CN113506295B - Strip steel surface hot rolling slip defect detection method based on deep learning - Google Patents

Strip steel surface hot rolling slip defect detection method based on deep learning Download PDF

Info

Publication number
CN113506295B
CN113506295B CN202111058596.7A CN202111058596A CN113506295B CN 113506295 B CN113506295 B CN 113506295B CN 202111058596 A CN202111058596 A CN 202111058596A CN 113506295 B CN113506295 B CN 113506295B
Authority
CN
China
Prior art keywords
pixel point
rgb image
probability fluctuation
pixel
layer structure
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.)
Active
Application number
CN202111058596.7A
Other languages
Chinese (zh)
Other versions
CN113506295A (en
Inventor
王柱
吴恩旗
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
QIDONG HAIXIN MACHINERY CO Ltd
Original Assignee
QIDONG HAIXIN MACHINERY CO Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by QIDONG HAIXIN MACHINERY CO Ltd filed Critical QIDONG HAIXIN MACHINERY CO Ltd
Priority to CN202111058596.7A priority Critical patent/CN113506295B/en
Publication of CN113506295A publication Critical patent/CN113506295A/en
Application granted granted Critical
Publication of CN113506295B publication Critical patent/CN113506295B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/187—Segmentation; Edge detection involving region growing; involving region merging; involving connected component labelling
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/194—Segmentation; Edge detection involving foreground-background segmentation
    • 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
    • G06T2207/10—Image acquisition modality
    • G06T2207/10024—Color image
    • 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
    • G06T2207/20081—Training; Learning
    • 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]
    • 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
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30204—Marker

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a method for detecting hot rolling slippage defects on the surface of strip steel based on deep learning. The method comprises the steps of obtaining a receptive field of each pixel point in an RGB image on the surface of strip steel through a pooling template to obtain a corresponding characteristic map, stacking the characteristic maps corresponding to the receptive fields with different sizes into a characteristic map layer structure, obtaining a classification result of the pixel points in each characteristic map in the characteristic map layer structure, and obtaining a probability fluctuation curve of each pixel point in the RGB image according to the classification result to confirm that the defective pixel points obtain a final defect area. And detecting the defective pixel points according to the segmentation result change of each pixel point in the characteristic diagram and the local characteristics corresponding to the receptive fields with different sizes so as to obtain an accurate defect detection result, avoid missing detection and false detection, reduce the error of hot rolling slip defect detection and improve the surface quality of the strip steel.

