CN111524117A - Tunnel surface defect detection method based on characteristic pyramid network - Google Patents

Tunnel surface defect detection method based on characteristic pyramid network Download PDF

Info

Publication number
CN111524117A
CN111524117A CN202010312482.XA CN202010312482A CN111524117A CN 111524117 A CN111524117 A CN 111524117A CN 202010312482 A CN202010312482 A CN 202010312482A CN 111524117 A CN111524117 A CN 111524117A
Authority
CN
China
Prior art keywords
network
tunnel
feature
image
detection
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
Application number
CN202010312482.XA
Other languages
Chinese (zh)
Inventor
汪俊
隆昆
李大伟
李虎
刘树亚
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.)
Nanjing University of Aeronautics and Astronautics
Original Assignee
Nanjing University of Aeronautics and Astronautics
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 Nanjing University of Aeronautics and Astronautics filed Critical Nanjing University of Aeronautics and Astronautics
Priority to CN202010312482.XA priority Critical patent/CN111524117A/en
Publication of CN111524117A publication Critical patent/CN111524117A/en
Pending legal-status Critical Current

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
    • G06T7/0006—Industrial image inspection using a design-rule based approach
    • 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
    • G06N3/045—Combinations of 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
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20016—Hierarchical, coarse-to-fine, multiscale or multiresolution image processing; Pyramid transform
    • 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
    • G06T2207/20024—Filtering details

Abstract

The invention discloses a tunnel surface defect detection method based on a characteristic pyramid network, which comprises the steps of collecting tunnel images with surface defects, constructing a data set, and dividing the images in the data set into a training set and a testing set; preprocessing a tunnel surface image; marking defects on the surface of the tunnel, and marking cracks and hollow areas in the image by adopting a marking tool; constructing a significance detection network for defect detection; and constructing a training loss function, and detecting the tunnel surface defects by using a detection network of the member. The method has the advantages of strong robustness and significance detection effectiveness, and can well locate the correct tunnel surface defects.

