CN111721770A - Automatic crack detection method based on frequency division convolution - Google Patents

Automatic crack detection method based on frequency division convolution Download PDF

Info

Publication number
CN111721770A
CN111721770A CN202010540557.XA CN202010540557A CN111721770A CN 111721770 A CN111721770 A CN 111721770A CN 202010540557 A CN202010540557 A CN 202010540557A CN 111721770 A CN111721770 A CN 111721770A
Authority
CN
China
Prior art keywords
convolution
neural network
frequency
deep
training
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
CN202010540557.XA
Other languages
Chinese (zh)
Inventor
范衠
陈颖
李冲
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.)
Shantou University
Original Assignee
Shantou University
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 Shantou University filed Critical Shantou University
Priority to CN202010540557.XA priority Critical patent/CN111721770A/en
Publication of CN111721770A publication Critical patent/CN111721770A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/84—Systems specially adapted for particular applications
    • G01N21/88—Investigating the presence of flaws or contamination
    • G01N21/8851—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges
    • 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
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/84—Systems specially adapted for particular applications
    • G01N21/88—Investigating the presence of flaws or contamination
    • G01N21/8851—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges
    • G01N2021/8887—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges based on image processing techniques

Abstract

The embodiment of the invention discloses an automatic crack detection method based on frequency division convolution, which comprises the following steps: shooting a road image by using a camera, and creating a training set and a test set of the road crack image; creating a deep convolution neural network comprising a frequency division convolution, a frequency division transposition convolution, a hole convolution module and a jump connection structure; training the deep convolutional neural network by using the established training set; and testing the trained deep convolution neural network model by using the test set, and outputting a crack image. The method has the advantages of simple detection process, high detection efficiency, low labor intensity, convenience in carrying, strong operability and the like.

