CN108520516A - A kind of bridge pavement Crack Detection and dividing method based on semantic segmentation - Google Patents

A kind of bridge pavement Crack Detection and dividing method based on semantic segmentation Download PDF

Info

Publication number
CN108520516A
CN108520516A CN201810309151.3A CN201810309151A CN108520516A CN 108520516 A CN108520516 A CN 108520516A CN 201810309151 A CN201810309151 A CN 201810309151A CN 108520516 A CN108520516 A CN 108520516A
Authority
CN
China
Prior art keywords
layers
crack
denseblock
convolution
semantic segmentation
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
CN201810309151.3A
Other languages
Chinese (zh)
Inventor
李良福
孙瑞赟
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.)
Shaanxi Normal University
Original Assignee
Shaanxi Normal 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 Shaanxi Normal University filed Critical Shaanxi Normal University
Priority to CN201810309151.3A priority Critical patent/CN108520516A/en
Publication of CN108520516A publication Critical patent/CN108520516A/en
Pending legal-status Critical Current

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/0002Inspection of images, e.g. flaw detection
    • G06T7/0004Industrial image inspection
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/10Segmentation; Edge detection
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/10Image acquisition modality
    • G06T2207/10004Still image; Photographic image
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20081Training; Learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20084Artificial neural networks [ANN]
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/30Subject of image; Context of image processing
    • G06T2207/30108Industrial image inspection

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Quality & Reliability (AREA)
  • Health & Medical Sciences (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

Invention belongs to computer vision, deep learning field herein;More particularly to a kind of bridge pavement Crack Detection and dividing method based on semantic segmentation, the sample concentrated to data carries out artificial semantic segmentation, makes the label of training sample;Secondly, the amount of images concentrated to data by data enhancing expands;Then, ready training set input FC DenseNet103 network models are trained, finally carry out crack extract using the crack image of collected test set;Traditional Crack Detection mostly uses greatly the methods of edge detection, morphology or thresholding, need artificially setting and adjusting parameter, the deep learning method being currently known is established affected by noise small, crack target clearly on the basis of, underestimate the complexity of bridge pavement image, it is difficult to meet the needs of engineer application;Present invention combination semantic segmentation algorithm provides a kind of bridge pavement crack suitable under complex background detection and dividing method automatically.