Description

Tunnel surface defect detection method based on characteristic pyramid network
Technical Field
The invention relates to a tunnel surface defect detection method, in particular to a tunnel surface defect detection method based on a characteristic pyramid network, and belongs to the field of computer vision and image processing.
Background
In recent years, with the gradual saturation of ground road traffic of various large cities, the situation of traffic jam at a peak is more and more serious, and subways gradually become main public transport means of the cities due to the advantages of convenience, quickness and the like. The development prospect of the subway is good, and each large city is also tightening and widening a subway network and increasing subway lines. The quality of the tunnel is unqualified due to various reasons in the subway construction process, and the quality problems of cracks and the like in the tunnel due to the surrounding construction of the tunnel or natural disasters in the subway tunnel use process can further cause major accidents such as tunnel collapse. Therefore, the defect detection is carried out in the subway construction process and the subway using process, so that the occurrence of major traffic safety accidents can be effectively avoided and prevented. The traditional defect detection method mainly depends on human eyes for detection, consumes a large amount of time, manpower and financial resources, and is not suitable for the requirement of the modern society on the fault detection efficiency, so that the development of a high-precision and high-efficiency tunnel surface defect detection method plays an important role in the development of modern tunnel traffic.
At present, the tunnel surface defect detection method except for manual identification is mainly a method based on the traditional digital image processing technology, and comprises the following steps: (1) a detection method based on threshold processing; (2) detection methods based on edge detection and morphological operations. The detection method based on threshold processing is mainly realized by searching an interested area through gray level, but the detection effect in a complex background is poor, and false detection is easy to occur when the gray level value of a target to be detected is close to that of noise. The processing method based on edge detection and morphological operation mainly comprises the steps of obtaining a defect edge through an edge detection operator, and then determining a defect area through morphological operation.
Disclosure of Invention
The invention aims to provide a tunnel surface defect detection method based on a characteristic pyramid network, and improve the efficiency and the precision of tunnel surface defect detection.
In order to solve the technical problems, the technical scheme adopted by the invention is as follows:
a tunnel surface defect detection method based on a characteristic pyramid network is characterized by comprising the following steps:
the method comprises the following steps: acquiring tunnel images with surface defects, constructing a data set, and dividing the images in the data set into a training set and a testing set;
step two: preprocessing a tunnel surface image;
step three: marking defects on the surface of the tunnel, and marking cracks and hollow areas in the image by adopting a marking tool;
step four: constructing a significance detection network for defect detection;
step five: and constructing a training loss function, and detecting the tunnel surface defects by using the constructed detection network.
Further, the first step is specifically to collect a tunnel image with surface defects, cut key information of the image, construct an image library, perform scaling, turning and mirror image expansion on the image in the image library, and divide the image in the expanded image library into a training sample set and a testing sample set according to a ratio of 4: 1.
Further, the second step is specifically to perform preprocessing on the tunnel surface image and remove image noise by using guided filtering, and the specific process is as follows:
taking the image as a two-dimensional function, assuming that the output and input of the function satisfy a linear relationship within a two-dimensional window:
Figure BDA0002458375170000031
where q is the value of the output pixel, I is the input image, I and k are the pixel indices, ωkIs a window, akAnd bkIs the coefficient of the linear function when the window is located at k; taking the gradient on two sides of the above formula:
Figure BDA0002458375170000032
the difference between the fitting function output value and the true value is:
Figure BDA0002458375170000033
wherein p is a true value, and e represents a minimum value;
the coefficients of the image function are obtained by the least squares method:
Figure BDA0002458375170000034
Figure BDA0002458375170000035
where | ω | is the window ω |kNumber of middle pixels, μkIs that I is in the window omegakThe average value of (a) is,
Figure BDA0002458375170000036
is that I is in the window omegakThe variance in (a) is greater than or equal to,
Figure BDA0002458375170000037
is that the image p to be filtered is in the window omegakThe average value of (1);
when calculating the linear coefficients of each window, since each pixel is described by a plurality of linear functions, when specifically calculating the output value of a certain point, all the linear functions including the point are averaged:
Figure BDA0002458375170000038
in the above formula IiIs the intensity of the pixel i, ωkIs the total window containing pixel i, k is its central position,
Figure BDA0002458375170000041
represents the slope parameter of the linear regression,
Figure BDA0002458375170000042
representing the linear regression intercept parameter.