Description

Automatic crack detection method based on frequency division convolution
Technical Field
The invention relates to the field of structural health detection and evaluation, in particular to an automatic crack detection method based on frequency division convolution.
Background
With the rapid development of Chinese economy, the popularization and construction of the Chinese road network have been rapidly developed, and the perfectness and flatness of the road surface are important factors for ensuring the running of vehicles on the highway. Cracks are important signs of road damage, if the road surface is uneven, cracks and the like, the service life of the road and the safety of drivers are seriously influenced, and the health condition of the drivers needs to be evaluated regularly, so that the cracks of the road and the bridge are detected to be of great importance.
At present, the crack detection method of the road and the bridge is mainly based on the traditional image processing algorithm and human eye recognition. The crack detection and identification are carried out by human eyes alone, and the efficiency is not high. The image processing method is mainly used for detecting cracks of background images of the same material and texture, and the color images cannot be directly subjected to crack detection. The road crack detection based on the deep learning framework can realize the crack detection processing of the color image, can realize the end-to-end image processing and does not need the sliding block processing of the convolutional neural network. Therefore, the road crack detection method based on the deep learning frame can realize the automatic detection of the road crack. Therefore, how to improve the monitoring efficiency and effect of pavement crack detection is a technical problem to be overcome in the field of pavement crack detection.
Disclosure of Invention
Based on the above, the invention aims to provide an automatic crack detection method based on frequency division convolution. The method can solve the problems of low positioning precision, large error and the like in human eye observation and image processing crack detection.
In order to solve the above-mentioned problems of the prior art, an embodiment of the present invention provides an automatic crack detection method based on frequency division convolution, which specifically includes the following steps:
s1, shooting a road image by a camera, and creating a training set and a test set of the road crack image;
s2, creating a deep convolution neural network comprising a frequency division convolution, a frequency division transposition convolution, a hole convolution module and a jump connection structure;
s3, training the deep convolutional neural network by using the established training set;
and S4, testing the trained deep convolution neural network model by using the test set, and outputting a crack image.
Further, the step S1 specifically includes:
s11, shooting a crack image by using all intelligent terminals of the user, or dividing the crack image into a training set and a testing set by using a common crack image data set CFD, AigleRN and other crack image data sets;
s12, constructing a crack image database by the collected surface crack images of different structures, performing data enhancement on the constructed crack image database, expanding a data set, performing artificial label marking on the crack area of the crack image in the expanded crack image database, and then dividing the image in the crack image database into a training set and a test set.
Further, the step S2 specifically includes:
s21, building a deep neural network structure model: determining the number of layers of an encoder and a decoder in the deep convolutional neural network volume, the number of feature maps contained in the high frequency and the low frequency of each frequency division convolutional layer, the number of layers of pooling layers, the size and the training step length of a sampling kernel in each pooling layer, the number of layers of frequency division transpose convolutional layers, the number of feature maps contained in the high frequency and the low frequency of each deconvolution layer, a connection mode of jump connection and the size of a hollow ratio in a hollow convolutional module;
s22, selecting a training strategy of the deep neural network: selecting a cost function in the deep neural network training as a cross entropy loss function and Relu of an activation function, adding a weight attenuation regularization item into the loss cost function, and adding dropout into a convolutional layer to reduce overfitting, wherein an optimization algorithm SGD is used in the deep neural network training;
s23, building frequency division convolution layer X ═ XH,XLY ═ YH，YLDenotes input and output, where YL＝YH →L+YL→LAnd Y isH＝YH→H+YL→HWhich is indicative of a change in the output frequency,
WH＝[WH→H，WL→H]，WL＝[WH→L，WL→L]representing the variation of the frequency of the convolution kernel, the variation of high and low frequencies in the frequency division convolution operation is represented by the following formula:
Figure BDA0002537019710000031
Figure BDA0002537019710000032
wherein (p, q) represents the position of the pixel point, k represents the size of the convolution kernel, σ (·) represents the activation function, b represents the bias variation, and XH，XLHigh and low frequency profiles, Y, representing the input profile, respectivelyH，YLHigh-frequency and low-frequency feature maps respectively representing the output feature map, H → L representing the feature map converted from high frequency to low frequency, L → H representing the feature map converted from low frequency to high frequency, H → H representing the feature map converted from high frequency to high frequency, L → L representing the feature map converted from low frequency to low frequency, m and n being used to determine the range of the local receptive field with (p, q) as the pixel center point on the input X;
s24, the built frequency division transpose convolution layer X ═ XH,XLAnd
Figure BDA0002537019710000033
representing an input and an output, wherein
Figure BDA0002537019710000034
And
Figure BDA0002537019710000035