Description

A kind of bridge pavement Crack Detection and dividing method based on semantic segmentation
Technical field
The invention belongs to computer vision, deep learning fields, and in particular to a kind of bridge pavement based on semantic segmentation Crack Detection and dividing method.
Background technology
Effective means are taken to be detected and divide Bridge Crack, to ensuring the safety and normal operation of public transport It plays a very important role, has received the extensive concern of domestic and international academia for a long time.From traditional image procossing The big heat of technology machine learning and deep learning by now, domestic and foreign scholars are constantly by new technology for bridge pavement crack Detection and segmentation, and achieve some outstanding achievements in research;Kirschke etc. proposes a kind of based on image histogram Pavement crack partitioning algorithm, the algorithm extract crack information according to image histogram feature, using Threshold segmentation;Sun Liang et al. is carried Gone out the crack detection method based on adaptive threshold Canny algorithms, this method for can only artificial selected threshold the shortcomings that into Improvement is gone, in conjunction with the Grad of each pixel in Harris feature detection algorithms and image, fracture image extracts; Landstrom et al. proposes a kind of full-automatic Crack Monitoring system, algorithm combining form processing and logistic regression algorithm Fracture is detected, and is filtered out noise using Statistical Classification method, is improved accuracy of detection;With the development of science and technology, will The algorithm that deep learning is applied to Bridge Crack detection occurs immediately;Chen Yao et al. propose the bridge based on Climbing Robot The detection in beam crack and sorting technique, the support in non-contact capture image and machine learning in this method application machine vision Vector machine algorithm, and wavelet transformation is combined to complete the extraction and identification in crack with morphological analysis, start to depth The exploration in habit field;Liu Hong public affairs et al. propose the detection of the Bridge Crack based on convolutional neural networks and identification, which combines Machine vision and convolutional neural networks technology, improve the convolutional neural networks model (CNN) of classification of rifts, final to propose that one kind is new Intelligent Crack Detection scheme;Zhang lei et al. propose the pavement crack detection algorithm based on depth convolutional neural networks, The algorithm has trained the depth convolutional neural networks of supervision, classifies to each image block in the image that is collected into, and Generate good result;Why Bridge Crack detection algorithm set forth above achieves good experiment effect, is because of acquisition The contrast of image is very high, and noise is very low, and scene is fairly simple, and any barrier is not present;And this reason in actual life The case where thinking is fewer, and actual bridge pavement often has the barriers such as water stain, lane line, fallen leaves or pavement texture Extremely complex, conventional method is difficult to meet requirement of engineering.
For the deficiency studied above, it is proposed that the bridge pavement seam under the complex background based on semantic segmentation detects and divides Segmentation method.
Invention content
In order to solve it is existing in the prior art restricted by environmental factor, under complex background detection and dividing method Inaccurate problem, the present invention provides a kind of bridge pavement Crack Detection and dividing method based on semantic segmentation;The present invention Technical problems to be solved are achieved through the following technical solutions:
A kind of bridge pavement Crack Detection and dividing method based on semantic segmentation, include the following steps:
Step 1:Dataset acquisition, by continuously taking pictures along pavement crack direction, obtaining crack image pattern, and people The corresponding label of making of work carries out semantic segmentation mark to the crack image, and crack is labeled as one kind in fracture image Solid color, all chaff interferents and background in the image of crack except crack are all set to another unified single face Color;Data set is expanded, and is training set and test set by data set random assortment after amplification;
Step 2:By crack image in training set, input FC-Densenet103 network models are trained by several times, specifically Method is as follows:
Step 1:Crack image in training set is carried out to the convolution of a 3*3;
Step 2:And convolution results are inputted to the DenseBlock modules for including 4 layers layers;
Step 3:Step 2 result is subjected to Transition Down operations, reduces crack image resolution ratio;
Step 4:It is 5 layers, 7 layers, 10 layers, 12 layers that the DenseBlock modules layers layer numbers, which are set gradually, according to It is secondary to be repeated 4 times step 2 and step 3;
Step 5:The result of the step 4 is inputted to the Bottleneck being made of 15 layers, completes to adopt under whole Sample, and carry out the attended operation of multiple features;
Step 6:Upper layer output result is inputted into the up-sampling channel being made of Transition Up and DenseBlock, It is 12 layers that DenseBlock, which corresponds to the layers numbers of plies in down-sampling,;
Step 7:The layers numbers of plies of DenseBlock in the step 6 are set as 10,7,5,4 successively, are repeated 4 times step 6;
Step 8:1*1 convolution operations are carried out to the output result of the step 7;
Step 9:Step 8 result input is judged for softmax layers, the probability in crack and non-crack is exported;
Step 3:After the completion of the step 2 training, by trained FC-Densenet103 network models to test It concentrates crack image to be tested, obtains test result.
Further, in the step 2 103 network models of FC-DenseNet include by DenseBlock and The down-sampling path of Transition Down composition, and the up-sampling road that is made of DenseBlock and Transition Up Diameter and softmax functions.
Further, 103 network models of the FC-DenseNet are made of 103 convolutional layers altogether:First convolution is straight It connects and acts on input picture, there is 38 convolutional layers in the down-sampling path being made of DenseBlock, in bottleneck Bottleneck There are 15 convolutional layers, there is 38 convolutional layers, 103 nets of the FC-DenseNet in the up-sampling path being made of DenseBlock Also include that 5 Transition Down, each Transition Down include a convolution and 5 in network model Transition Up, each Transition Up include the 1*1 convolution of last layer in a transposition convolution and network.