Further, the fourth step is specifically to construct a saliency detection network, and construct a feature pyramid network based on the VGG16 network. Two low-level blocks Conv1-2 and Conv2-2 of the VG 16 network are used for extracting low-level features of the image, and because background regions which interfere with the generation of the saliency map exist in the low-level features, a space attention mechanism is adopted to filter out some background details according to the high-level features so as to focus more on the foreground regions;
taking the characteristics generated by three high-level blocks Conv3-3, Conv4-5 and Conv5-3 of the VGG16 network as initial high-level characteristics, and in order to enable the finally extracted high-level characteristics to contain scale and shape invariance, adopting the convolution of holes with expansion rates of 3, 5 and 7 respectively to obtain the context information of multiple receptive fields; then, a channel attention mechanism is adopted to allocate a larger weight value for a channel with high response to the obvious target; cross-channel cascade connection is carried out on the feature mapping of the void convolution layers with different expansion rates and the 1 x 1 dimensional dimension reduction features to obtain three features with different scales and context perception information, and two smaller scale features are promoted to the scale with the same maximum feature scale through up-sampling; and finally, combining the high-level features and the low-level features through cross-channel cascade connection to serve as the output of the context-aware pyramid feature network.
Further, a channel attention mechanism is added after the context perception pyramid features are extracted, and the weighted multi-scale and multi-perception wild features are extracted:
CA＝F(vh,W)＝σ1(fc2((fc1(vh,W1)),W2)) (7)
where CA stands for channel attention, vhRepresenting the channel feature vector, W refers to the parameter in the attention block of each channel, σ1Sigmoid operation is indicated, fc is a full connection layer, and an activation function ReLU is indicated; the final output of the channel attention module is defined as the high level feature f of the feature pyramidhThe weighting value of (1):
Figure BDA0002458375170000051
the saliency map of the low-level features contains a large amount of detail information, and in order to obtain a detailed boundary between a salient object and a background, a spatial attention method is adopted to concentrate attention on a foreground area; to increase the receptive field to obtain global information without increasing parameters, convolutional layers with two kernels of 1 × k and k × 1, respectively, are used for high-level feature capture spatial interest:
Figure BDA0002458375170000052
Figure BDA0002458375170000053
wherein C is1、C2Representing intermediate results of two different orders of convolution operations, conv1、conv2Which represents a convolution operation, is a function of,
Figure BDA0002458375170000054
the parameters in the spatial attention module are represented,
Figure BDA0002458375170000055
an output of the channel attention module;
then, carrying out normalization processing on the coded spatial feature map mapped to [0,1] by using sigmoid operation:
Figure BDA0002458375170000056
where SA denotes spatial attention, W denotes a parameter in the spatial attention module, σ2Representing sigmoid operation;
the final output of the spatial attention module is the low-level feature flThe weighting value of (1):
Figure BDA0002458375170000057
further, the fifth step is to define a loss function in the significance target detection by using the cross entropy between the final significance detection result graph and the theoretical result:
Figure BDA0002458375170000061
where Y represents the theoretical result, P represents the significance of the network output, αsA balance parameter representing a positive sample and a negative sample;
the above loss function only provides a general guidance for generating a saliency map, and in order to highlight the generation of the boundary details of a salient object, a laplacian operator is used to obtain a theoretical boundary of the salient object and a boundary of the salient object in the saliency map output by a network:
Figure BDA0002458375170000062
Figure BDA0002458375170000063
wherein f represents a feature, x, y are cartesian coordinates of the xy plane, the laplace algorithm is executed using the gradient of the image, directly invoking the convolution operation conv () internally, tanh () is an activation function, KlaplaceIs laplacian, abs () represents absolute value operations;
the generation of the significant object boundary is then supervised by cross entropy loss:
Figure BDA0002458375170000064
finally, defining the total loss function as the weighted sum of the significant object detection loss and the significant object boundary detection loss:
L＝αLS+(1-α)LB(17)
wherein L is a loss function, LSFor significant target detection loss, LBFor significant object boundary detection loss α is a weight.
Further, α in the formula (13)s＝0.528。
Further, in the model training process, firstly, α is set to 1.0, the initial learning rate is 0.01, a rough saliency map is generated through training, then α is adjusted to refine the boundaries of the saliency map, and the adjusted value of α is 0.7.