representing variation of the output of the frequency-dividing transpose, WH＝[WH→H，WL→H]，WL＝[WH→L，WL→L]The variation of the high and low frequencies of the convolution is expressed, and the variation of the high and low frequencies in the frequency division transposition convolution operation is expressed by the following formula:
Figure BDA0002537019710000036
Figure BDA0002537019710000037
Figure BDA0002537019710000038
and
Figure BDA0002537019710000039
respectively representing high-frequency and low-frequency characteristic graphs of a frequency division transposition convolution output characteristic graph, wherein values of m and n are used for determining the range of a local receptive field taking (p, q) as a pixel central point on an input X, and k represents the size of a convolution kernel;
s25, connecting the encoder and the decoder in the deep convolutional neural network through jump connection;
s26, in the deep convolutional neural network, the input image and the encoder part and all encoders are connected through jump connection, so that the transfer of image information can be realized;
s27, in the cavity convolution module in the deep convolution neural network, the input of the cavity convolution module is the output of the feature map in the last convolution layer of the encoder, the cavity convolution module is composed of convolution layers with different cavity rates, and the output of the cavity convolution module is obtained by superposition and fusion of feature maps obtained by convolution with different cavity rates;
s28, using a deep learning library package in the deep convolutional neural network: caffe, Tensorflow and PyTorch realize the deep neural network structure, model training is carried out according to the divided training set and the divided testing set, parameters of the deep neural network are learned by continuously reducing function values of the loss function, and parameter values in the deep neural network model are determined.
Further, the step S3 specifically includes:
and S31, training the deep convolutional neural network by using a training set according to the steps S21, S22, S23, S24, S25, S26, S27 and S28, continuously optimizing parameters of the neural network through backward propagation, reducing the value of a loss function, optimizing the network, and realizing end-to-end training.
Further, the step S4 specifically includes:
s41, testing the trained neural network model by using a test set according to the step S31;
and S42, normalizing the output value of the neural network model and outputting a probability chart of the crack image.
Drawings
FIG. 1 is a flow chart of an automated crack detection method based on void convolution according to the present invention;
FIG. 2 is a flow chart of a deep convolutional neural network model according to an embodiment of the present invention;
FIG. 3 is a flow diagram of a frequency-division convolution model according to an embodiment of the present invention;
FIG. 4 is a flow diagram of a frequency division transpose convolution model according to an embodiment of the present invention;
FIG. 5 is a flow diagram of a hole convolution module in accordance with an embodiment of the present invention;
FIG. 6 is a diagram of the output of the deep convolutional neural network in accordance with an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the accompanying drawings.
The experimental environment of the embodiment of the invention is an outdoor environment which is an experimental building, a wall and a road surface in a highway. In this embodiment, the crack image is selected as a public area of the outdoor environment.
In this embodiment, a PC including an Nvidia video card is used. The implementation method is an Ubuntu method, a Tensorflow method platform is built, and an open source software library in Tensorflow is adopted.
Referring to fig. 1, an automatic crack detection method based on void convolution according to an embodiment of the present invention includes the following steps:
and S1, shooting road images by using a camera, and creating a training set and a testing set of road crack images.
In the present example, a common data set CFD is used, which contains 118 original color images and 118 label data images, and the data set is divided into a training set test set, wherein each of the training sets contains 100 original color images and corresponding 100 label data images, and the test set contains 18 original color images and corresponding 18 label data images.
Meanwhile, in order to expand the image data volume and perform data enhancement on the crack images in the CFD data set, the original color images and the label data images in each piece of divided data are rotated and cut to increase the number of the crack images in the embodiment of the invention.
And S2, creating a deep convolutional neural network comprising an encoder, a decoder, a hole convolutional module and a jump connection structure.
The deep convolutional neural network model adopted in the embodiment of the invention is based on a U-net model, and the network model is improved. Please refer to fig. 2 for a flowchart of a deep convolutional neural network model used in an embodiment of the present invention.
The establishment of the deep neural network large model structure comprises the steps of determining the number of layers of an encoder and a decoder in a deep convolutional neural network volume, the number of characteristic graphs contained in high frequency and low frequency in each frequency division convolutional layer, the number of layers of pooling layers, the size and training step length of a sampling kernel in each pooling layer, the number of layers of frequency division transpose convolutional layers, the number of characteristic graphs contained in high frequency and low frequency in each deconvolution layer, the connection mode of jump connection and the size of hollow ratio in a hollow convolutional module.
Selecting a training strategy of the deep neural network: the cost function in the deep neural network training is selected as a cross entropy loss function and Relu of an activation function, meanwhile, a weight attenuation regularization item is added into the loss cost function, dropout is added into a convolutional layer to reduce overfitting, and an optimization algorithm SGD is used in the deep neural network training.
