiring a picture to be detected, wherein the picture to be detected is a picture with a concrete crack;
step S120: searching and positioning a crack region on a picture to be detected based on Mask R-CNN, and segmenting a crack target on the picture to be detected;
step S130: performing edge detection on the crack target by combining a Sobel edge filtering operator, a Laplacian filtering operator and a gaussian smoothing filter to generate a crack mask;
step S140: and calculating the length, the width and the crack of the crack according to the pole coordinate of the generated crack mask and the minimum circumscribed rectangle of the mask.
The Mask R-CNN is an extension of the Faster R-CNN, can synchronously perform semantic segmentation, task segmentation, positioning and classification tasks, and work of extracting characteristic regions and generating masks, and is a Mask R-CNN model shown in FIG. 2, and has the following maximum characteristics: independently extracting frame information of a crack target of a picture to be detected; combining pyramid characteristics, inputting the length and width of a picture, and performing pyramid ROI processing; and aligning by ROI to obtain a crack target.
In the embodiment, Mask R-CNN adopts ResNet101, ResNet152 and MobileNet to combine with a main network of FPN, and improves example segmentation performance of crack target features by fusing multilayer crack feature information. The ResNet152 training model memory is 60M, the ResNet101 training model memory is 245M, and the MobileNet training model memory is 93M, which shows that the MobileNet model is small in size and easy to transplant to mobile equipment, but the accuracy is not high enough when overlapped target detection is carried out. In order to enhance the example segmentation effect of the target, the Mask R-CNN adopts ResNet101 and ResNet152 according to the principle that the deeper network effect is better, and the segmentation detection of the ResNet101 and the ResNet152 is more accurate.
The method comprises the steps of performing segmentation processing on targets with different sizes in a visual recognition field image to obtain basic challenges, and segmenting a crack target by adopting a scheme of constructing a multi-scale pyramid by MaskR-CNN, wherein in the detection of the crack target, the size of the crack is unpredictable, the segmentation result of the crack target with larger size is output in the 1 st Layer of the FPN and the segmentation result of the crack target with the next larger size is output in the 2 nd Layer by adopting an FPN (characteristic pyramid), and the segmentation result of the crack target with smaller size is output in the 3 rd Layer by adopting an FPN (characteristic pyramid). when a crack target example is segmented by using a Mask R-CNN model, firstly, inputting a crack image into a ResNet depth residual error network for processing, performing characteristic up-sampling operation from Layer4 correspondingly after deep convolution operation, performing dimension reduction operation by adding a Layer of Layer 1 on a Layer, performing semantic down-sampling operation on a rolling Layer, adding Layer 3656 and Layer 4656 as feature element pairs, performing semantic addition operation to obtain a result, and providing low-level sampling information, so that the characteristic position detection error can be accurately detected by comparing the low-sampling operation with a high-level sampling network, and providing high-level sampling error detection efficiency detection method, which can be more accurate and can be achieved by adding the low-positioning method.
The Mask R-CNN also uses an ROI Align layer for correctly aligning extracted features with input, the ROI Align layer plays a role in accurately positioning ship targets, and is very important in a Mask R-CNN system, CNNs are connected behind the ROI-Align layer, CNNs are connected behind the ROI-Align layer for extracting input features for classification and Mask generation, 7 × 7 alignment feature maps are obtained from the ROI-Align layer by the target classification, 14 × 14 alignment feature maps are obtained from the ROI Align layer by the Mask generation, four constituent convolution layers and two completely connected convolution layers are connected behind the ROI-Align layer, and Relu activation and maximum pooling sampling are carried out after each convolution layer, so that Mask edge detection branches are added after Mask R-CNN generation, and the problem of complete coverage of crack edges can be effectively solved.
According to the characteristics of Mask R-CNN, after a crack target is segmented, masks (masks) are punched on the crack target, but an original Mask R-CNN model can generate incomplete or worse masks, and particularly the phenomenon is serious in early trainingMask-Rcnn＝Lcls+Lbox+Lmask+Ledge(ii) a Wherein L iscls、Lbox、Lmask、LedgeThe method comprises a classification loss function, a prediction frame loss function, a mask generation loss function and an edge detection loss function.
Since a lot of information is lost when a real-world three-dimensional space is mapped to a two-dimensional space displayed by an image, and interference of parts like illumination, scenes and the like is also added, when a crack is identified, the crack edge is mainly searched, which is that characteristics which are more used in image identification are used, a place with severe gray value change can be considered as an edge, a Sobel operator makes own specification on edge detection, and the Sobel operator mainly has the function of processing edge information input and output of images, the traditional Sobel edge filtering operator comprises convolution templates in horizontal and vertical directions, and has a good effect on edge detection in two directions, but the convolution template of the Sobel edge filtering operator is expanded into 4 to 8 directions due to unpredictable directions of the crack, and the following definitions are respectively:
aiming at the irregularity of the crack image, the horizontal direction of the original sobel algorithm is 0 degree (S)0°) And the vertical direction 90 ° (S)90°) On the basis of increasing 45 degrees of direction (S)45°) 135 degree direction (S)135°) 180 deg. direction (S)180°) 225 ° direction (S)225°) 227 deg. direction (S)227°) 315 deg. direction (S)315°) And the edge positioning precision is improved.