Compared with the prior art, the invention has the following advantages and effects: the invention provides a tunnel surface defect detection method based on a significance detection network aiming at the problems that the defect detection of the surface of a subway tunnel is difficult to finish with high efficiency and high precision, and the like, wherein a characteristic pyramid attention network is designed and established by considering different characteristics of different levels of characteristics, a context perception pyramid characteristic extraction module is designed aiming at high-level characteristics, and multi-scale void convolution and a channel attention mechanism are used for obtaining the high-level semantic characteristics; using a spatial attention mechanism for low-level features to suppress background noise, attention is focused on salient objects. Experiments show that the network has stronger robustness and significance detection effectiveness, and the method can well locate correct tunnel surface defects.
Drawings
FIG. 1 is a flowchart of a method for detecting defects on a tunnel surface based on a feature pyramid network according to the present invention.
FIG. 2 is a schematic diagram of tunnel surface defect test data according to an embodiment of the present invention.
Fig. 3 is a schematic diagram of a tunnel surface defect detection result according to an embodiment of the present invention.
FIG. 4 is a schematic diagram of a feature pyramid attention network structure according to an embodiment of the present invention.
Detailed Description
To further clarify the working principle and working process of the present invention, the following description is given in conjunction with the accompanying drawings and specific embodiments for facilitating the understanding and implementation of the present invention for those skilled in the art, and the following detailed description is given for the purpose of illustration and explanation, and it is to be understood that the embodiments described herein are only used for the purpose of illustration and explanation, and are not to be construed as limiting the present invention.
As shown in fig. 1, a method for detecting defects on a tunnel surface based on a feature pyramid network of the present invention is characterized by comprising the following steps:
the method comprises the following steps: acquiring tunnel images with surface defects, constructing a data set, and dividing the images in the data set into a training set and a testing set;
the method comprises the steps of collecting tunnel images with surface defects, cutting key information of the images, constructing an image library, carrying out scaling, turning and mirror image expansion on the images in the image library, and dividing the images in the expanded image library into a training sample set and a testing sample set according to the ratio of 4: 1.
Step two: preprocessing a tunnel surface image;
the method comprises the following steps of preprocessing a tunnel surface image, and removing image noise by using guide filtering, wherein the specific process comprises the following steps:
taking the image as a two-dimensional function, assuming that the output and input of the function satisfy a linear relationship within a two-dimensional window:
Figure BDA0002458375170000081
where q is the value of the output pixel, I is the input image, I and k are the pixel indices, ωkIs a window, akAnd bkIs the coefficient of the linear function when the window is located at k; taking the gradient on two sides of the above formula:
Figure BDA0002458375170000082
the difference between the fitting function output value and the true value is:
Figure BDA0002458375170000083
wherein p is a true value, and e represents a minimum value;
the coefficients of the image function are obtained by the least squares method:
Figure BDA0002458375170000091
Figure BDA0002458375170000092
where | ω | is the window ω |kNumber of middle pixels, μkIs that I is in the window omegakThe average value of (a) is,
Figure BDA0002458375170000093
is that I is in the window omegakThe variance in (a) is greater than or equal to,
Figure BDA0002458375170000094
is that the image p to be filtered is in the window omegakThe average value of (1);
when calculating the linear coefficients of each window, since each pixel is described by a plurality of linear functions, when specifically calculating the output value of a certain point, all the linear functions including the point are averaged:
Figure BDA0002458375170000095
in the above formula IiIs the intensity of the pixel i, ωkIs the total window containing pixel i, k is its central position,
Figure BDA0002458375170000096
represents the slope parameter of the linear regression,
Figure BDA0002458375170000097
representing the linear regression intercept parameter.
Step three: marking defects on the surface of the tunnel, and marking cracks and hollow areas in the image by adopting a marking tool;
step four: constructing a significance detection network for defect detection;
and constructing a significance detection network, and constructing a feature pyramid network based on the VGG16 network. Two low-level blocks Conv1-2 and Conv2-2 of the VG 16 network are used for extracting low-level features of the image, and because background regions which interfere with the generation of the saliency map exist in the low-level features, a space attention mechanism is adopted to filter out some background details according to the high-level features so as to focus more on the foreground regions;
taking the characteristics generated by three high-level blocks Conv3-3, Conv4-5 and Conv5-3 of the VGG16 network as initial high-level characteristics, and in order to enable the finally extracted high-level characteristics to contain scale and shape invariance, adopting the convolution of holes with expansion rates of 3, 5 and 7 respectively to obtain the context information of multiple receptive fields; then, a channel attention mechanism is adopted to allocate a larger weight value for a channel with high response to the obvious target; cross-channel cascade connection is carried out on the feature mapping of the void convolution layers with different expansion rates and the 1 x 1 dimensional dimension reduction features to obtain three features with different scales and context perception information, and two smaller scale features are promoted to the scale with the same maximum feature scale through up-sampling; and finally, combining the high-level features and the low-level features through cross-channel cascade connection to serve as the output of the context-aware pyramid feature network.