In the embodiment of the present invention, a frequency division convolution layer (as shown in fig. 3) X ═ XH,XLY ═ YH,YLDenotes input and output, where YL＝YH→L+YL→LAnd Y isH＝YH→H+YL→HWhich is indicative of a change in the output frequency,
WH＝[WH→H，WL→H]，WL＝[WH→L，WL→L]representing the variation of the frequency of the convolution kernel, the variation of high and low frequencies in the frequency division convolution operation is represented by the following formula:
Figure BDA0002537019710000061
Figure BDA0002537019710000062
wherein (p, q) represents the position of the pixel point, k represents the size of the convolution kernel, σ (·) represents the activation function, b represents the bias variation, and XH，XLHigh and low frequency profiles, Y, representing the input profile, respectivelyH，YLHigh-frequency and low-frequency feature maps respectively representing the output feature map, H → L representing the feature map converted from high frequency to low frequency, L → H representing the feature map converted from low frequency to high frequency, H → H representing the feature map converted from high frequency to high frequency, L → L representing the feature map converted from low frequency to low frequency, m and n being used to determine the range of the local receptive field with (p, q) as the pixel center point on the input X;
in the embodiment of the invention, the frequency division transpose convolution layer (as shown in fig. 4) is built, wherein X is { X ═ XH，XLAnd
Figure BDA0002537019710000063
representing an input and an output, wherein
Figure BDA0002537019710000064
And
Figure BDA0002537019710000065
representing variation of the output of the frequency-dividing transpose, WH＝[WH→H，WL→H]，WL＝[WH→L，WL→L]The variation of the high and low frequencies of the convolution is expressed, and the variation of the high and low frequencies in the frequency division transposition convolution operation is expressed by the following formula:
Figure BDA0002537019710000066
Figure BDA0002537019710000067
Figure BDA0002537019710000068
and
Figure BDA0002537019710000069
the high-frequency characteristic diagram and the low-frequency characteristic diagram respectively represent a frequency division transposition convolution output characteristic diagram, the values of m and n are used for determining the range of a local receptive field taking (p, q) as the central point of a pixel on an input X, and k represents the size of a convolution kernel.
In the embodiment of the invention, an activation function adopted by a convolution layer in a deep neural network large model is a ReLU, a sigmoid activation function is adopted in the output of the last layer to output a logit, and a loss function formula used in the embodiment of the invention is as follows:
Figure BDA0002537019710000071
where α and β are hyper-parameters,
Figure BDA0002537019710000072
is the true value of the tag data and,
Figure BDA0002537019710000073
is a predicted value of the original image through the depth network. Meanwhile, the embodiment of the invention uses an Adam optimization algorithm for optimization, and the learning rate is 0.001 to minimize the loss function.
In the embodiment of the invention, the encoder part and the decoder part in the U-net structure in the deep convolutional neural network are connected through a contract, and the jump connection function can realize the transmission of the texture information of the image to the decoder, thereby avoiding the loss of image characteristics caused by a pooling layer or downsampling.
Meanwhile, in the deep convolutional neural network, the input image and the encoder part and all the encoders are connected through jumping connection, so that the transmission of image information can be realized, the input image can still keep the original characteristic information of the input image through jumping connection input after a series of convolution and pooling, and the loss of image texture information is avoided.
The deep learning library of the deep neural network used in the embodiment of the invention is TensorFlow, cross validation is carried out according to the divided training set and validation set by utilizing the deep learning library, the parameter of the deep neural network is learned by continuously reducing the loss function, and the value of the parameter in the large model of the deep neural network is determined.
In the hole convolution module (as shown in fig. 5) in the deep convolution neural network, the period input is the output of the feature map in the last convolution layer of the encoder, and the output of the hole convolution module is obtained by superposition and fusion of the feature maps obtained by convolution with different hole rates.
The deep convolutional neural network structure is realized by using a deep learning library comprising Caffe and Tensorflow, model training is carried out according to a divided training set and a verification set, parameters of the deep neural network are learned by continuously reducing function values of a loss function, and parameter values in a deep neural network model are determined.
And S3, training the deep convolutional neural network by using the created training set.
The deep convolutional neural network is trained by utilizing a training set, parameters of the neural network are continuously optimized through back propagation, the value of a loss function is reduced, the network is optimized, and end-to-end training is realized.
And S4, testing the trained deep convolution neural network model by using the test set, and outputting a crack image.
Testing the trained neural network model by using the test set, then normalizing the output value of the neural network model, and outputting a probability map of the crack image, referring to fig. 6, which sequentially comprises from left to right: real images, labels, prediction results.
The above examples only represent the preferred embodiments of the present invention, and the description thereof is more specific and detailed, but not to be construed as limiting the scope of the present invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the inventive concept, which falls within the scope of the present invention. Therefore, the protection scope of the present patent shall be subject to the appended claims.