Description

Strip steel surface hot rolling slip defect detection method based on deep learning
Technical Field
The invention relates to the technical field of artificial intelligence, in particular to a method for detecting hot rolling slippage defects on the surface of strip steel based on deep learning.
Background
In the production process of strip steel and stainless steel, the surface of the strip steel is defective due to improper operation of each process. In the production and processing processes of strip steel and stainless steel, the strip steel coiled by a winch is in strong contact with a winding drum and is wound up when the coiling is started, so that the surface defect of hot rolling slippage is generated on the surface of the strip steel, the defect is a flaw with short linear shape and unobvious characteristics, and a detection method of the hot rolling slippage surface defect is needed in order to guarantee the surface quality of the strip steel.
Disclosure of Invention
In order to solve the technical problems, the invention aims to provide a method for detecting the hot rolling slip defect on the surface of the strip steel based on deep learning, and the adopted technical scheme is as follows:
the embodiment of the invention provides a deep learning-based method for detecting hot rolling slip defects on the surface of strip steel, which comprises the following specific steps:
performing semantic segmentation on an RGB image on the surface of the strip steel to obtain a semantic segmentation image, acquiring a circumscribed rectangle of each defect area in the semantic segmentation image, and obtaining the maximum width and the maximum height by using the circumscribed rectangle;
respectively acquiring the receptive fields of each pixel point in the RGB image in the direction of a horizontal axis and the direction of a vertical axis by using two pooling templates with set sizes according to set step lengths to obtain a characteristic diagram corresponding to the receptive fields, wherein the characteristic diagram comprises a horizontal axis characteristic diagram and a vertical axis characteristic diagram; stacking the feature graphs corresponding to the receptive fields with different sizes into a feature layer structure, wherein the height of the feature layer structure is obtained from the maximum width and the maximum height; performing semantic segmentation on each feature map in the feature map layer structure to obtain a classification result of each pixel point, wherein the classification result is a probability value of a defective pixel point;
and obtaining a probability fluctuation curve of each pixel point in the RGB image based on the classification result of the pixel point in each feature map, and confirming the defect pixel point in the RGB image by the probability fluctuation curve to obtain a final defect area.
Preferably, the height of the feature layer structure is the maximum of the maximum width and the maximum height.
Preferably, the feature layer structure includes two branches, one branch is formed by stacking the horizontal-axis feature maps, and the other branch is formed by stacking the vertical-axis feature maps.
Preferably, the method for obtaining the probability fluctuation curve of each pixel point in the RGB image based on the classification result of the pixel point in each feature map includes:
acquiring a first height corresponding to the feature layer structure based on the receptive field size of the pixel points in the RGB image;
establishing a two-dimensional plane coordinate system by taking pixel points in the RGB image as an origin, wherein the abscissa of the two-dimensional plane coordinate system represents different receptive field sizes, and the ordinate represents the classification result corresponding to the pixel points in the characteristic diagram at the first height; based on the two-dimensional plane coordinate system, the probability fluctuation curve of each pixel point in the RGB image is obtained by respectively corresponding the classification results of the pixel points in the horizontal axis feature map and the vertical axis feature map under the receptive fields with different sizes.
Preferably, the method for confirming the defect pixel point in the RGB image by the probability fluctuation curve to obtain the final defect region includes:
calculating a probability fluctuation index of each point according to the classification result of each point on the probability fluctuation curve;
obtaining the marking value of each pixel point position in the RGB image according to the probability fluctuation index;
and confirming the defect pixel points in the RGB image based on the marking values, and obtaining the final defect area by the confirmed defect pixel points.
Preferably, the probability volatility indicator is derived from the difference in classification between adjacent points on the probability fluctuation curve.
Preferably, the method for obtaining the mark value of each pixel point position in the RGB image by the probability fluctuation indicator includes:
and when the probability fluctuation index is larger than zero, acquiring the position of the pixel point of the point in the RGB image, and adding one to the mark value of the position.
Preferably, the method for confirming the defective pixel point in the RGB image based on the mark value includes:
and setting a marking value threshold, and when the marking value of the pixel point position is greater than the marking value threshold, determining the pixel point as the defective pixel point.
Preferably, the initial value of the mark value for each pixel point position in the RGB image is zero.