Adding a channel attention mechanism after the context perception pyramid feature extraction, and extracting the weighted multi-scale and multi-perception wild features:
CA＝F(vh,W)＝σ1(fc2((fc1(vh,W1)),W2)) (7)
where CA stands for channel attention, vhRepresenting the channel feature vector, W refers to the parameter in the attention block of each channel, σ1Sigmoid operation is indicated, fc is a full connection layer, and an activation function ReLU is indicated; the final output of the channel attention module is defined as the high level feature f of the feature pyramidhThe weighting value of (1):
Figure BDA0002458375170000101
the saliency map of the low-level features contains a large amount of detail information, and in order to obtain a detailed boundary between a salient object and a background, a spatial attention method is adopted to concentrate attention on a foreground area; to increase the receptive field to obtain global information without increasing parameters, convolutional layers with two kernels of 1 × k and k × 1, respectively, are used for high-level feature capture spatial interest:
Figure BDA0002458375170000102
Figure BDA0002458375170000103
wherein C is1、C2Representing intermediate results of two different orders of convolution operations, conv1、conv2Which represents a convolution operation, is a function of,
Figure BDA0002458375170000104
the parameters in the spatial attention module are represented,
Figure BDA0002458375170000111
an output of the channel attention module;
then, carrying out normalization processing on the coded spatial feature map mapped to [0,1] by using sigmoid operation:
Figure BDA0002458375170000112
where SA denotes spatial attention, W denotes a parameter in the spatial attention module, σ2Representing sigmoid operation;
the final output of the spatial attention module is the low-level feature flThe weighting value of (1):
Figure BDA0002458375170000113
step five: and constructing a training loss function, and detecting the tunnel surface defects by using a detection network of the member.
The cross entropy between the final saliency detection result map and the theoretical result is typically used in saliency target detection to define a loss function:
Figure BDA0002458375170000114
where Y represents the theoretical result, P represents the significance of the network output, αsRepresenting the balance parameters of the positive and negative samples, α in this examples＝0.528。
The above loss function only provides a general guidance for generating a saliency map, and in order to highlight the generation of the boundary details of a salient object, a laplacian operator is used to obtain a theoretical boundary of the salient object and a boundary of the salient object in the saliency map output by a network:
Figure BDA0002458375170000115
Figure BDA0002458375170000116
wherein f represents a feature, x, y are cartesian coordinates of the xy plane, the laplace algorithm is executed using the gradient of the image, directly invoking the convolution operation conv () internally, tanh () is an activation function, KlaplaceIs laplacian, abs () represents absolute value operations;
the generation of the significant object boundary is then supervised by cross entropy loss:
Figure BDA0002458375170000121
finally, defining the total loss function as the weighted sum of the significant object detection loss and the significant object boundary detection loss:
L＝αLS+(1-α)LB(17)
wherein L is a loss function, LSFor significant target detection loss, LBFor significant object boundary detection loss α is a weight.
In the model training process, firstly, alpha is set to be 1.0, the initial learning rate is 0.01, a rough saliency map is generated through training, then alpha is adjusted to refine the boundaries of the saliency map, and the adjusted optimal value of alpha is 0.7.
And finally, inputting the data to be tested into the network model with the good component to automatically detect the defects of the tunnel surface.
The invention provides a tunnel surface defect detection method based on a significance detection network aiming at the problems that the defect detection of the surface of a subway tunnel is difficult to finish with high efficiency and high precision, and the like, wherein a characteristic pyramid attention network is designed and established by considering different characteristics of different levels of characteristics, a context perception pyramid characteristic extraction module is designed aiming at high-level characteristics, and multi-scale void convolution and a channel attention mechanism are used for obtaining the high-level semantic characteristics; using a spatial attention mechanism for low-level features to suppress background noise, attention is focused on salient objects. Experiments show that the network has stronger robustness and significance detection effectiveness, and the method can well locate correct tunnel surface defects.
The foregoing shows and describes the basic principles, principal steps and advantages of the present invention. It is to be understood that the invention is not limited to the embodiments described above, which are described in the foregoing embodiments and description only to illustrate the principles of the invention, but rather that various changes and modifications may be made without departing from the spirit and scope of the invention, as defined by the appended claims, the description and equivalents thereof.