Claims (5)

1. An automatic crack detection method based on frequency division convolution is characterized by specifically comprising the following steps of:
s1, shooting a road image by a camera, and creating a training set and a test set of the road crack image;
s2, creating a deep convolution neural network comprising a frequency division convolution, a frequency division transposition convolution, a hole convolution module and a jump connection structure;
s3, training the deep convolutional neural network by using the established training set;
and S4, testing the trained deep convolution neural network model by using the test set, and outputting a crack image.
2. The method according to claim 1, wherein the step S1 specifically includes:
s11, shooting a crack image by using an intelligent terminal, or dividing the crack image into a training set and a testing set by using a common crack image data set CFD and AigleRN;
s12, constructing a crack image database by the collected surface crack images of different structures, performing data enhancement on the constructed crack image database, expanding a data set, performing artificial label marking on the crack area of the crack image in the expanded crack image database, and then dividing the image in the crack image database into a training set and a test set.
3. The method according to claim 2, wherein the step S2 specifically includes:
s21, building a deep neural network structure model: determining the number of layers of an encoder and a decoder in the deep convolutional neural network volume, the number of feature maps contained in the high frequency and the low frequency of each frequency division convolutional layer, the number of layers of pooling layers, the size and the training step length of a sampling kernel in each pooling layer, the number of layers of frequency division transpose convolutional layers, the number of feature maps contained in the high frequency and the low frequency of each deconvolution layer, a connection mode of jump connection and the size of a hollow ratio in a hollow convolutional module;
s22, selecting a training strategy of the deep neural network: selecting a cost function in the deep neural network training as a cross entropy loss function and Relu of an activation function, adding a weight attenuation regularization item into the loss cost function, and adding dropout into a convolutional layer to reduce overfitting, wherein an optimization algorithm SGD is used in the deep neural network training;
s23, building frequency division convolution layer X ═ XH,XLY ═ YH,YLDenotes input and output, where YL＝YH→L+YL →LAnd YH＝YH→H+YL→HRepresenting a change in output frequency, WH＝[WH→HWL→H]，WL＝[WH→L，WL→L]Representing the variation of the frequency of the convolution kernel, the variation of high and low frequencies in the frequency division convolution operation is represented by the following formula:
Figure FDA0002537019700000021
Figure FDA0002537019700000022
wherein (p, q) represents the position of the pixel point, k represents the size of the convolution kernel, σ (·) represents the activation function, b represents the bias variation, and XH，XLHigh and low frequency profiles, Y, representing the input profile, respectivelyH，YLRespectively representing output characteristic diagramsHigh-frequency and low-frequency feature maps, wherein H → L indicates that the feature map is converted from high frequency to low frequency, L → H indicates that the feature map is converted from low frequency to high frequency, H → H indicates that the feature map is converted from high frequency to high frequency, L → L indicates that the feature map is converted from low frequency to low frequency, and m and n are used for determining the range of the local receptive field with (p, q) as the central point of the pixel on the input X;
s24, the built frequency division transpose convolution layer X ═ XH,XLAnd
Figure FDA0002537019700000023
representing an input and an output, wherein
Figure FDA0002537019700000024
And
Figure FDA0002537019700000025
representing variation of the output of the frequency-dividing transpose, WH＝[WH→H，WL →H]，WL＝[WH→L，WL→L]The variation of the high and low frequencies of the convolution is expressed, and the variation of the high and low frequencies in the frequency division transposition convolution operation is expressed by the following formula:
Figure FDA0002537019700000026
Figure FDA0002537019700000027
wherein the content of the first and second substances,
Figure FDA0002537019700000028
and
Figure FDA0002537019700000029
high and low frequency profiles representing the frequency division transposed convolution output profile, respectively, the values of m and n being used to determine the range of the local field of view on the input X with (p, q) as the pixel center, kRepresents the convolution kernel size;
s25, connecting the encoder and the decoder in the deep convolutional neural network through jump connection;
s26, in the deep convolutional neural network, the input image and the encoder part and all encoders are connected through jump connection, so that the transfer of image information can be realized;
s27, in the cavity convolution module in the deep convolution neural network, the input of the cavity convolution module is the output of the feature map in the last convolution layer of the encoder, the cavity convolution module is composed of convolution layers with different cavity rates, and the output of the cavity convolution module is obtained by superposition and fusion of feature maps obtained by convolution with different cavity rates;
s28, using a deep learning library package in the deep convolutional neural network: caffe, Tensorflow and PyTorch realize the deep neural network structure, model training is carried out according to the divided training set and the divided testing set, parameters of the deep neural network are learned by continuously reducing function values of the loss function, and parameter values in the deep neural network model are determined.
4. The method according to claim 3, wherein the step S3 specifically includes:
and S31, training the deep convolutional neural network by using a training set according to the steps S21, S22, S23, S24, S25, S26, S27 and S28, continuously optimizing parameters of the neural network through backward propagation, reducing the value of a loss function, optimizing the network, and realizing end-to-end training.
5. The method according to claim 4, wherein the step S4 specifically includes:
s41, testing the trained neural network model by using a test set according to the step S31;
and S42, normalizing the output value of the neural network model and outputting a probability chart of the crack image.
CN202010540557.XA 2020-06-12 2020-06-12 Automatic crack detection method based on frequency division convolution Pending CN111721770A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010540557.XA CN111721770A (en) 2020-06-12 2020-06-12 Automatic crack detection method based on frequency division convolution

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010540557.XA CN111721770A (en) 2020-06-12 2020-06-12 Automatic crack detection method based on frequency division convolution

Publications (1)

Publication Number Publication Date
CN111721770A true CN111721770A (en) 2020-09-29