The embodiment of the invention at least has the following beneficial effects: the method comprises the steps of obtaining different size receptive fields of each pixel point in an image by utilizing a pooling template with a set size to obtain corresponding characteristic graphs, semantically segmenting the characteristic graphs corresponding to the different size receptive fields, and detecting defective pixel points according to segmentation result changes of each pixel point and local characteristics corresponding to the different size receptive fields so as to obtain accurate defect detection results, avoid missing detection and false detection, reduce errors of hot rolling slippage defect detection, and improve the surface quality of strip steel.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions and advantages of the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without creative efforts.
FIG. 1 is a schematic illustration of a hot rolling slip defect provided in an embodiment of the present invention;
FIG. 2 is a flowchart illustrating steps of a method for detecting hot rolling slip defects on a surface of a strip steel based on deep learning according to an embodiment of the present invention;
fig. 3 is a schematic diagram of an acquisition process of a horizontal axis feature map and a vertical axis feature map in a feature layer structure according to an embodiment of the present invention.
FIG. 4 is a diagram of a pixel point in an RGB image according to an embodiment of the present invention
Figure DEST_PATH_IMAGE002
Schematic diagrams of corresponding different size receptive fields.
Detailed Description
In order to further illustrate the technical means and effects of the present invention adopted to achieve the predetermined invention purpose, the following detailed description, with reference to the accompanying drawings and preferred embodiments, describes specific implementation, structure, features and effects of a deep learning based method for detecting hot rolling slip defects on a strip steel surface according to the present invention. In the following description, different "one embodiment" or "another embodiment" refers to not necessarily the same embodiment. Furthermore, the particular features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.
Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
The specific scheme of the strip steel surface hot rolling slip defect detection method based on deep learning provided by the invention is specifically described below by combining the attached drawings.
The embodiment of the invention aims at the following specific scenes: the short linear and non-obvious flaw appeared in the hot rolling process of the strip production process is shown in figure 1.
Referring to fig. 2, a flowchart of steps of a deep learning based method for detecting hot rolling slip defects on a strip steel surface according to an embodiment of the present invention is shown, where the method includes the following steps:
and S001, performing semantic segmentation on the RGB image on the surface of the strip steel to obtain a semantic segmentation image, acquiring an external rectangle of each defect area in the semantic segmentation image, and obtaining the maximum width and the maximum height from the external rectangle.
Specifically, RGB images of the surface of the strip steel in the hot rolling process are collected and sent into a semantic segmentation network to obtain a semantic segmentation image, the semantic segmentation image is a binary image, a foreground area represents a defect area, and other areas are background areas. The training data of the semantic segmentation network is an acquired RGB image; the label data is obtained by artificial labeling: marking the pixel value of the pixel point of the defect area as 1, and marking the pixel value of the pixel point of other areas as 0; the loss function is a cross-entropy loss function.
Preferably, the embodiment of the present invention collects a semantic segmentation network of an encoder-decoder structure, and an implementer may use an existing network such as pnet, deplab v3, and the like.
Further, segmenting images for semantic meaningAnalyzing the connected domains to obtain the circumscribed rectangle of each connected domain in the foreground region, namely the circumscribed rectangle of the defect region, and further obtaining the width and height of each circumscribed rectangle
Figure DEST_PATH_IMAGE004
Wherein, in the step (A),
Figure DEST_PATH_IMAGE006
is the width of the circumscribed rectangle,
Figure DEST_PATH_IMAGE008
the maximum width is obtained from the width and height of each circumscribed rectangle
Figure DEST_PATH_IMAGE010
And maximum height
Figure DEST_PATH_IMAGE012
。
Step S002, respectively acquiring the receptive fields of each pixel point in the RGB image in the horizontal axis direction and the longitudinal axis direction by two pooling templates with set sizes according to set step lengths to obtain characteristic graphs corresponding to the receptive fields, wherein the characteristic graphs comprise a horizontal axis characteristic graph and a longitudinal axis characteristic graph; stacking feature graphs corresponding to different scale receptive fields into a feature layer structure, wherein the height of the feature layer structure is obtained by the maximum width and the maximum height; and performing semantic segmentation on each feature map in the feature map layer structure to obtain a classification result of each pixel point, wherein the classification result is the probability value of the defective pixel point.
Specifically, in order to obtain neighborhood characteristics of each pixel point in the RGB image in different sizes, the receptive fields of each pixel point in the RGB image in the horizontal axis direction and the longitudinal axis direction are respectively obtained by using two pooling templates with set sizes and set step lengths to obtain characteristic maps corresponding to the receptive fields, wherein the characteristic maps comprise a horizontal axis characteristic map and a longitudinal axis characteristic map; stacking feature graphs corresponding to different scale receptive fields into a feature layer structure, wherein the height of the feature layer structure is obtained from the maximum width and the maximum height, and the feature layer structure is obtained by the following steps:
(1) the two pooling templates used in the embodiment of the invention have the size of
Figure DEST_PATH_IMAGE014
And the step size is set to 1. Wherein
Figure DEST_PATH_IMAGE016
The pooling template is changed by the size of the characteristic diagram in the horizontal axis direction, namely the width of the horizontal axis characteristic diagram is reduced by 1 compared with that of the RGB image, and each pixel point in the horizontal axis characteristic diagram corresponds to the receptive field of the pixel point in the RGB image and also reflects the pixel point in the RGB image
Figure 58926DEST_PATH_IMAGE016
Neighborhood characteristics of (1); in the same way, the method for preparing the composite material,
Figure DEST_PATH_IMAGE018
the pooling template changes the feature map size in the longitudinal axis direction, namely the height of the longitudinal axis feature map is reduced by 1 compared with that of the RGB image, and each pixel point in the longitudinal axis feature map corresponds to the receptive field of the pixel point in the RGB image and also reflects the pixel point in the RGB image
Figure 507225DEST_PATH_IMAGE018
The neighborhood characteristics of (2).
The size of the RGB image is
Figure DEST_PATH_IMAGE020
Wherein
Figure DEST_PATH_IMAGE022
In order to be the width of the sheet,
Figure DEST_PATH_IMAGE024
is the height.
(2) With reference to FIG. 3, respectively
Figure 739492DEST_PATH_IMAGE014
The two pooling templates have a step length of 1 pair RProcessing the GB image to obtain two characteristic graphs of a second layer of the characteristic layer structure, namely a transverse-axis characteristic graph
Figure DEST_PATH_IMAGE026
And longitudinal axis feature map
Figure DEST_PATH_IMAGE028
And the size of the feature map of the horizontal axis is
Figure DEST_PATH_IMAGE030
The dimension of the feature map of the vertical axis is
Figure DEST_PATH_IMAGE032
。
In addition, the horizontal axis characteristic diagram
Figure 133038DEST_PATH_IMAGE026
And longitudinal axis feature map
Figure 275306DEST_PATH_IMAGE028
Each pixel point in the RGB image respectively corresponds to the pixel point in the RGB image with the size of
Figure 674932DEST_PATH_IMAGE014
The receptive field of (1).
(3) Further, utilize
Figure 145228DEST_PATH_IMAGE016
Cross-axis profile of pooled template to second layer
Figure 672024DEST_PATH_IMAGE026
Processing to obtain a cross-axis feature map of the third layer
Figure DEST_PATH_IMAGE034
Horizontal axis feature diagram
Figure 888373DEST_PATH_IMAGE034
Has a size of
Figure DEST_PATH_IMAGE036
(ii) a By using
Figure 789726DEST_PATH_IMAGE018
Longitudinal axis profile of pooled template to second layer
Figure 12897DEST_PATH_IMAGE028
Processing to obtain the longitudinal axis characteristic diagram of the third layer
Figure DEST_PATH_IMAGE038
Feature diagram of vertical axis
Figure DEST_PATH_IMAGE040
。
In addition, the horizontal axis characteristic diagram
Figure 484330DEST_PATH_IMAGE034
And longitudinal axis feature map
Figure 210977DEST_PATH_IMAGE038
Each pixel point in the RGB image respectively corresponds to the pixel point in the RGB image with the size of
Figure DEST_PATH_IMAGE042
The receptive field of (1).
(4) And analogizing in sequence, stacking the feature graphs corresponding to the receptive fields with different sizes into a feature graph layer structure, wherein the feature graph layer structure comprises two branches, and one branch is formed by stacking the feature graphs of the transverse axis
Figure 811461DEST_PATH_IMAGE022
One branch, the other branch being formed by stacking longitudinal axis feature maps
Figure 866136DEST_PATH_IMAGE024
And (4) branching. Because the width and height size information of the circumscribed rectangle of the defect region in the RGB image are different, the maximum width of the circumscribed rectangle in the RGB image is determined
Figure 610101DEST_PATH_IMAGE010
And maximum height
Figure 581468DEST_PATH_IMAGE012
Obtaining the heights corresponding to the two branches in the feature layer structure respectively
Figure 496728DEST_PATH_IMAGE022
The height of the feature layer structure corresponding to the branch is
Figure DEST_PATH_IMAGE044
、
Figure 756808DEST_PATH_IMAGE024
The height of the feature layer structure corresponding to the branch is
Figure DEST_PATH_IMAGE046
Wherein
Figure DEST_PATH_IMAGE048
Is an rounding-up function. Because the height information of the two branches is different, the maximum height information is the final height of the feature layer structure.
It should be noted that the height information of the two branches limits the number of feature maps in the feature map layer structure, and reduces the amount of calculation.
Further, all feature maps in the feature map layer structure are sent to the semantic segmentation network in the step S001 to obtain a semantic segmentation result of each feature map, the semantic segmentation result is a classification result of each pixel point, the classification result is a probability value that the pixel point belongs to a defective pixel point or a non-defective pixel point, and the classification result is set as a probability value
Figure DEST_PATH_IMAGE050
Wherein, in the step (A),
Figure DEST_PATH_IMAGE052
probability value of defective pixel point;
Figure DEST_PATH_IMAGE054
probability value of non-defect pixel point;
Figure DEST_PATH_IMAGE056
。
and S003, obtaining a probability fluctuation curve of each pixel point based on the classification result of the pixel points in each characteristic graph, and confirming the defect pixel points in the RGB image by the probability fluctuation curve to obtain a final defect area.
Specifically, local features of different-size receptive fields corresponding to each pixel point in the RGB image can be reflected by using a pixel point of a feature map in the feature map layer structure, and therefore, the receptive field size of each pixel point in the RGB image is obtained based on each feature map in the feature map layer structure.