Further, after the step 5 completes down-sampling, operation is attached to multiple output features, is embodied as Such as formula (1):
Xl=Hl([X0X1..., Xl-1] (1)
L indicates the number of plies, X in formulalIndicate l layers of output, [X0X1...Xl-1] indicate to connect 0 Dao l-1 layers output characteristic pattern It connects, Hl () indicates the combination of the convolution of Batch Normalization, ReLU and 3*3.
Further, 4 layers layers of DenseBlock modules are included in the step 2, layer layers by Batch Normalization, ReLU, 3*3 convolution sum Dropout are constituted, and the Dropout refers to training in deep learning network It is temporarily abandoned, so that each mini-batch neural network unit by Cheng Zhong according to certain probability from network Networks all different in training, wherein Dropout=0.2.
Further, the Batch Normalization specific algorithms are as follows:
The each layer of input of Batch Normalization algorithms in each iteration is all normalized, will be defeated Entering the distributions of data, to be normalized to mean value be 0, and the distribution that variance is 1 is specific such as formula (2):
Wherein, xkIndicate the kth dimension of input data, E [xk] indicate the average value that k is tieed up,Indicate standard deviation;
Batch Normalization algorithms are provided with two the variable γ and β that can learn, specifically such as formula (3),
γ and β is for restoring the data distribution that last layer should be acquired.
Further, the ReLU include continuous nonlinear activation function Activation Function with Rectifier, it is specific to calculate as shown in formula (4):
Rectifier (x)=max (0, x) (4).
Further, the Transition Down are used to reduce the Spatial Dimension of characteristic pattern, by Batch Normalization, RELU, 1*1 convolution and the operation of 2*2 pondizations form, and wherein the convolution of 1*1 is used to preserve the number of characteristic pattern Amount, the pondization of 2*2 operate the resolution ratio for reducing characteristic pattern, and the Transition Up are made of a transposition convolution, are used In the spatial resolution for restoring input picture, the transposition convolution only uses the characteristic pattern of the last one DenseBlock.
Further, the Bottleneck is by 15 layers layers of DenseBlock constituted.
Further, 103 network models of the FC-DenseNet use Filter Concatenation characteristic pattern Get up by deep linking.
Compared with prior art, beneficial effects of the present invention:
Traditional Crack Detection mostly uses greatly the methods of edge detection, morphology or thresholding, need artificially be arranged and Adjusting parameter.With the fast development of deep learning, this method has been successfully applied to Bridge Crack detection field, though with certainly Adaptability, it is no longer necessary to artificial setting and adjusting parameter, but the deep learning method being currently known is established affected by noise It is small, crack target clearly on the basis of, underestimate the complexity of bridge pavement image, it is difficult to meet the needs of engineer application, The present invention for bridge pavement Crack Detection and extracts DenseNet structures, and achieves remarkable result, has broken original bridge The single limitation of background, more with practical value in beam Crack Detection.
Description of the drawings
Fig. 1 is bridge pavement crack image.
Fig. 2 is the image after crack image and the visualization of crack image tag.
Fig. 3 is data images after amplification.
Fig. 4 is DenseBlock schematic diagrames.
Fig. 5 is the concrete structure of 103 network models of FC-DenseNet.
Fig. 6 is layers layers of concrete structure.
Fig. 7 is the concrete structure of Transition Down.
Fig. 8 is the concrete structure of Transition Up.
Fig. 9 is partial test result figure.
Specific implementation mode
Further detailed description is done to the present invention with reference to specific embodiment, but embodiments of the present invention are not limited to This.
A kind of bridge pavement Crack Detection and dividing method based on semantic segmentation, include the following steps:
Step 1:Dataset acquisition is flown along pavement crack direction by unmanned plane, and continuously taken pictures, and crack pattern is obtained Picture;Fracture image carries out semantic segmentation, the corresponding label of making for needing the sample concentrated to data artificial;Specifically marking During note, crack is labeled as a kind of solid color in fracture image, all chaff interferents in the image of crack except crack with And background is all set to another unified solid color;The detection of bridge pavement seam and dividing method based on semantic segmentation It realizes, needs the pavement crack image of label a large amount of, with semantic classes as training set and test set;But at present Until, disclosed not yet, with class label, for bridge pavement crack image semantic segmentation the data acquisition system in the whole world;Cause This, it is necessary to oneself creates the data acquisition system for bridge pavement crack image detection and segmentation;Due to making image tag manually Also there is sizable workload, therefore we should use the data set amplification method of efficiency highest, calculation amount minimum;It is specific to use Data enhancing method be:
A. 224 × 224 random fragment is extracted from 256 × 256 image;
B. horizontal reflection and vertical reflection are carried out to the fragment cut at random;
And data set random assortment is training set and test set after expanding;
Step 2:By crack image in training set, input FC-Densenet103 network models are trained by several times, specifically Method is as follows:
Step 1:Crack image in training set is carried out to the convolution of a 3*3;
Step 2:And convolution results are inputted to the DenseBlock modules for including 4 layers layers;
Step 3:Step 2 result is subjected to Transition Down operations, reduces crack image resolution ratio;
Step 4:It is 5 layers, 7 layers, 10 layers, 12 layers that DenseBlock module layers layer numbers, which are set gradually, is weighed successively Multiple 4 steps 2 and step 3;