Since the Laplace operator is the simplest isotropic differential operator in all directions, it has an edge detection operator independent of the edge direction, with which the image can be sharpened, enhancing the crack edge contrast.
The difference between the second derivatives of Laplace operator in x and y directions is defined as follows:
the difference form of the Laplace operator is obtained from the above two formulas as follows:
the form that can be converted into a Laplacian template is as follows:
in order to realize full-angle correspondence of the Laplacian template, the form of the Laplacian template is rewritten into expansion templates, and the form is defined as follows:
in order to reduce the influence of noise on the edge detection result as much as possible, it is necessary to filter out the noise to prevent erroneous detection caused by the noise. To smooth the image, a gaussian filter is used to convolve with the image, which smoothes the image to reduce the effects of noise apparent on the edge detector-here, a two-dimensional gaussian distribution function is used to compute the gaussian convolution operator template. The two-dimensional gaussian distribution function equation is defined as follows:
substituting the gaussian template into the two-dimensional gaussian distribution function equation to obtain a gaussian filter G with a size of 3 × 3, which is defined as follows:
based on the above basis, a weighted gaussian filter is constructed, and since the sum of the weight coefficients must be 1, we perform a normalization process on the gaussian filter G to obtain:
and further obtaining a Gaussian convolution operator template for smoothing the image.
In order to calculate the final loss rate, network branches are added in the mask R-CNN, which are called edge detection protocol branches, and the step of performing edge detection on a crack target by combining a Sobel edge filter operator, a Laplacian filter operator and a gaussian smoothing filter to generate a crack mask specifically comprises the following steps:
defining a convolution template of the Sobel edge filter operator in four directions, a convolution template of the laplacian edge filter operator and a convolution template of the gaussian smoothing filter, and respectively converting the convolution templates into convolution kernels of 3 x 3;
splicing the Sobel edge filtering operator and the laplacian edge filtering operator into edge filters;
performing Gaussian smoothing on the crack target through a gaussian smoothing filter, using the crack target as the input of an edge filter, and performing convolution through a Sobel edge filtering operator and a laplacian edge filtering operator to predict an edge;
performing Gaussian smoothing on the crack target through a Gaussian smoothing filter, using the smooth as the input of the GroudTruth, and performing convolution through a Sobel edge filter operator and a laplacian edge filter operator to obtain the GroudTruth;
the mean square error between the predicted edge and the GroudTruth is used to determine the mask edge.
According to the above steps, the generated edge detection loss function can be positioned as follows:
LP(yP,yt,P)＝KMean(yP-yt)Pwherein y isPAs edge detection value, ytIs groupThe value of dTruth, P is power, P is 2, and the missing regions of the mask edge can be reduced by applying gaussian smoothing, which has higher accuracy in the overall coverage of the target mask edge.
The damage level of the crack to the reinforced concrete structure is determined by the length, width or occupied area of the crack. And generating the pole coordinates of the crack Mask and the minimum Mask circumscribed rectangle through the optimized Mask R-CNN model, and calculating the length, width and area of the crack. The damage degree is defined as 3 grades, and the damage degree is used for determining the damage degree of the crack and can be used as an important basis for measuring the safety grade of the building. Wherein, the length measurement is to count the total number of pixels of the crack Mask generated by Mask R-CNN, and the crack length measurement is defined as: l ═ λ · Pl(ii) a Where L is the crack length, λ is the side length of a single pixel, PlThe length of the minimum circumscribed rectangle of the crack mask is obtained, wherein the side length of a single pixel is 0.001mm, and the crack length measurement formula is mainly used for calculating the length of the crack with the longitudinal direction and the transverse direction. And measuring the width, taking the maximum coordinate and the minimum coordinate of the mask pixel points according to the generated coordinates of the crack mask pixel points, calculating the linear distance between two polar coordinates to obtain the width of the crack, and taking a polar left coordinate point C1 and a polar right coordinate point Cr of the crack mask to obtain a polar left polar distance PWSpecifically defined as: pW＝|Cr-C1L, |; and defining the side length of the pixel as lambda to obtain the crack width W, which is specifically defined as: w ═ λ · PW. As shown in fig. 4a, the polar coordinates of the longitudinal crack are (497,724), (516,929) and the polar coordinates of the crack mask are (668.5,724), (708.5,929), and the width of the longitudinal crack is the difference of the abscissa between two coordinate points, i.e. the crack width, so that W1 is 0.17mm and W2 is 0.19mm, and in order to show early warning on crack control in building safety, a larger value W2 is selected as the crack width of the crack mask in fig. 4a, and similarly, the crack width W3 in the longitudinal crack in fig. 4b can be calculated to be 0.16 mm.