It should be noted that, because the value of the receptive field size and the height information of the two branches of the feature layer structure are determined, the receptive field size range corresponding to each pixel point in the RGB image is obtained based on the height of the feature layer structure.
The embodiment of the invention uses the pixel points in the RGB image
Figure 632360DEST_PATH_IMAGE002
For example, a pixel
Figure 458233DEST_PATH_IMAGE002
The size of the receptive field comprises a horizontal axis direction and a vertical axis direction, so that the pixel point
Figure 809974DEST_PATH_IMAGE002
The receptive field size of (a) includes a receptive field size in the horizontal axis direction and a receptive field size in the vertical axis direction. For the
Figure 822929DEST_PATH_IMAGE022
Pixel point on branch
Figure 174276DEST_PATH_IMAGE002
The receptive field size range of (a):
Figure DEST_PATH_IMAGE058
wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE060
is a pixel point
Figure 572765DEST_PATH_IMAGE002
Position in the direction of the transverse axis.
In the same way, for
Figure 718576DEST_PATH_IMAGE024
Pixel point on branch
Figure 235139DEST_PATH_IMAGE002
The receptive field size range of (a):
Figure DEST_PATH_IMAGE062
wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE064
is a pixel point
Figure 817606DEST_PATH_IMAGE002
Position in the direction of the longitudinal axis.
It should be noted that the pixel points
Figure 634383DEST_PATH_IMAGE002
The position information of
Figure DEST_PATH_IMAGE066
Then it is of a size of
Figure DEST_PATH_IMAGE068
And
Figure DEST_PATH_IMAGE070
respectively correspond to local characteristics of the receptive field
Figure 184051DEST_PATH_IMAGE022
In a branch is first
Figure DEST_PATH_IMAGE072
Pixel points in layer cross-axis feature map
Figure 391172DEST_PATH_IMAGE066
And
Figure 943377DEST_PATH_IMAGE024
in a branch is first
Figure DEST_PATH_IMAGE074
Pixel points in layer longitudinal axis feature map
Figure 912863DEST_PATH_IMAGE066
Then pixel point
Figure 400476DEST_PATH_IMAGE002
The corresponding different size receptive fields are shown in fig. 4.
Further, analyzing classification results of each pixel point in the RGB image in the feature maps corresponding to the receptive fields with different sizes, namely acquiring a first height corresponding to the feature layer structure based on the receptive field sizes of the pixel points in the RGB image; establishing a two-dimensional plane coordinate system by taking pixel points in the RGB image as an origin, wherein the abscissa of the two-dimensional plane coordinate system represents different receptive field sizes, and the ordinate represents a classification result of the pixel points in the corresponding characteristic diagram at the first height; based on a two-dimensional plane coordinate system, obtaining probability fluctuation curves of each pixel point in the RGB image respectively corresponding to classification results of the pixel points in the horizontal axis feature map and the vertical axis feature map under different size receptive fields, and confirming the defect pixel points in the RGB image by the probability fluctuation curves to obtain a final defect area, wherein the method for obtaining the final defect area comprises the following steps:
(1) the embodiment of the invention uses the pixel points in the RGB image
Figure 891631DEST_PATH_IMAGE002
For example, first, pixel points are mapped
Figure 247526DEST_PATH_IMAGE002
Analysis of the receptive field along the transverse axis: based on pixel points
Figure 897951DEST_PATH_IMAGE002
The corresponding receptive field size range is defined by pixel points
Figure 664787DEST_PATH_IMAGE002
Establishing a two-dimensional plane coordinate system for the origin, the abscissa in the two-dimensional plane coordinate system
Figure DEST_PATH_IMAGE076
Representing the dimension in the direction of the transverse axis
Figure DEST_PATH_IMAGE078
Reception field, ordinate of
Figure DEST_PATH_IMAGE080
The horizontal axis corresponds to the pixel points in the horizontal axis characteristic diagram under the size receptive field
Figure 502293DEST_PATH_IMAGE002
The classification result of (1).
Preferably, the embodiment of the invention selects the probability value of the defective pixel point in the classification result
Figure 648497DEST_PATH_IMAGE052
As the value of the ordinate.
(2) According to pixel point
Figure 278062DEST_PATH_IMAGE002
The receptive fields with different sizes in the direction of the transverse axis and the probability values of the corresponding pixel points
Figure 841898DEST_PATH_IMAGE052
Obtaining pixel points
Figure 307646DEST_PATH_IMAGE002
Probability fluctuation curve of
Figure DEST_PATH_IMAGE082
. Probability fluctuation curve
Figure 582507DEST_PATH_IMAGE082
Reflecting the influence of local characteristics of different size receptive fields on defect classification probability, calculating the probability fluctuation index of each point according to the classification result of each point on the probability fluctuation curve, wherein the probability fluctuation index is obtained by the difference of the classification result between adjacent points on the probability fluctuation curve, and the probability fluctuation curve
Figure 941944DEST_PATH_IMAGE082
Midpoint
Figure 801316DEST_PATH_IMAGE076
The calculation method of the probability fluctuation index comprises the following steps:
Figure DEST_PATH_IMAGE084
wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE086
as the midpoint of the probability fluctuation curve
Figure 865611DEST_PATH_IMAGE076
A probability fluctuation index of;
Figure DEST_PATH_IMAGE088
is a point
Figure 835841DEST_PATH_IMAGE076
A probability value of (d);
Figure DEST_PATH_IMAGE090
is a point
Figure DEST_PATH_IMAGE092
The probability value of (2).
The probability fluctuation index is the probability value change condition of the pixel point when the receptive field changes, and the mark value of each pixel point position in the RGB image is obtained by the probability fluctuation index, namely when the probability fluctuation index