Step 5:The result of step 4 is inputted to the Bottleneck being made of 15 layers, completes whole down-samplings, and Carry out the attended operation of multiple features;
Step 6:Upper layer output result is inputted into the up-sampling channel being made of Transition Up and DenseBlock, It is 12 layers that DenseBlock, which corresponds to the layers numbers of plies in down-sampling,;
Step 7:The layers numbers of plies of DenseBlock in step 6 are set as 10,7,5,4 successively, are repeated 4 times step 6;
Step 8:1*1 convolution operations are carried out to the output result of step 7;
Step 9:The input of step 8 result is judged for softmax layers, the probability in crack and non-crack is exported;
Step 3:After the completion of step 2 training, by trained FC-Densenet103 network models in test set Crack image is tested, and test result is obtained.
As shown in Fig. 5, Fig. 6, Fig. 7, Fig. 8,103 network models of FC-DenseNet include by DenseBlock in step 2 And the down-sampling path of Transition Down composition, and the up-sampling that is made of DenseBlock and Transition Up The up-sampling path of path and softmax functions, DenseBlock and Transition Up compositions is inputted for restoring Image spatial resolution, wherein m represent the number of characteristic pattern, and c represents last classification number.
103 network models of FC-DenseNet are made of 103 convolutional layers altogether:First convolution directly acts on input figure Picture has 38 convolutional layers in the down-sampling path being made of DenseBlock, there is 15 convolutional layers in bottleneck Bottleneck, by There are 38 convolutional layers in the up-sampling path of DenseBlock compositions, also includes 5 in 103 network models of FC-DenseNet Transition Down, each Transition Down include a convolution and 5 Transition Up, each Transition Up include the 1*1 convolution of last layer in a transposition convolution and network.
Dense Convolutional Network (DenseNet) are a kind of convolutional Neural nets with intensive connection Network.In the network, there is direct connection between any two layers, that is to say, that the input that each layer of network is all front institute There is the union that layer exports, and the characteristic pattern that this layer is learnt can also be directly passed to and be used as input for all layers behind.In tradition Convolutional neural networks in, if you have L layers, just have L connection, but in DenseNet, have L (L+1)/2 A connection is embodied as following formula (1):
Xl=Hl([X0X1..., Xl-1]) (1)
L indicates the number of plies, X in formulalIndicate l layers of output, [X0X1...Xl-1] indicate to connect 0 Dao l-1 layers output characteristic pattern It connects;Hl() indicates the combination of the convolution of Batch Normalization, ReLU and 3*3.
It is well known that network model is deeper to a certain extent, the effect of acquirement is better, however the network the deep is often more difficult to With training;Because convolutional network, during training, the parametric variations of preceding layer the variation of back layer, and this Influence can constantly amplify with the increase of network depth.It is most all using under batch gradient when convolutional network is trained Drop method, then with input data it is continuous variation and network in parameter constantly adjust, each layer input data of network point Cloth then can constantly change, then each layer just needs continuous change to adapt to this new data distribution during training, To cause network training difficult and the problem of be difficult to be fitted;For this problem, the present invention introduces in the training process Normalization layers of Batch;
The each layer of input of Batch Normalization algorithms in each iteration is all normalized, will be defeated Entering the distributions of data, to be normalized to mean value be 0, and the distribution that variance is 1 is specific such as formula (2):
Wherein, xkIndicate the kth dimension of input data, E [xk] indicate the average value that k is tieed up,Indicate standard deviation;
Batch Normalization algorithms are provided with two the variable γ and β that can learn, specifically such as formula (3),
γ and β indicates digital output value for restoring the data distribution that last layer should be acquired, y.
In order to enhance the ability to express of network, deep learning introduces continuous nonlinear activation function Activation Function, ReLU include continuous nonlinear activation function Activation Function and Rectifier functions;Network In the activation primitive that generally uses have Sigmod functions and Rectifier functions, it is specific to calculate as shown in formula (4):
Rectifier (x)=max (0, x) (4).
Since activation primitive ReLU is generally acknowledged to the explanation on biology, and ReLU has been proved to than Sigmod letter Several fitting effects is more preferable;Therefore, the activation primitive selection in model uses ReLU.
According to (1) formula, we need to be attached operation to multiple output characteristic patterns, and be attached operation must It is the in the same size of characteristic pattern to want condition;Down-sampling layer is essential in convolutional network, its effect is to pass through change The size of characteristic pattern carries out dimensionality reduction;Therefore, for the ease of can be carried out in our architecture down-sampling and smoothly it is complete At attended operation, we split the network into intensive piece of DenseBlock of multiple intensive connections, in each DenseBlock The size of characteristic pattern is identical.
Include 4 layers layers of DenseBlock modules in step 2, layer layers by Batch Normalization, ReLU, 3*3 convolution sum Dropout are constituted, and Dropout refers to the training process in deep learning network In, for neural network unit, it is temporarily abandoned from network according to certain probability, so that each mini-batch In the different network of training, wherein Dropout=0.2 can effectively prevent over-fitting using Dropout layers, it is accurate to improve experiment True rate.
Transition Down operations are used for reducing the Spatial Dimension of characteristic pattern, and such conversion is by Batch Normalization, RELU, 1*1 convolution sum 2*2 pondizations operation composition;Wherein use the convolution of 1*1 for preserving characteristic pattern Quantity operates the resolution ratio for reducing characteristic pattern using the pondization of 2*2.As the increase of the number of plies causes feature quantity linearly to increase It is long, however, pondization operation can effectively reduce the resolution ratio of characteristic pattern, therefore be operated by pondization to reduce spatial resolution, Increasing for as caused by number of plies increase characteristic pattern quantity is compensated with this.