If the slits are net-like or block-like, slits are shown in FIG. 5The method comprises the steps of determining the harmfulness of a crack by calculating the area of a minimum external rectangle of the crack, wherein the coordinates of the upper left corner of the minimum external rectangle in FIG. 5 are (0,22), the coordinates of the lower right corner of the minimum external rectangle are (595,455), the coordinates of the two points are the coordinates of the vertex of the minimum external rectangle of a crack target Mask generated by a Mask R-CNN model, the length and the width of the rectangle are respectively obtained by the distance difference between the two coordinates, so that the area of the external rectangle is calculated, and a measurement formula is defined as: s ═ λ (| X)2-X1|)·λ·(|Y2-Y1I)); wherein, the lambda is the side length of a single pixel and takes the value of 0.001mm, and X2Is the abscissa of the lower right corner, X1Is the abscissa of the upper left corner, Y2Is the ordinate of the lower right corner, Y1The upper left ordinate, whereby the minimum circumscribed rectangular area S of the reticular fracture in FIG. 5 was calculated to be 0.26mm2. The block crack shown in FIG. 6 has the coordinates of the minimum bounding rectangle with the upper left corner (32,26) and the lower right corner (1024,984), and the minimum bounding rectangle area of the block crack shown in FIG. 6 is calculated to be 0.95mm2。
The inventor also provides another technical solutions, wherein the storage media store computer programs, and when the computer programs are executed by a processor, the computer programs execute the following steps:
acquiring a picture to be detected, wherein the picture to be detected is a picture with a concrete crack;
searching and positioning a crack region on a picture to be detected based on Mask R-CNN, and segmenting a crack target on the picture to be detected;
performing edge detection on the crack target by combining a Sobel edge filtering operator, a Laplacian filtering operator and a gaussian smoothing filter to generate a crack mask;
and calculating the length, the width and the crack of the crack according to the pole coordinate of the generated crack mask and the minimum circumscribed rectangle of the mask.
The Mask R-CNN is an extension of the Faster R-CNN, can synchronously perform semantic segmentation, task segmentation, positioning and classification tasks, and work of extracting characteristic regions and generating masks, and is a Mask R-CNN model shown in FIG. 2, and has the following maximum characteristics: independently extracting frame information of a crack target of a picture to be detected; combining pyramid characteristics, inputting the length and width of a picture, and performing pyramid ROI processing; and aligning by ROI to obtain a crack target.
In the embodiment, Mask R-CNN adopts ResNet101, ResNet152 and MobileNet to combine with a main network of FPN, and improves example segmentation performance of crack target features by fusing multilayer crack feature information. The ResNet152 training model memory is 60M, the ResNet101 training model memory is 245M, and the MobileNet training model memory is 93M, which shows that the MobileNet model is small in size and easy to transplant to mobile equipment, but the accuracy is not high enough when overlapped target detection is carried out. In order to enhance the example segmentation effect of the target, the Mask R-CNN adopts ResNet101 and ResNet152 according to the principle that the deeper network effect is better, and the segmentation detection of the ResNet101 and the ResNet152 is more accurate.
The method comprises the steps of performing segmentation processing on targets with different sizes in a visual recognition field image to obtain basic challenges, and segmenting a crack target by adopting a scheme of constructing a multi-scale pyramid by MaskR-CNN, wherein in the detection of the crack target, the size of the crack is unpredictable, the segmentation result of the crack target with larger size is output in the 1 st Layer of the FPN and the segmentation result of the crack target with the next larger size is output in the 2 nd Layer by adopting an FPN (characteristic pyramid), and the segmentation result of the crack target with smaller size is output in the 3 rd Layer by adopting an FPN (characteristic pyramid). when a crack target example is segmented by using a Mask R-CNN model, firstly, inputting a crack image into a ResNet depth residual error network for processing, performing characteristic up-sampling operation from Layer4 correspondingly after deep convolution operation, performing dimension reduction operation by adding a Layer of Layer 1 on a Layer, performing semantic down-sampling operation on a rolling Layer, adding Layer 3656 and Layer 4656 as feature element pairs, performing semantic addition operation to obtain a result, and providing low-level sampling information, so that the characteristic position detection error can be accurately detected by comparing the low-sampling operation with a high-level sampling network, and providing high-level sampling error detection efficiency detection method, which can be more accurate and can be achieved by adding the low-positioning method.
The Mask R-CNN also uses an ROI Align layer for correctly aligning extracted features with input, and the ROIAlign layer plays a role in accurately positioning ship targets, and is very important in a Mask R-CNN system, CNNs are connected behind the ROI-Align layer, CNNs are connected behind the ROI-Align layer for extracting input features for classification and Mask generation, 7 × 7 alignment feature maps are obtained from the ROI-Align layer by the target classification, 14 × 14 alignment feature maps are obtained from the ROI Align layer by the Mask generation, four component volume layers and two completely connected volume layers are connected behind the ROI-Align layer, Relu activation and maximum ROI sampling are carried out behind each volume layer, Mask edge detection branches are added after Mask R-CNN generation Mask branches, and the problem of complete coverage of crack edges can be effectively solved.