Figure DEST_PATH_IMAGE094
At the same time, the pixel point is represented
Figure 767894DEST_PATH_IMAGE002
Covering the receptive field from the normal area to the defect area, and acquiring the abscissa at the moment
Figure 63746DEST_PATH_IMAGE076
(ii) a Further obtain the abscissa
Figure 238507DEST_PATH_IMAGE076
Corresponding to the pixel position in the RGB image
Figure DEST_PATH_IMAGE096
And the marking value at the position of the pixel point is added with 1.
It should be noted that each pixel point in the RGB image has a corresponding mark value, and the initial value of the mark value is 0.
(3) Similarly, the pixel point is processed in the same way
Figure 604240DEST_PATH_IMAGE002
Analysis of the receptive field along the longitudinal axis: based on pixel points
Figure 672690DEST_PATH_IMAGE002
The corresponding receptive field size range is defined by pixel points
Figure 139443DEST_PATH_IMAGE002
Establishing a two-dimensional plane coordinate system for the origin, the abscissa in the two-dimensional plane coordinate system
Figure 801500DEST_PATH_IMAGE076
Representing the dimension in the direction of the longitudinal axis
Figure DEST_PATH_IMAGE098
Reception field, ordinate of
Figure 425117DEST_PATH_IMAGE080
A pixel point under a receptive field is a dimension corresponding to the abscissa in the longitudinal axis characteristic diagram
Figure 738287DEST_PATH_IMAGE002
The classification result of (i.e. probability value of defective pixel)
Figure 251308DEST_PATH_IMAGE052
. And then according to the pixel point
Figure 666240DEST_PATH_IMAGE002
The receptive fields with different sizes in the direction of the longitudinal axis and the probability values of the corresponding pixel points
Figure 47542DEST_PATH_IMAGE052
Obtaining pixel points
Figure 936257DEST_PATH_IMAGE002
Probability fluctuation curve of
Figure DEST_PATH_IMAGE100
And (3) updating the mark value of each pixel point position in the RGB image by using the method in the step (2).
(4) By counting pixel points
Figure 948076DEST_PATH_IMAGE002
The local characteristics of the receptive fields with different sizes are analyzed to obtain pixel points
Figure 319145DEST_PATH_IMAGE002
Updating the mark value of each pixel point position in the RGB image. And (4) completing the updating of the marking value of each pixel point to each pixel point position in the RGB image by using the methods from the step (1) to the step (3) to obtain a marking value image.
（5）In the marking value image, the marking value of the pixel point position can be regarded as a voting result of whether the pixel point belongs to a defective pixel point according to the local characteristics of the pixel point in the receptive fields with different sizes, and the larger the marking value is, the higher the possibility that the pixel point belongs to the defective pixel point is. Setting a flag value threshold
Figure DEST_PATH_IMAGE102
When any pixel point position
Figure DEST_PATH_IMAGE104
Is marked with a value
Figure DEST_PATH_IMAGE106
Marker value threshold
Figure DEST_PATH_IMAGE108
And if so, determining that the pixel point belongs to a defective pixel point, and further obtaining a final defective area according to the confirmed defective pixel point.
Note that the flag value threshold value
Figure 425510DEST_PATH_IMAGE102
To achieve the empirical threshold, the practitioner may make modifications based on the accuracy requirements of the defect detection.
In summary, the embodiment of the present invention provides a method for detecting a hot rolling slip defect on a strip steel surface based on deep learning, the method performs semantic segmentation on an RGB image of the strip steel surface to obtain a semantic segmentation image, and obtains a circumscribed rectangle of each defect region in the semantic segmentation image to obtain a maximum width and a maximum height; performing pooling operation on the RGB image in the transverse axis direction and the longitudinal axis direction by using a pooling template to obtain transverse axis feature maps and longitudinal axis feature maps corresponding to different size receptive fields, stacking the transverse axis feature maps and the longitudinal axis feature maps into a feature map layer structure, wherein the feature map layer structure is obtained by the maximum width and the maximum height, and performing semantic segmentation on each feature map in the feature map layer structure to obtain a classification result of each pixel point; and obtaining a probability fluctuation curve of each pixel point in the RGB image based on the classification result of the pixel points in each characteristic graph, and confirming the defective pixel points in the RGB image by the probability fluctuation curve to obtain a final defective area. By semantically segmenting the feature maps corresponding to the receptive fields with different sizes and detecting the defective pixel points according to the segmentation result change of each pixel point and the local features corresponding to the receptive fields with different sizes, the accurate defect detection result is obtained, the error of hot rolling slippage defect detection is reduced, and the surface quality of the strip steel is improved.
It should be noted that: the precedence order of the above embodiments of the present invention is only for description, and does not represent the merits of the embodiments. And specific embodiments thereof have been described above. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims may be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing may also be possible or may be advantageous.
The embodiments in the present specification are described in a progressive manner, and the same and similar parts among the embodiments are referred to each other, and each embodiment focuses on the differences from the other embodiments.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.