The effect of Transition Up operations is to restore the spatial resolution of input picture, and such conversion is turned by one Convolution composition is set, transposition convolution only uses the characteristic pattern of the last one DenseBlock, because of the last one DenseBlock Combine the information of all DenseBlock before.
Last layer in down-sampling path is referred to as Bottleneck;Bottlenek is really by 15 layers layers of structures At DenseBlock and greatly reduce calculation amount its advantage is that gradient disappearance can be alleviated.
103 network models of FC-DenseNet are played characteristic pattern by deep linking using Filter Concatenation Come.
Another embodiment provided by the present invention is:
The specific acquisition method of dataset acquisition process is that unmanned plane is allowed to hover near pavement crack, is then led to Cross the posture of the holder adjustment area array cameras on unmanned plane so that the camera lens of camera is parallel to the surface of pavement crack, and wants The surface substantially 30cm for seeking the distance of camera lens pavement crack of camera, adjust camera posture and distance after, right unmanned plane from Floating state is converted into along the smooth flight of pavement crack direction, is continuously taken pictures, and is taken pictures as shown in Figure 1.
The image, semantic segmentation annotation tool that the present embodiment uses is the image, semantic segmentation annotation tool of pixel scale LabelMe;Specific mask method is that the crack in image is labeled as green i.e. RGB color for (0,255,0), in image All chaff interferents and background except crack are all set to black, i.e. RGB color is (0,0,0);Crack image and right Image after the visualization of crack image tag is as shown in Fig. 2, wherein the first behavior artwork, after the second behavior crack label visualization Image.
Database amplification is carried out to data set by data enhancing, it is original 4096 times to make database amplification.Part is expanded Data set after increasing is illustrated in fig. 3 shown below, wherein the first behavior artwork, the second behavior carries out artwork the image after random cropping, Third behavior carries out image after cutting the image after flip horizontal, and fourth line is after carrying out flip vertical to the image after cutting Image.
The data set handled well input FC-Densenet103 models are trained, use the convolution kernel of 3*3 to defeated first Enter data to be handled.Since artwork is coloured image, there are tri- channels RGB, therefore the m of input picture is 3, by 3*3's After convolution algorithm, m becomes 48.
Image input after process of convolution is included to 4 layers layers of DenseBlock modules, each DenseBlock It is the iteration cascade of preceding features figure, therefore m is the characteristic pattern number and newly-increased feature figure after the convolution of upper layer after this module The sum of number, i.e. 48+4*16=112.
Carry out Transition Down;The feature quantity linear increase after DenseBlock carries out Transition Down can effectively reduce the resolution ratio of characteristic pattern, and the rapid increasing of the characteristic pattern quantity as caused by number of plies increase is compensated with this It is long, it can effectively avoid information explosion;Transition Down only reduce characteristic pattern resolution ratio, do not change quantity, therefore m is still It is 112.
Only change the layers numbers of plies of DenseBlock, it is 5 layers, 7 layers, 10 layers, 12 layers that the number of plies, which is set gradually, repeats 4 Secondary image input DenseBlock modules and progress Transition Down steps;Under same calculating principle, m can be obtained successively It is 192,304,464,656.
Input Bottleneck;Bottleneck is really a DenseBlock being made of 15 layers of layers;Therefore, The feature cascade of characteristic pattern, i.e. m=896+15*16=896 before characteristic pattern number remains as.
Upper layer output result is inputted into the up-sampling channel being made of Transition Up and DenseBlock; It is 12 layers that DenseBlock, which corresponds to the layers numbers of plies in down-sampling,.A upper DenseBlock is up-sampled by transposition convolution to generate Characteristic pattern, then with during down-sampling equal resolution characteristic pattern carry out parallel link, for making up pond process The minutia of middle loss;In order to avoid characteristic pattern explodes, the input of DenseBlock at this time is not attached to its output; Therefore, m at this time is made of three parts altogether, characteristic pattern quantity from TU, the equal resolution from down-sampling path The characteristic pattern quantity generated in characteristic pattern quantity and new DenseBlock;That is m=15*16+656+12*16=1088.
The layers numbers of plies of DenseBlock are followed successively by 10,7,5,4;Repeat by upper layer output result input by The up-sampling channel of Transition Up and DenseBlock compositions;It counts counted m and is followed successively by 816,578,384,256.
1*1 convolution operations are carried out to image, m is reduced to classification number, since we are used for the extraction of Bridge Crack, are only divided For crack and non-crack, therefore, m=2 at this time.
It recently enters softmax layers to be judged, exports the probability in crack and non-crack;If probability of cracks is more than non-crack Probability then judges the pixel for crack;If probability of cracks is less than non-probability of cracks, judge the pixel for non-crack.
Using trained model, test set is tested.Obtained partial test result is illustrated in fig. 9 shown below;Wherein First and third row is original image, and second, four rows are test results.
The above content is a further detailed description of the present invention in conjunction with specific preferred embodiments, and it cannot be said that The specific implementation of the present invention is confined to these explanations.For those of ordinary skill in the art to which the present invention belongs, exist Under the premise of not departing from present inventive concept, a number of simple deductions or replacements can also be made, all shall be regarded as belonging to the present invention's Protection domain.