Claims (8)

1. A tunnel surface defect detection method based on a characteristic pyramid network is characterized by comprising the following steps:
the method comprises the following steps: acquiring tunnel images with surface defects, constructing a data set, and dividing the images in the data set into a training set and a testing set;
step two: preprocessing a tunnel surface image;
step three: marking defects on the surface of the tunnel, and marking cracks and hollow areas in the image by adopting a marking tool;
step four: constructing a significance detection network for defect detection;
step five: and constructing a training loss function, and detecting the tunnel surface defects by using the constructed detection network.
2. The method for detecting defects on the surface of a tunnel based on a feature pyramid network as claimed in claim 1, wherein: the first step is to collect tunnel images with surface defects, cut key information of the images, construct an image library, zoom, turn and mirror image expansion the images in the image library, and divide the images in the expanded image library into a training sample set and a testing sample set according to a ratio of 4: 1.
3. The method for detecting defects on the surface of a tunnel based on a feature pyramid network as claimed in claim 1, wherein: the second step is specifically to carry out pretreatment on the tunnel surface image and remove image noise by using guide filtering, and the specific process is as follows:
taking the image as a two-dimensional function, assuming that the output and input of the function satisfy a linear relationship within a two-dimensional window:
Figure FDA0002458375160000011
where q is the value of the output pixel, I is the input image, I and k are the pixel indices, ωkIs a window, akAnd bkIs the coefficient of the linear function when the window is located at k; taking the gradient on two sides of the above formula:
Figure FDA0002458375160000021
the difference between the fitting function output value and the true value is:
Figure FDA0002458375160000022
wherein p is a true value, and e represents a minimum value;
the coefficients of the image function are obtained by the least squares method:
Figure FDA0002458375160000023
Figure FDA0002458375160000024
where | ω | is the window ω |kNumber of middle pixels, μkIs that I is in the window omegakThe average value of (a) is,
Figure FDA0002458375160000025
is that I is in the window omegakThe variance in (a) is greater than or equal to,
Figure FDA0002458375160000026
is that the image p to be filtered is in the window omegakThe average value of (1);
when calculating the linear coefficients of each window, since each pixel is described by a plurality of linear functions, when specifically calculating the output value of a certain point, all the linear functions including the point are averaged:
Figure FDA0002458375160000027
in the above formula IiIs the intensity of the pixel i, ωkIs the total window containing pixel i, k is its central position,
Figure FDA0002458375160000028
represents the slope parameter of the linear regression,
Figure FDA0002458375160000029
representing the linear regression intercept parameter.
4. The method for detecting the defects of the tunnel surface based on the feature pyramid network as claimed in claim 1, wherein: and the fourth step is specifically to construct a significance detection network, and construct a feature pyramid network based on the VGG16 network. Two low-level blocks Conv1-2 and Conv2-2 of the VG 16 network are used for extracting low-level features of the image, and because background regions which interfere with the generation of the saliency map exist in the low-level features, a space attention mechanism is adopted to filter out some background details according to the high-level features so as to focus more on the foreground regions;
taking the characteristics generated by three high-level blocks Conv3-3, Conv4-5 and Conv5-3 of the VGG16 network as initial high-level characteristics, and in order to enable the finally extracted high-level characteristics to contain scale and shape invariance, adopting the convolution of holes with expansion rates of 3, 5 and 7 respectively to obtain the context information of multiple receptive fields; then, a channel attention mechanism is adopted to allocate a larger weight value for a channel with high response to the obvious target; cross-channel cascade connection is carried out on the feature mapping of the void convolution layers with different expansion rates and the 1 x 1 dimensional dimension reduction features to obtain three features with different scales and context perception information, and two smaller scale features are promoted to the scale with the same maximum feature scale through up-sampling; and finally, combining the high-level features and the low-level features through cross-channel cascade connection to serve as the output of the context-aware pyramid feature network.
5. The method for detecting defects on the surface of a tunnel based on a feature pyramid network as claimed in claim 4, wherein: adding a channel attention mechanism after the context perception pyramid feature extraction, and extracting the weighted multi-scale and multi-perception wild features:
CA＝F(vh，W)＝σ1(fc2((fc1(vh，W1))，W2)) (7)
where CA stands for channel attention, vhRepresenting the channel feature vector, W refers to the parameter in the attention block of each channel, σ1Sigmoid operation is indicated, fc is a full connection layer, and an activation function ReLU is indicated; the final output of the channel attention module is defined as the high level feature f of the feature pyramidhThe weighting value of (1):
Figure FDA0002458375160000031
the saliency map of the low-level features contains a large amount of detail information, and in order to obtain a detailed boundary between a salient object and a background, a spatial attention method is adopted to concentrate attention on a foreground area; to increase the receptive field to obtain global information without increasing parameters, convolutional layers with two kernels of 1 × k and k × 1, respectively, are used for high-level feature capture spatial interest:
Figure FDA0002458375160000041
Figure FDA0002458375160000042
wherein C is1、C2Representing intermediate results of two different orders of convolution operations, conv1、conv2Which represents a convolution operation, is a function of,
Figure FDA0002458375160000043
the parameters in the spatial attention module are represented,
Figure FDA0002458375160000044
an output of the channel attention module;
then, carrying out normalization processing on the coded spatial feature map mapped to [0,1] by using sigmoid operation:
Figure FDA0002458375160000045
where SA denotes spatial attention, W denotes a parameter in the spatial attention module, σ2Representing sigmoid operation;
the final output of the spatial attention module is the low-level feature flThe weighting value of (1):
Figure FDA0002458375160000046
6. the method for detecting defects on the surface of a tunnel based on a feature pyramid network as claimed in claim 1, wherein: the fifth step is to define a loss function by using the cross entropy between the final significance detection result graph and the theoretical result in the significance target detection:
Figure FDA0002458375160000047
where Y represents the theoretical result, P represents the significance of the network output, αsA balance parameter representing a positive sample and a negative sample;
the above loss function only provides a general guidance for generating a saliency map, and in order to highlight the generation of the boundary details of a salient object, a laplacian operator is used to obtain a theoretical boundary of the salient object and a boundary of the salient object in the saliency map output by a network:
Figure FDA0002458375160000051
Figure FDA0002458375160000052
wherein f represents a feature, x, y are cartesian coordinates of the xy plane, the laplace algorithm is executed using the gradient of the image, directly invoking the convolution operation conv () internally, tanh () is an activation function, KlaplaceIs laplacian, abs () represents absolute value operations;
the generation of the significant object boundary is then supervised by cross entropy loss:
Figure FDA0002458375160000053
finally, defining the total loss function as the weighted sum of the significant object detection loss and the significant object boundary detection loss:
L＝αLs+(1-α)LB(17)
wherein L is a loss function, LSFor significant target detection loss, LBFor significant object boundary detection loss α is a weight.
7. The method as claimed in claim 6, wherein α in formula (13) is defined ass＝0.528。
8. The method for detecting defects on the surface of a tunnel based on a feature pyramid network as claimed in claim 6, wherein: in the model training process, firstly, alpha is set to be 1.0, the initial learning rate is 0.01, a rough saliency map is generated through training, then alpha is adjusted to refine the boundaries of the saliency map, and the adjusted value of alpha is 0.7.
CN202010312482.XA 2020-04-20 2020-04-20 Tunnel surface defect detection method based on characteristic pyramid network Pending CN111524117A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010312482.XA CN111524117A (en) 2020-04-20 2020-04-20 Tunnel surface defect detection method based on characteristic pyramid network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010312482.XA CN111524117A (en) 2020-04-20 2020-04-20 Tunnel surface defect detection method based on characteristic pyramid network