Claims (9)

1. A deep learning-based method for detecting hot rolling slip defects on the surface of strip steel is characterized by comprising the following steps:
performing semantic segmentation on an RGB image on the surface of the strip steel to obtain a semantic segmentation image, acquiring a circumscribed rectangle of each defect area in the semantic segmentation image, and obtaining the maximum width and the maximum height by using the circumscribed rectangle;
respectively acquiring the receptive fields of each pixel point in the RGB image in the direction of a horizontal axis and the direction of a vertical axis by using two pooling templates with set sizes according to set step lengths to obtain a characteristic diagram corresponding to the receptive fields, wherein the characteristic diagram comprises a horizontal axis characteristic diagram and a vertical axis characteristic diagram; stacking the feature graphs corresponding to the receptive fields with different sizes into a feature layer structure, wherein the height of the feature layer structure is obtained from the maximum width and the maximum height; performing semantic segmentation on each feature map in the feature map layer structure to obtain a classification result of each pixel point, wherein the classification result is a probability value of a defective pixel point;
and obtaining a probability fluctuation curve of each pixel point in the RGB image based on the classification result of the pixel point in each feature map, and confirming the defect pixel point in the RGB image by the probability fluctuation curve to obtain a final defect area.
2. The method according to claim 1, wherein the height of the feature layer structure is the maximum of the maximum width and the maximum height.
3. The method of claim 1, wherein the feature layer structure includes two branches, one branch being stacked from the cross-axis feature map and the other branch being stacked from the vertical-axis feature map.
4. The method as claimed in claim 1, wherein the method for obtaining the probability fluctuation curve of each pixel point in the RGB image based on the classification result of the pixel point in each feature map comprises:
acquiring a first height corresponding to the feature layer structure based on the receptive field size of the pixel points in the RGB image;
establishing a two-dimensional plane coordinate system by taking pixel points in the RGB image as an origin, wherein the abscissa of the two-dimensional plane coordinate system represents different receptive field sizes, and the ordinate represents the classification result corresponding to the pixel points in the characteristic diagram at the first height; based on the two-dimensional plane coordinate system, the probability fluctuation curve of each pixel point in the RGB image is obtained by respectively corresponding the classification results of the pixel points in the horizontal axis feature map and the vertical axis feature map under the receptive fields with different sizes.
5. The method of claim 1, wherein said identifying said defective pixels in said RGB image from said probability fluctuation curve to obtain a final defective region comprises:
calculating a probability fluctuation index of each point according to the classification result of each point on the probability fluctuation curve;
obtaining the marking value of each pixel point position in the RGB image according to the probability fluctuation index;
and confirming the defect pixel points in the RGB image based on the marking values, and obtaining the final defect area by the confirmed defect pixel points.
6. The method of claim 5, wherein the probability volatility indicator is derived from differences in the classification results between adjacent points on the probability fluctuation curve.
7. The method as claimed in claim 5, wherein the method for obtaining the mark value of each pixel point position in the RGB image by the probability fluctuation indicator comprises:
and when the probability fluctuation index is larger than zero, acquiring the position of the pixel point of the point in the RGB image, and adding one to the mark value of the position.
8. The method of claim 5, wherein said method of validating said defective pixel in said RGB image based on said marker value comprises:
and setting a marking value threshold, and when the marking value of the pixel point position is greater than the marking value threshold, determining the pixel point as the defective pixel point.
9. The method of claim 5, wherein the initial value of the marker value for each pixel point location in the RGB image is zero.
CN202111058596.7A 2021-09-10 2021-09-10 Strip steel surface hot rolling slip defect detection method based on deep learning Active CN113506295B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111058596.7A CN113506295B (en) 2021-09-10 2021-09-10 Strip steel surface hot rolling slip defect detection method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111058596.7A CN113506295B (en) 2021-09-10 2021-09-10 Strip steel surface hot rolling slip defect detection method based on deep learning