Claims (10)

1. a kind of bridge pavement Crack Detection and dividing method based on semantic segmentation, it is characterised in that:Include the following steps:
Step 1:Dataset acquisition, by continuously taking pictures along pavement crack direction, obtaining crack image pattern, and artificial Corresponding label is made, to crack image progress semantic segmentation mark, crack is labeled as a kind of single in fracture image Color, all chaff interferents and background in the image of crack except crack are all set to another unified solid color;It is right Data set is expanded, and is training set and test set by data set random assortment after amplification;
Step 2:By crack image in training set, input FC-Densenet103 network models are trained by several times, specific method It is as follows:
Step 1:Crack image in training set is carried out to the convolution of a 3*3;
Step 2:And convolution results are inputted to the DenseBlock modules for including 4 layers layers;
Step 3:Step 2 result is subjected to Transition Down operations, reduces crack image resolution ratio;
Step 4:It is 5 layers, 7 layers, 10 layers, 12 layers that the DenseBlock modules layers layer numbers, which are set gradually, is weighed successively Multiple 4 steps 2 and step 3;
Step 5:The result of the step 4 is inputted to the Bottleneck being made of 15 layers, completes whole down-samplings, and Carry out the attended operation of multiple features;
Step 6:Upper layer output result is inputted into the up-sampling channel being made of Transition Up and DenseBlock, It is 12 layers that DenseBlock, which corresponds to the layers numbers of plies in down-sampling,;
Step 7:The layers numbers of plies of DenseBlock in the step 6 are set as 10,7,5,4 successively, are repeated 4 times step 6;
Step 8:1*1 convolution operations are carried out to the output result of the step 7;
Step 9:Step 8 result input is judged for softmax layers, the probability in crack and non-crack is exported;
Step 3:After the completion of the step 2 training, by trained FC-Densenet103 network models in test set Crack image is tested, and test result is obtained.
2. a kind of bridge pavement Crack Detection and dividing method based on semantic segmentation according to claim 1, feature It is:103 network models of FC-DenseNet include by DenseBlock and Transition Down groups in the step 2 At down-sampling path, and by the up-sampling path formed DenseBlock and Transition Up and softmax functions.
3. a kind of bridge pavement Crack Detection and dividing method based on semantic segmentation according to claim 2, feature It is:103 network models of the FC-DenseNet are made of 103 convolutional layers altogether:First convolution directly acts on input Image has 38 convolutional layers in the down-sampling path being made of DenseBlock, there is 15 convolutional layers in bottleneck Bottleneck, There are 38 convolutional layers in the up-sampling path being made of DenseBlock, is also wrapped in 103 network models of the FC-DenseNet Include a convolution and 5 Transition Up containing 5 Transition Down, each Transition Down, often A Transition Up include the 1*1 convolution of last layer in a transposition convolution and network.
4. a kind of bridge pavement Crack Detection and dividing method based on semantic segmentation according to claim 1, feature It is:After the step 5 completes down-sampling, operation is attached to multiple output features, is embodied as such as formula (1):
Xl=Hl([X0X1..., Xl-1]) (1)
L indicates the number of plies, X in formulalIndicate l layers of output, [X0X1...Xl-1] indicate to connect 0 Dao l-1 layers output characteristic pattern, Hl () indicates the combination of the convolution of Batch Normalization, ReLU and 3*3.
5. a kind of bridge pavement Crack Detection and dividing method based on semantic segmentation according to claim 1, feature It is:Include 4 layers layers of DenseBlock modules in the step 2, layer layers by Batch Normalization, ReLU, 3*3 convolution sum Dropout are constituted, and the Dropout refers to training in deep learning network It is temporarily abandoned, so that each mini-batch neural network unit by Cheng Zhong according to certain probability from network Networks all different in training, wherein Dropout=0.2.
6. a kind of bridge pavement Crack Detection and dividing method based on semantic segmentation according to claim 5, feature It is:The Batch Normalization specific algorithms are as follows:
The each layer of input of Batch Normalization algorithms in each iteration is all normalized, and number will be inputted According to distribution to be normalized to mean value be 0, the distribution that variance is 1 is specific such as formula (2):
Wherein, xkIndicate the kth dimension of input data, E [xk] indicate the average value that k is tieed up,Indicate standard deviation;
Batch Normalization algorithms are provided with two the variable γ and β that can learn, specifically such as formula (3),
γ and β is for restoring the data distribution that last layer should be acquired.
7. a kind of bridge pavement Crack Detection and dividing method based on semantic segmentation according to claim 5, feature It is:The ReLU includes continuous nonlinear activation function Activation Function and Rectifier, specific to calculate As shown in formula (4):
Rectifier (x)=max (0, x) (4)
8. a kind of bridge pavement Crack Detection and dividing method based on semantic segmentation according to claim 2, feature It is:The Transition Down are used to reduce the Spatial Dimension of characteristic pattern, by Batch Normalization, RELU, 1*1 convolution is formed with the operation of 2*2 pondizations, and wherein the convolution of 1*1 is used to preserve the quantity of characteristic pattern, and the pondization operation of 2*2 is used for The resolution ratio of characteristic pattern is reduced, the Transition Up are made of a transposition convolution, the space for restoring input picture Resolution ratio, the transposition convolution only use the characteristic pattern of the last one DenseBlock.
9. a kind of bridge pavement Crack Detection and dividing method based on semantic segmentation according to claim 1, feature It is:The Bottleneck is by 15 layers layers of DenseBlock constituted.
10. a kind of bridge pavement Crack Detection and dividing method based on semantic segmentation according to claim 3, feature It is:103 network models of the FC-DenseNet are played characteristic pattern by deep linking using Filter Concatenation Come.
CN201810309151.3A 2018-04-09 2018-04-09 A kind of bridge pavement Crack Detection and dividing method based on semantic segmentation Pending CN108520516A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810309151.3A CN108520516A (en) 2018-04-09 2018-04-09 A kind of bridge pavement Crack Detection and dividing method based on semantic segmentation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810309151.3A CN108520516A (en) 2018-04-09 2018-04-09 A kind of bridge pavement Crack Detection and dividing method based on semantic segmentation