According to the characteristics of Mask R-CNN, after a crack target is segmented, masks (masks) are punched on the crack target, but an original Mask R-CNN model can generate incomplete or worse masks, and particularly the phenomenon is serious in early trainingMask-Rcnn＝Lcls+Lbox+Lmask+Ledge(ii) a Wherein L iscls、Lbox、Lmask、LedgeThe method comprises a classification loss function, a prediction frame loss function, a mask generation loss function and an edge detection loss function.
Since a lot of information is lost when a real-world three-dimensional space is mapped to a two-dimensional space displayed by an image, and interference of parts like illumination, scenes and the like is also added, when a crack is identified, the crack edge is mainly searched, which is that characteristics which are more used in image identification are used, a place with severe gray value change can be considered as an edge, a Sobel operator makes own specification on edge detection, and the Sobel operator mainly has the function of processing edge information input and output of images, the traditional Sobel edge filtering operator comprises convolution templates in horizontal and vertical directions, and has a good effect on edge detection in two directions, but the convolution template of the Sobel edge filtering operator is expanded into 4 to 8 directions due to unpredictable directions of the crack, and the following definitions are respectively:
aiming at the irregularity of the crack image, the horizontal direction of the original sobel algorithm is 0 degree (S)0And 90 DEG (S) from the vertical90°) On the basis of increasing 45 degrees of direction (S)45°) 135 degree direction (S)135°) 180 deg. direction (S)180°) 225 ° direction (S)225°) 227 deg. direction (S)227°) 315 deg. direction (S)315°) And the edge positioning precision is improved.
Since the Laplace operator is the simplest isotropic differential operator in all directions, it has an edge detection operator independent of the edge direction, with which the image can be sharpened, enhancing the crack edge contrast.
The difference between the second derivatives of Laplace operator in x and y directions is defined as follows:
the difference form of the Laplace operator is obtained from the above two formulas as follows:
the form that can be converted into a Laplacian template is as follows:
in order to realize full-angle correspondence of the Laplacian template, the form of the Laplacian template is rewritten into expansion templates, and the form is defined as follows:
in order to reduce the influence of noise on the edge detection result as much as possible, it is necessary to filter out the noise to prevent erroneous detection caused by the noise. To smooth the image, a gaussian filter is used to convolve with the image, which smoothes the image to reduce the effects of noise apparent on the edge detector-here, a two-dimensional gaussian distribution function is used to compute the gaussian convolution operator template. The two-dimensional gaussian distribution function equation is defined as follows:
substituting the gaussian template into the two-dimensional gaussian distribution function equation to obtain a gaussian filter G with a size of 3 × 3, which is defined as follows:
based on the above basis, a weighted gaussian filter is constructed, and since the sum of the weight coefficients must be 1, we perform a normalization process on the gaussian filter G to obtain:
and further obtaining a Gaussian convolution operator template for smoothing the image.
In order to calculate the final loss rate, network branches are added in the mask R-CNN, which are called edge detection protocol branches, and the step of performing edge detection on a crack target by combining a Sobel edge filter operator, a Laplacian filter operator and a gaussian smoothing filter to generate a crack mask specifically comprises the following steps:
defining a convolution template of the Sobel edge filter operator in four directions, a convolution template of the laplacian edge filter operator and a convolution template of the gaussian smoothing filter, and respectively converting the convolution templates into convolution kernels of 3 x 3;
splicing the Sobel edge filtering operator and the laplacian edge filtering operator into edge filters;
performing Gaussian smoothing on the crack target through a gaussian smoothing filter, using the crack target as the input of an edge filter, and performing convolution through a Sobel edge filtering operator and a laplacian edge filtering operator to predict an edge;
performing Gaussian smoothing on the crack target through a Gaussian smoothing filter, using the smooth as the input of the GroudTruth, and performing convolution through a Sobel edge filter operator and a laplacian edge filter operator to obtain the GroudTruth;
the mean square error between the predicted edge and the GroudTruth is used to determine the mask edge.
According to the above steps, the generated edge detection loss function can be positioned as follows:
LP(yP,yt,P)＝KMean(yP-yt)Pwherein y isPAs edge detection value, ytIs the GroudTruth value, P is the power, and P is 2, and Gaussian smoothing can be appliedThe missing areas of the mask edge are reduced, with higher accuracy in the overall target mask edge coverage.