Publications (1)

Publication Number Publication Date
CN111524117A true CN111524117A (en) 2020-08-11

Family

ID=71903745

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010312482.XA Pending CN111524117A (en) 2020-04-20 2020-04-20 Tunnel surface defect detection method based on characteristic pyramid network

Country Status (1)

Country Link
CN (1) CN111524117A (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113313668A (en) * 2021-04-19 2021-08-27 石家庄铁道大学 Subway tunnel surface disease feature extraction method
CN113313669A (en) * 2021-04-23 2021-08-27 石家庄铁道大学 Method for enhancing semantic features of top layer of surface disease image of subway tunnel
CN114441546A (en) * 2022-04-08 2022-05-06 湖南万航科技有限公司 Crack measurement endoscope
CN114841998A (en) * 2022-07-01 2022-08-02 山东嘉路包装科技有限公司 Artificial intelligence-based packaging printing abnormity monitoring method
CN117094999A (en) * 2023-10-19 2023-11-21 南京航空航天大学 Cross-scale defect detection method
CN117576109A (en) * 2024-01-19 2024-02-20 成都数之联科技股份有限公司 Defect detection method, device, equipment and storage medium

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109934821A (en) * 2019-03-22 2019-06-25 杭州睿工科技有限公司 A kind of part defect detection method and system
CN110084249A (en) * 2019-04-24 2019-08-02 哈尔滨工业大学 The image significance detection method paid attention to based on pyramid feature
CN110119687A (en) * 2019-04-17 2019-08-13 浙江工业大学 Detection method based on the road surface slight crack defect that image procossing and convolutional neural networks combine