Publications (2)

Publication Number Publication Date
CN113506295A CN113506295A (en) 2021-10-15
CN113506295B true CN113506295B (en) 2021-11-26

Family

ID=78017083

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111058596.7A Active CN113506295B (en) 2021-09-10 2021-09-10 Strip steel surface hot rolling slip defect detection method based on deep learning

Country Status (1)

Country Link
CN (1) CN113506295B (en)

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107833220A (en) * 2017-11-28 2018-03-23 河海大学常州校区 Fabric defect detection method based on depth convolutional neural networks and vision significance
CN111401436A (en) * 2020-03-13 2020-07-10 北京工商大学 Streetscape image segmentation method fusing network and two-channel attention mechanism

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107833220A (en) * 2017-11-28 2018-03-23 河海大学常州校区 Fabric defect detection method based on depth convolutional neural networks and vision significance
CN111401436A (en) * 2020-03-13 2020-07-10 北京工商大学 Streetscape image segmentation method fusing network and two-channel attention mechanism

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
基于深度学习语义分割的导光板缺陷检测方法;柳锋等;《计算机系统应用》;20201231;第26卷(第06期);全文 *
基于语义分割的火车车厢位置检测研究;卢进南等;《工程设计学报》;20201031;第27卷(第05期);全文 *

Also Published As

Publication number Publication date
CN113506295A (en) 2021-10-15

Similar Documents

Publication Publication Date Title
US11551341B2 (en) Method and device for automatically drawing structural cracks and precisely measuring widths thereof
CN113239930B (en) Glass paper defect identification method, system, device and storage medium
CN109635814B (en) Forest fire automatic detection method and device based on deep neural network
CN114897896B (en) Building wood defect detection method based on gray level transformation
CN106780455A (en) A kind of product surface detection method based on the local neighborhood window for sliding
CN110276285A (en) A kind of shipping depth gauge intelligent identification Method in uncontrolled scene video
CN111476120B (en) Unmanned aerial vehicle intelligent ship water gauge identification method and device
CN110334709B (en) License plate detection method based on end-to-end multi-task deep learning
CN110223310B (en) Line structure light center line and box edge detection method based on deep learning
CN116740070A (en) Plastic pipeline appearance defect detection method based on machine vision
CN114782391A (en) Method, system and device for constructing defect detection model of few-sample industrial image
CN115147418B (en) Compression training method and device for defect detection model
CN112215907A (en) Automatic extraction method for weld defects
CN110991374B (en) Fingerprint singular point detection method based on RCNN
CN115775236A (en) Surface tiny defect visual detection method and system based on multi-scale feature fusion
CN113506295B (en) Strip steel surface hot rolling slip defect detection method based on deep learning
CN116189194A (en) Drawing enhancement segmentation method for engineering modeling
CN113469984B (en) Method for detecting appearance of display panel based on YOLO structure
CN114418968A (en) Power transmission line small target defect detection method based on deep learning
CN111429411B (en) X-ray defect image sample generation method for carbon fiber composite core wire
CN113256563A (en) Method and system for detecting surface defects of fine product tank based on space attention mechanism
CN114322793B (en) Workpiece size measuring method and device based on global segmentation network and storage medium
CN116596878B (en) Strip steel surface defect detection method, system, electronic equipment and medium
CN116403204B (en) Intelligent reading method for round pointer meter
CN117541582B (en) IGBT insulation quality detection method for high-frequency converter

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
GR01 Patent grant
GR01 Patent grant