The damage level of the crack to the reinforced concrete structure is determined by the length, width or occupied area of the crack. And generating the pole coordinates of the crack Mask and the minimum Mask circumscribed rectangle through the optimized Mask R-CNN model, and calculating the length, width and area of the crack. The damage degree is defined as 3 grades, and the damage degree is used for determining the damage degree of the crack and can be used as an important basis for measuring the safety grade of the building. Wherein, the length measurement is to count the total number of pixels of the crack Mask generated by Mask R-CNN, and the crack length measurement is defined as: l ═ λ · Pl(ii) a Where L is the crack length, λ is the side length of a single pixel, PlThe length of the minimum circumscribed rectangle of the crack mask is obtained, wherein the side length of a single pixel is 0.001mm, and the crack length measurement formula is mainly used for calculating the length of the crack with the longitudinal direction and the transverse direction. And measuring the width, taking the maximum coordinate and the minimum coordinate of the mask pixel points according to the generated coordinates of the crack mask pixel points, calculating the linear distance between two polar coordinates to obtain the width of the crack, and taking a polar left coordinate point C1 and a polar right coordinate point Cr of the crack mask to obtain a polar left polar distance PWSpecifically defined as: pW＝|Cr-C1L, |; and defining the side length of the pixel as lambda to obtain the crack width W, which is specifically defined as: w ═ λ · PW. As shown in fig. 4a, the polar coordinates of the crack mask of the longitudinal crack are (497,724), (516,929), and the polar coordinates of the crack mask are (668.5,724), (708.5,929), and the width is the difference of the abscissa between two coordinate points, that is, the crack width, because the crack is a longitudinal crack, W1 is 0.17mm, and W2 is 0.19mm, in order to show early warning for crack control in building safety, a larger value W2 is selected as the crack width of the crack mask in fig. 4a, and similarly, the crack width W3 in the longitudinal crack in fig. 4b can be calculated to be 0.16 mm.
If the cracks are net-like or block-like, such as net-like cracks in FIG. 5, we calculate the minimum of cracksThe area of the external rectangle is used for determining the damage degree of the crack, the coordinates of the upper left corner of the minimum external rectangle in fig. 5 are (0,22), the coordinates of the lower right corner are (595,455), the coordinates of the two points are the vertex coordinates of the minimum external rectangle of the crack target Mask generated by the Mask R-CNN model, the length and the width of the rectangle are respectively obtained according to the distance difference between the two coordinates, so that the area of the external rectangle is calculated, and the measurement formula is defined as: s ═ λ (| X)2-X1|)·λ·(|Y2-Y1I)); wherein, the lambda is the side length of a single pixel and takes the value of 0.001mm, and X2Is the abscissa of the lower right corner, X1Is the abscissa of the upper left corner, Y2Is the ordinate of the lower right corner, Y1The upper left ordinate, whereby the minimum circumscribed rectangular area S of the network of cracks in fig. 5 was calculated to be 0.26mm 2. For the block crack shown in FIG. 6, the coordinates of the top left corner and the bottom right corner of the minimum bounding rectangle are (32,26) and (1024,984), respectively, and the area of the minimum bounding rectangle of the block crack shown in FIG. 6 is calculated to be 0.95mm 2. It should be noted that, although the above embodiments have been described herein, the invention is not limited thereto. Therefore, based on the innovative concepts of the present invention, the technical solutions of the present invention can be directly or indirectly applied to other related technical fields by making changes and modifications to the embodiments described herein, or by using equivalent structures or equivalent processes performed in the content of the present specification and the attached drawings, which are included in the scope of the present invention.
Claims (10)
1, reinforced concrete crack identification and measurement methods based on Mask R-CNN, characterized by comprising the following steps:
acquiring a picture to be detected, wherein the picture to be detected is a picture with a concrete crack;
searching and positioning a crack region on a picture to be detected based on Mask R-CNN, and segmenting a crack target on the picture to be detected;
performing edge detection on the crack target by combining a Sobel edge filtering operator, a Laplacian filtering operator and a gaussian smoothing filter to generate a crack mask;
and calculating the length, the width and the crack of the crack according to the pole coordinate of the generated crack mask and the minimum circumscribed rectangle of the mask.
2. The method for identifying and measuring the reinforced concrete crack based on the Mask R-CNN as claimed in claim 1, wherein the step of performing edge detection on the crack target by combining a Sobel edge filter operator, a Laplacian filter operator and a gaussian smoothing filter to generate the crack Mask specifically comprises the following steps:
defining a convolution template of the Sobel edge filter operator in four directions, a convolution template of the laplacian edge filter operator and a convolution template of the gaussian smoothing filter, and respectively converting the convolution templates into convolution kernels of 3 x 3;
splicing the Sobel edge filtering operator and the laplacian edge filtering operator into edge filters;
performing Gaussian smoothing on the crack target through a gaussian smoothing filter, using the crack target as the input of an edge filter, and performing convolution through a Sobel edge filtering operator and a laplacian edge filtering operator to predict an edge;
performing Gaussian smoothing on the crack target through a Gaussian smoothing filter, using the smooth as the input of the GroudTruth, and performing convolution through a Sobel edge filter operator and a laplacian edge filter operator to obtain the GroudTruth;
the mean square error between the predicted edge and the GroudTruth is used to determine the mask edge.
3. The Mask R-CNN-based reinforced concrete crack identification and measurement method according to claim 1, wherein the Sobel edge filter operator is an eight-direction Sobel edge filter operator.