Publications (1)

Publication Number Publication Date
CN108520516A true CN108520516A (en) 2018-09-11

Family

ID=63430737

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810309151.3A Pending CN108520516A (en) 2018-04-09 2018-04-09 A kind of bridge pavement Crack Detection and dividing method based on semantic segmentation

Country Status (1)

Country Link
CN (1) CN108520516A (en)

Cited By (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109493320A (en) * 2018-10-11 2019-03-19 苏州中科天启遥感科技有限公司 Method for extracting remote sensing image road and system, storage medium, electronic equipment based on deep learning
CN109669474A (en) * 2018-12-21 2019-04-23 国网安徽省电力有限公司淮南供电公司 The adaptive hovering position optimization algorithm of multi-rotor unmanned aerial vehicle based on priori knowledge
CN109784386A (en) * 2018-12-29 2019-05-21 天津大学 A method of it is detected with semantic segmentation helpers
CN109872318A (en) * 2019-02-22 2019-06-11 中国石油大学(华东) A kind of geology for deep learning is appeared crack data set production method
CN109919942A (en) * 2019-04-04 2019-06-21 哈尔滨工业大学 Bridge Crack intellectualized detection method based on high-precision noise reduction theory
CN110070008A (en) * 2019-04-04 2019-07-30 中设设计集团股份有限公司 Bridge disease identification method adopting unmanned aerial vehicle image
CN110147714A (en) * 2019-03-28 2019-08-20 中国矿业大学 Coal mine gob crack identification method and detection system based on unmanned plane
CN110197477A (en) * 2019-05-07 2019-09-03 北京邮电大学 The method, apparatus and system of pavement crack detection
CN110349122A (en) * 2019-06-10 2019-10-18 长安大学 A kind of pavement crack recognition methods based on depth convolution fused neural network
CN111626092A (en) * 2020-03-26 2020-09-04 陕西陕北矿业韩家湾煤炭有限公司 Unmanned aerial vehicle image ground crack identification and extraction method based on machine learning
CN111832437A (en) * 2020-06-24 2020-10-27 万翼科技有限公司 Building drawing identification method, electronic equipment and related product
CN112989981A (en) * 2021-03-05 2021-06-18 五邑大学 Pavement crack detection method, system and storage medium
CN113092502A (en) * 2021-04-14 2021-07-09 成都理工大学 Unmanned aerial vehicle pavement damage detection method and system
CN113222904A (en) * 2021-04-21 2021-08-06 重庆邮电大学 Concrete pavement crack detection method for improving PoolNet network structure
CN113409267A (en) * 2021-06-17 2021-09-17 西安热工研究院有限公司 Pavement crack detection and segmentation method based on deep learning
CN113506281A (en) * 2021-07-23 2021-10-15 西北工业大学 Bridge crack detection method based on deep learning framework
CN113610778A (en) * 2021-07-20 2021-11-05 武汉工程大学 Bridge surface crack detection method and system based on semantic segmentation
CN114359272A (en) * 2022-03-11 2022-04-15 科大天工智能装备技术(天津)有限公司 DenseNet-based bridge steel cable breakage detection method and system
CN117152163A (en) * 2023-11-01 2023-12-01 安徽乾劲企业管理有限公司 Bridge construction quality visual detection method
CN117612021A (en) * 2023-10-19 2024-02-27 广州大学 Remote sensing extraction method and system for agricultural plastic greenhouse

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106910186A (en) * 2017-01-13 2017-06-30 陕西师范大学 A kind of Bridge Crack detection localization method based on CNN deep learnings
CN107133960A (en) * 2017-04-21 2017-09-05 武汉大学 Image crack dividing method based on depth convolutional neural networks

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106910186A (en) * 2017-01-13 2017-06-30 陕西师范大学 A kind of Bridge Crack detection localization method based on CNN deep learnings
CN107133960A (en) * 2017-04-21 2017-09-05 武汉大学 Image crack dividing method based on depth convolutional neural networks

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
SIMON JEGOU等: ""The One Hundred Layers Tiramisu:Fully Convolutional DenseNets for Semantic Segmentation"", 《2017 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION WORKSHOPS》 *