Family

ID=72566787

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010540557.XA Pending CN111721770A (en) 2020-06-12 2020-06-12 Automatic crack detection method based on frequency division convolution

Country Status (1)

Country Link
CN (1) CN111721770A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112233105A (en) * 2020-10-27 2021-01-15 江苏科博空间信息科技有限公司 Road crack detection method based on improved FCN
CN113506281A (en) * 2021-07-23 2021-10-15 西北工业大学 Bridge crack detection method based on deep learning framework

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2015047466A2 (en) * 2013-06-05 2015-04-02 Innersense, Inc. Bi-phasic applications of real & imaginary separation, and reintegration in the time domain
CN104574362A (en) * 2014-12-01 2015-04-29 汕头大学 Passive visual system-based molten pool edge extraction method
CN105844630A (en) * 2016-03-21 2016-08-10 西安电子科技大学 Binocular visual image super-resolution fusion de-noising method
CN110619309A (en) * 2019-09-19 2019-12-27 天津天地基业科技有限公司 Embedded platform face detection method based on octave convolution sum YOLOv3
CN111179244A (en) * 2019-12-25 2020-05-19 汕头大学 Automatic crack detection method based on cavity convolution

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2015047466A2 (en) * 2013-06-05 2015-04-02 Innersense, Inc. Bi-phasic applications of real & imaginary separation, and reintegration in the time domain
CN104574362A (en) * 2014-12-01 2015-04-29 汕头大学 Passive visual system-based molten pool edge extraction method
CN105844630A (en) * 2016-03-21 2016-08-10 西安电子科技大学 Binocular visual image super-resolution fusion de-noising method
CN110619309A (en) * 2019-09-19 2019-12-27 天津天地基业科技有限公司 Embedded platform face detection method based on octave convolution sum YOLOv3
CN111179244A (en) * 2019-12-25 2020-05-19 汕头大学 Automatic crack detection method based on cavity convolution

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
莫嘉杰: "基于分频卷积神经网络的视网膜血管分割", 《万方学位论文》 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112233105A (en) * 2020-10-27 2021-01-15 江苏科博空间信息科技有限公司 Road crack detection method based on improved FCN
CN113506281A (en) * 2021-07-23 2021-10-15 西北工业大学 Bridge crack detection method based on deep learning framework
CN113506281B (en) * 2021-07-23 2024-02-27 西北工业大学 Bridge crack detection method based on deep learning framework

Similar Documents

Publication Publication Date Title
CN111179244B (en) Automatic crack detection method based on cavity convolution
CN111127449B (en) Automatic crack detection method based on encoder-decoder
CN111126258B (en) Image recognition method and related device
CN112668494A (en) Small sample change detection method based on multi-scale feature extraction
CN111738054B (en) Behavior anomaly detection method based on space-time self-encoder network and space-time CNN
CN113283356B (en) Multistage attention scale perception crowd counting method
CN111611861B (en) Image change detection method based on multi-scale feature association
CN110717886A (en) Pavement pool detection method based on machine vision in complex environment
CN113177560A (en) Universal lightweight deep learning vehicle detection method
CN111721770A (en) Automatic crack detection method based on frequency division convolution
CN112733693B (en) Multi-scale residual error road extraction method for global perception high-resolution remote sensing image
CN116596151B (en) Traffic flow prediction method and computing device based on time-space diagram attention
CN111199539A (en) Crack detection method based on integrated neural network
CN116524189A (en) High-resolution remote sensing image semantic segmentation method based on coding and decoding indexing edge characterization
CN115810149A (en) High-resolution remote sensing image building extraction method based on superpixel and image convolution
WO2022100607A1 (en) Method for determining neural network structure and apparatus thereof
CN112132867B (en) Remote sensing image change detection method and device
CN111160282B (en) Traffic light detection method based on binary Yolov3 network
CN111738324B (en) Multi-frequency and multi-scale fusion automatic crack detection method based on frequency division convolution
CN111079811A (en) Sampling method for multi-label classified data imbalance problem
CN115497006B (en) Urban remote sensing image change depth monitoring method and system based on dynamic mixing strategy
CN116778318A (en) Convolutional neural network remote sensing image road extraction model and method
CN115880557A (en) Pavement crack extraction method and device based on deep learning
CN115578693A (en) Construction safety early warning method and device based on significance neural network model
CN111091061B (en) Vehicle scratch detection method based on video analysis

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: 20200929