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109934821A (en) * 2019-03-22 2019-06-25 杭州睿工科技有限公司 A kind of part defect detection method and system
CN110119687A (en) * 2019-04-17 2019-08-13 浙江工业大学 Detection method based on the road surface slight crack defect that image procossing and convolutional neural networks combine
CN110084249A (en) * 2019-04-24 2019-08-02 哈尔滨工业大学 The image significance detection method paid attention to based on pyramid feature

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
KAIMING HE等: "Guided Image Filtering", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 *
TING ZHAO等: "Pyramid Feature Attention Network for Saliency detection", 《2019 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 *

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113313668A (en) * 2021-04-19 2021-08-27 石家庄铁道大学 Subway tunnel surface disease feature extraction method
CN113313668B (en) * 2021-04-19 2022-09-27 石家庄铁道大学 Subway tunnel surface disease feature extraction method
CN113313669A (en) * 2021-04-23 2021-08-27 石家庄铁道大学 Method for enhancing semantic features of top layer of surface disease image of subway tunnel
CN114441546A (en) * 2022-04-08 2022-05-06 湖南万航科技有限公司 Crack measurement endoscope
CN114441546B (en) * 2022-04-08 2022-06-24 湖南万航科技有限公司 Crack measurement endoscope
CN114841998A (en) * 2022-07-01 2022-08-02 山东嘉路包装科技有限公司 Artificial intelligence-based packaging printing abnormity monitoring method
CN114841998B (en) * 2022-07-01 2022-09-06 山东嘉路包装科技有限公司 Artificial intelligence-based packaging printing abnormity monitoring method
CN117094999A (en) * 2023-10-19 2023-11-21 南京航空航天大学 Cross-scale defect detection method
CN117094999B (en) * 2023-10-19 2023-12-22 南京航空航天大学 Cross-scale defect detection method
CN117576109A (en) * 2024-01-19 2024-02-20 成都数之联科技股份有限公司 Defect detection method, device, equipment and storage medium
CN117576109B (en) * 2024-01-19 2024-04-02 成都数之联科技股份有限公司 Defect detection method, device, equipment and storage medium

Similar Documents

Publication Publication Date Title
CN111047551B (en) Remote sensing image change detection method and system based on U-net improved algorithm
CN111524117A (en) Tunnel surface defect detection method based on characteristic pyramid network
CN109145922B (en) Automatic cutout system
CN105678213B (en) Dual-mode mask person event automatic detection method based on video feature statistics
CN112287832A (en) High-resolution remote sensing image-based urban illegal building detection method
CN111611861B (en) Image change detection method based on multi-scale feature association
CN105787930A (en) Sharpness-based significance detection method and system for virtual images
Xia et al. PANDA: Parallel asymmetric network with double attention for cloud and its shadow detection
CN114677601A (en) Dam crack detection method based on unmanned aerial vehicle inspection and combined with deep learning
CN112801075B (en) Automatic rural road boundary line extraction method based on aerial image
CN114596500A (en) Remote sensing image semantic segmentation method based on channel-space attention and DeeplabV3plus
CN110852327A (en) Image processing method, image processing device, electronic equipment and storage medium
CN111753682A (en) Hoisting area dynamic monitoring method based on target detection algorithm
CN115457277A (en) Intelligent pavement disease identification and detection method and system
CN115497015A (en) River floating pollutant identification method based on convolutional neural network
CN116403121A (en) Remote sensing image water area segmentation method, system and equipment for multi-path fusion of water index and polarization information
CN116524189A (en) High-resolution remote sensing image semantic segmentation method based on coding and decoding indexing edge characterization
CN116994140A (en) Cultivated land extraction method, device, equipment and medium based on remote sensing image
Fan et al. Application of YOLOv5 neural network based on improved attention mechanism in recognition of Thangka image defects
CN112800851B (en) Water body contour automatic extraction method and system based on full convolution neuron network
CN113378642A (en) Method for detecting illegal occupation buildings in rural areas
Meng et al. A modified fully convolutional network for crack damage identification compared with conventional methods
Yin et al. Road Damage Detection and Classification based on Multi-level Feature Pyramids.
Cheng et al. Power pole detection based on graph cut
CN114998866A (en) Traffic sign identification method based on improved YOLOv4

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
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20200811