Cited By (30)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109493320B (en) * 2018-10-11 2022-06-17 苏州中科天启遥感科技有限公司 Remote sensing image road extraction method and system based on deep learning, storage medium and electronic equipment
CN109493320A (en) * 2018-10-11 2019-03-19 苏州中科天启遥感科技有限公司 Method for extracting remote sensing image road and system, storage medium, electronic equipment based on deep learning
CN109669474A (en) * 2018-12-21 2019-04-23 国网安徽省电力有限公司淮南供电公司 The adaptive hovering position optimization algorithm of multi-rotor unmanned aerial vehicle based on priori knowledge
CN109669474B (en) * 2018-12-21 2022-02-15 国网安徽省电力有限公司淮南供电公司 Priori knowledge-based multi-rotor unmanned aerial vehicle self-adaptive hovering position optimization algorithm
CN109784386B (en) * 2018-12-29 2020-03-17 天津大学 Method for assisting object detection by semantic segmentation
CN109784386A (en) * 2018-12-29 2019-05-21 天津大学 A method of it is detected with semantic segmentation helpers
CN109872318A (en) * 2019-02-22 2019-06-11 中国石油大学(华东) A kind of geology for deep learning is appeared crack data set production method
CN110147714A (en) * 2019-03-28 2019-08-20 中国矿业大学 Coal mine gob crack identification method and detection system based on unmanned plane
CN110147714B (en) * 2019-03-28 2023-06-23 煤炭科学研究总院 Unmanned aerial vehicle-based coal mine goaf crack identification method and detection system
CN110070008A (en) * 2019-04-04 2019-07-30 中设设计集团股份有限公司 Bridge disease identification method adopting unmanned aerial vehicle image
CN109919942A (en) * 2019-04-04 2019-06-21 哈尔滨工业大学 Bridge Crack intellectualized detection method based on high-precision noise reduction theory
CN110197477A (en) * 2019-05-07 2019-09-03 北京邮电大学 The method, apparatus and system of pavement crack detection
CN110349122A (en) * 2019-06-10 2019-10-18 长安大学 A kind of pavement crack recognition methods based on depth convolution fused neural network
CN111626092B (en) * 2020-03-26 2023-04-07 陕西陕北矿业韩家湾煤炭有限公司 Unmanned aerial vehicle image ground crack identification and extraction method based on machine learning
CN111626092A (en) * 2020-03-26 2020-09-04 陕西陕北矿业韩家湾煤炭有限公司 Unmanned aerial vehicle image ground crack identification and extraction method based on machine learning
CN111832437B (en) * 2020-06-24 2024-03-01 万翼科技有限公司 Building drawing identification method, electronic equipment and related products
CN111832437A (en) * 2020-06-24 2020-10-27 万翼科技有限公司 Building drawing identification method, electronic equipment and related product
CN112989981A (en) * 2021-03-05 2021-06-18 五邑大学 Pavement crack detection method, system and storage medium
CN112989981B (en) * 2021-03-05 2023-10-17 五邑大学 Pavement crack detection method, system and storage medium
CN113092502A (en) * 2021-04-14 2021-07-09 成都理工大学 Unmanned aerial vehicle pavement damage detection method and system
CN113222904A (en) * 2021-04-21 2021-08-06 重庆邮电大学 Concrete pavement crack detection method for improving PoolNet network structure
CN113409267A (en) * 2021-06-17 2021-09-17 西安热工研究院有限公司 Pavement crack detection and segmentation method based on deep learning
CN113610778A (en) * 2021-07-20 2021-11-05 武汉工程大学 Bridge surface crack detection method and system based on semantic segmentation
CN113610778B (en) * 2021-07-20 2024-03-26 武汉工程大学 Bridge surface crack detection method and system based on semantic segmentation
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
CN114359272A (en) * 2022-03-11 2022-04-15 科大天工智能装备技术(天津)有限公司 DenseNet-based bridge steel cable breakage detection method and system
CN117612021A (en) * 2023-10-19 2024-02-27 广州大学 Remote sensing extraction method and system for agricultural plastic greenhouse
CN117152163A (en) * 2023-11-01 2023-12-01 安徽乾劲企业管理有限公司 Bridge construction quality visual detection method
CN117152163B (en) * 2023-11-01 2024-02-27 安徽乾劲企业管理有限公司 Bridge construction quality visual detection method

Similar Documents

Publication Publication Date Title
CN108520516A (en) A kind of bridge pavement Crack Detection and dividing method based on semantic segmentation
CN109977812B (en) Vehicle-mounted video target detection method based on deep learning
CN110363122B (en) Cross-domain target detection method based on multi-layer feature alignment
CN107134144B (en) A kind of vehicle checking method for traffic monitoring
CN108446617B (en) Side face interference resistant rapid human face detection method
CN109902806A (en) Method is determined based on the noise image object boundary frame of convolutional neural networks
CN107145889A (en) Target identification method based on double CNN networks with RoI ponds
CN102332092B (en) Flame detection method based on video analysis
CN109800736A (en) A kind of method for extracting roads based on remote sensing image and deep learning
CN103886308B (en) A kind of pedestrian detection method of use converging channels feature and soft cascade grader
CN104050471B (en) Natural scene character detection method and system
CN104599275B (en) The RGB-D scene understanding methods of imparametrization based on probability graph model
CN105512638B (en) A kind of Face datection and alignment schemes based on fusion feature
CN110348376A (en) A kind of pedestrian's real-time detection method neural network based
CN108427920A (en) A kind of land and sea border defense object detection method based on deep learning
CN111898406B (en) Face detection method based on focus loss and multitask cascade
CN108875608A (en) A kind of automobile traffic signal recognition method based on deep learning
CN109685115A (en) A kind of the fine granularity conceptual model and learning method of bilinearity Fusion Features
CN108460403A (en) The object detection method and system of multi-scale feature fusion in a kind of image
CN108229338A (en) A kind of video behavior recognition methods based on depth convolution feature
CN111881730A (en) Wearing detection method for on-site safety helmet of thermal power plant
CN111695514B (en) Vehicle detection method in foggy days based on deep learning
CN107463892A (en) Pedestrian detection method in a kind of image of combination contextual information and multi-stage characteristics
CN106651872A (en) Prewitt operator-based pavement crack recognition method and system
CN113052210A (en) Fast low-illumination target detection method based on convolutional neural network

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

Application publication date: 20180911

RJ01 Rejection of invention patent application after publication