4. The method for identifying and measuring the reinforced concrete crack based on the Mask R-CNN as claimed in claim 1, wherein the step of searching and positioning the crack region on the picture to be detected based on the Mask R-CNN specifically comprises the following steps:
independently extracting frame information of a crack target of a picture to be detected;
combining pyramid characteristics, inputting the length and width of a picture, and performing pyramid ROI processing;
and aligning by ROI to obtain a crack target.
5. The method for identifying and measuring the reinforced concrete cracks based on the Mask R-CNN as claimed in claim 1, wherein the basic main network of the Mask R-CNN is ResNet152 or MobileNet.
Storage medium , said storage medium storing a computer program, said computer program when executed by a processor performing the steps of:
acquiring a picture to be detected, wherein the picture to be detected is a picture with a concrete crack;
searching and positioning a crack region on a picture to be detected based on Mask R-CNN, and segmenting a crack target on the picture to be detected;
performing edge detection on the crack target by combining a Sobel edge filtering operator, a Laplacian filtering operator and a gaussian smoothing filter to generate a crack mask;
and calculating the length, the width and the crack of the crack according to the pole coordinate of the generated crack mask and the minimum circumscribed rectangle of the mask.
7. The storage medium of claim 6, wherein when the computer program is executed by the processor to perform the step of generating a crack mask by performing edge detection on a crack target through a combination of a Sobel edge filter operator, a Laplacian filter operator, and a gaussian smoothing filter, the following steps are specifically performed:
defining a convolution template of the Sobel edge filter operator in four directions, a convolution template of the laplacian edge filter operator and a convolution template of the gaussian smoothing filter, and respectively converting the convolution templates into convolution kernels of 3 x 3;
splicing the Sobel edge filtering operator and the laplacian edge filtering operator into edge filters;
performing Gaussian smoothing on the crack target through a gaussian smoothing filter, using the crack target as the input of an edge filter, and performing convolution through a Sobel edge filtering operator and a laplacian edge filtering operator to predict an edge;
performing Gaussian smoothing on the crack target through a Gaussian smoothing filter, using the smooth as the input of the GroudTruth, and performing convolution through a Sobel edge filter operator and a laplacian edge filter operator to obtain the GroudTruth;
the mean square error between the predicted edge and the GroudTruth is used to determine the mask edge.
8. The storage medium of claim 6, wherein the Sobel edge filter operator is an eight-way Sobel edge filter operator.
9. The storage medium according to claim 6, wherein when the computer program is executed by the processor to perform the step of searching and locating a crack region on the picture to be detected based on Mask R-CNN, the following steps are specifically performed:
independently extracting frame information of a crack target of a picture to be detected;
combining pyramid characteristics, inputting the length and width of a picture, and performing pyramid ROI processing;
and aligning by ROI to obtain a crack target.
10. The storage medium of claim 6, wherein the basic backbone network of Mask R-CNN is ResNet152 or MobileNet.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910949294.5A CN110738642A (en) | 2019-10-08 | 2019-10-08 | Mask R-CNN-based reinforced concrete crack identification and measurement method and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910949294.5A CN110738642A (en) | 2019-10-08 | 2019-10-08 | Mask R-CNN-based reinforced concrete crack identification and measurement method and storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110738642A true CN110738642A (en) | 2020-01-31 |
Family
ID=69268556
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910949294.5A Pending CN110738642A (en) | 2019-10-08 | 2019-10-08 | Mask R-CNN-based reinforced concrete crack identification and measurement method and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110738642A (en) |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111461110A (en) * | 2020-03-02 | 2020-07-28 | 华南理工大学 | Small target detection method based on multi-scale image and weighted fusion loss |
| CN111612786A (en) * | 2020-06-19 | 2020-09-01 | 国网湖南省电力有限公司 | Concrete defect detection method and device based on full convolution neural network and storage medium |
| CN111612787A (en) * | 2020-06-19 | 2020-09-01 | 国网湖南省电力有限公司 | Concrete crack high-resolution image lossless semantic segmentation method and device and storage medium |
| CN111724337A (en) * | 2020-03-05 | 2020-09-29 | 中冶赛迪重庆信息技术有限公司 | Cold bed top punching identification method and system, electronic equipment and medium |
| CN111932508A (en) * | 2020-07-31 | 2020-11-13 | 山东大学 | Steel bar size measuring method and system based on image processing |
| CN112037196A (en) * | 2020-08-31 | 2020-12-04 | 中冶赛迪重庆信息技术有限公司 | Cooling bed multiple-length detection method, system and medium |
| CN112053331A (en) * | 2020-08-28 | 2020-12-08 | 西安电子科技大学 | Bridge crack detection method based on image superposition and crack information fusion |
| CN112132884A (en) * | 2020-09-29 | 2020-12-25 | 中国海洋大学 | Sea cucumber length measuring method and system based on parallel laser and semantic segmentation |
| CN112509026A (en) * | 2020-11-06 | 2021-03-16 | 广东电网有限责任公司中山供电局 | Insulator crack length identification method |
| CN112686913A (en) * | 2021-01-11 | 2021-04-20 | 天津大学 | Object boundary detection and object segmentation model based on boundary attention consistency |
| CN113052106A (en) * | 2021-04-01 | 2021-06-29 | 重庆大学 | Airplane take-off and landing runway identification method based on PSPNet network |
| CN113240623A (en) * | 2021-03-18 | 2021-08-10 | 中国公路工程咨询集团有限公司 | Pavement disease detection method and device |
| CN113284107A (en) * | 2021-05-25 | 2021-08-20 | 重庆邮电大学 | Attention mechanism-induced improved U-net concrete crack real-time detection method |
| CN113392849A (en) * | 2021-06-30 | 2021-09-14 | 哈尔滨理工大学 | R-CNN-based complex pavement crack identification method |
| CN113409267A (en) * | 2021-06-17 | 2021-09-17 | 西安热工研究院有限公司 | Pavement crack detection and segmentation method based on deep learning |
| CN114763699A (en) * | 2022-05-23 | 2022-07-19 | 中建四局安装工程有限公司 | Embedded bolt positioning method, embedded bolt auxiliary fixing device and using method thereof |
| CN115393725A (en) * | 2022-10-26 | 2022-11-25 | 西南科技大学 | Bridge crack identification method based on feature enhancement and semantic segmentation |
| CN117078737A (en) * | 2023-10-17 | 2023-11-17 | 深圳市城市交通规划设计研究中心股份有限公司 | Linear crack length calculation method |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106910186A (en) * | 2017-01-13 | 2017-06-30 | 陕西师范大学 | A kind of Bridge Crack detection localization method based on CNN deep learnings |
| KR101772916B1 (en) * | 2016-12-30 | 2017-08-31 | 한양대학교 에리카산학협력단 | Device for measuring crack width of concretestructure |
| CN108492281A (en) * | 2018-03-06 | 2018-09-04 | 陕西师范大学 | A method of fighting Bridge Crack image detection of obstacles and the removal of network based on production |
| CN109978032A (en) * | 2019-03-15 | 2019-07-05 | 西安电子科技大学 | Bridge Crack detection method based on spatial pyramid cavity convolutional network |
-
2019
- 2019-10-08 CN CN201910949294.5A patent/CN110738642A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR101772916B1 (en) * | 2016-12-30 | 2017-08-31 | 한양대학교 에리카산학협력단 | Device for measuring crack width of concretestructure |
| CN106910186A (en) * | 2017-01-13 | 2017-06-30 | 陕西师范大学 | A kind of Bridge Crack detection localization method based on CNN deep learnings |
| CN108492281A (en) * | 2018-03-06 | 2018-09-04 | 陕西师范大学 | A method of fighting Bridge Crack image detection of obstacles and the removal of network based on production |
| CN109978032A (en) * | 2019-03-15 | 2019-07-05 | 西安电子科技大学 | Bridge Crack detection method based on spatial pyramid cavity convolutional network |
Non-Patent Citations (2)
| Title |
|---|
| LEANNE ATTARD ET AL.: "Automatic Crack Detection using Mask R-CNN", 《2019 11TH INTERNATIONAL SYMPOSIUM ON IMAGE AND SIGNAL PROCESSING AND ANALYSIS (ISPA)》 * |
| ROLAND S.ZIMMERMANN ET AL.: "Faster Training of Mask R-CNN by Focusing on Instance Boundaries", 《ARXIV:1809.07069V1》 * |
Cited By (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111461110A (en) * | 2020-03-02 | 2020-07-28 | 华南理工大学 | Small target detection method based on multi-scale image and weighted fusion loss |
| CN111461110B (en) * | 2020-03-02 | 2023-04-28 | 华南理工大学 | Small target detection method based on multi-scale image and weighted fusion loss |
| CN111724337A (en) * | 2020-03-05 | 2020-09-29 | 中冶赛迪重庆信息技术有限公司 | Cold bed top punching identification method and system, electronic equipment and medium |
| CN111612786A (en) * | 2020-06-19 | 2020-09-01 | 国网湖南省电力有限公司 | Concrete defect detection method and device based on full convolution neural network and storage medium |
| CN111612787A (en) * | 2020-06-19 | 2020-09-01 | 国网湖南省电力有限公司 | Concrete crack high-resolution image lossless semantic segmentation method and device and storage medium |
| CN111932508A (en) * | 2020-07-31 | 2020-11-13 | 山东大学 | Steel bar size measuring method and system based on image processing |
| CN111932508B (en) * | 2020-07-31 | 2023-10-10 | 山东大学 | Image processing-based steel bar size measurement method and system |
| CN112053331A (en) * | 2020-08-28 | 2020-12-08 | 西安电子科技大学 | Bridge crack detection method based on image superposition and crack information fusion |
| CN112053331B (en) * | 2020-08-28 | 2023-04-07 | 西安电子科技大学 | Bridge crack detection method based on image superposition and crack information fusion |
| CN112037196A (en) * | 2020-08-31 | 2020-12-04 | 中冶赛迪重庆信息技术有限公司 | Cooling bed multiple-length detection method, system and medium |
| CN112132884A (en) * | 2020-09-29 | 2020-12-25 | 中国海洋大学 | Sea cucumber length measuring method and system based on parallel laser and semantic segmentation |
| CN112509026A (en) * | 2020-11-06 | 2021-03-16 | 广东电网有限责任公司中山供电局 | Insulator crack length identification method |
| CN112686913A (en) * | 2021-01-11 | 2021-04-20 | 天津大学 | Object boundary detection and object segmentation model based on boundary attention consistency |
| CN113240623A (en) * | 2021-03-18 | 2021-08-10 | 中国公路工程咨询集团有限公司 | Pavement disease detection method and device |
| CN113240623B (en) * | 2021-03-18 | 2023-11-07 | 中国公路工程咨询集团有限公司 | Pavement disease detection method and device |
| CN113052106A (en) * | 2021-04-01 | 2021-06-29 | 重庆大学 | Airplane take-off and landing runway identification method based on PSPNet network |
| CN113284107A (en) * | 2021-05-25 | 2021-08-20 | 重庆邮电大学 | Attention mechanism-induced improved U-net concrete crack real-time detection method |
| CN113284107B (en) * | 2021-05-25 | 2022-10-11 | 重庆邮电大学 | Attention mechanism-introduced improved U-net concrete crack real-time detection method |
| CN113409267A (en) * | 2021-06-17 | 2021-09-17 | 西安热工研究院有限公司 | Pavement crack detection and segmentation method based on deep learning |
| CN113392849A (en) * | 2021-06-30 | 2021-09-14 | 哈尔滨理工大学 | R-CNN-based complex pavement crack identification method |
| CN114763699B (en) * | 2022-05-23 | 2022-11-08 | 中建四局安装工程有限公司 | Embedded bolt positioning method, embedded bolt auxiliary fixing device and using method thereof |
| CN114763699A (en) * | 2022-05-23 | 2022-07-19 | 中建四局安装工程有限公司 | Embedded bolt positioning method, embedded bolt auxiliary fixing device and using method thereof |
| CN115393725A (en) * | 2022-10-26 | 2022-11-25 | 西南科技大学 | Bridge crack identification method based on feature enhancement and semantic segmentation |
| CN115393725B (en) * | 2022-10-26 | 2023-03-07 | 西南科技大学 | Bridge crack identification method based on feature enhancement and semantic segmentation |
| CN117078737A (en) * | 2023-10-17 | 2023-11-17 | 深圳市城市交通规划设计研究中心股份有限公司 | Linear crack length calculation method |
| CN117078737B (en) * | 2023-10-17 | 2024-03-29 | 深圳市城市交通规划设计研究中心股份有限公司 | Linear crack length calculation method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110738642A (en) | Mask R-CNN-based reinforced concrete crack identification and measurement method and storage medium | |
| CN109544456B (en) | Panoramic environment sensing method based on two-dimensional image and three-dimensional point cloud data fusion | |
| CN109615611B (en) | Inspection image-based insulator self-explosion defect detection method | |
| Zhu et al. | Detection of large-scale concrete columns for automated bridge inspection | |
| WO2019042232A1 (en) | Fast and robust multimodal remote sensing image matching method and system | |
| CN110866871A (en) | Text image correction method and device, computer equipment and storage medium | |
| Dal Poz et al. | Automated extraction of road network from medium-and high-resolution images | |
| CN111062915A (en) | Real-time steel pipe defect detection method based on improved YOLOv3 model | |
| CN108230376B (en) | Remote sensing image processing method and device and electronic equipment | |
| CN107767456A (en) | A kind of object dimensional method for reconstructing based on RGB D cameras | |
| Kim et al. | Automated concrete crack evaluation using stereo vision with two different focal lengths | |
| CN111709387B (en) | Building segmentation method and system for high-resolution remote sensing image | |
| Xu et al. | A 3D reconstruction method for buildings based on monocular vision | |
| CN114677601A (en) | Dam crack detection method based on unmanned aerial vehicle inspection and combined with deep learning | |
| CN103700082B (en) | Image split-joint method based on dual quaterion relative orientation | |
| CN112381062A (en) | Target detection method and device based on convolutional neural network | |
| CN111932506A (en) | Method for extracting discontinuous straight line in image | |
| CN113705564B (en) | Pointer type instrument identification reading method | |
| CN114742864A (en) | Belt deviation detection method and device | |
| Mojidra et al. | Vision-based fatigue crack detection using global motion compensation and video feature tracking | |
| CN113989604A (en) | Tire DOT information identification method based on end-to-end deep learning | |
| Ebrahimikia et al. | True orthophoto generation based on unmanned aerial vehicle images using reconstructed edge points | |
| CN117227247A (en) | Intelligent positioning control method for carton processing | |
| CN116958837A (en) | Municipal facilities fault detection system based on unmanned aerial vehicle | |
| US20120275709A1 (en) | Building texture extracting apparatus and method thereof |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20200131 |