CN108416307B - Method, device and equipment for detecting pavement cracks of aerial images - Google Patents
Method, device and equipment for detecting pavement cracks of aerial images Download PDFInfo
- Publication number
- CN108416307B CN108416307B CN201810205751.5A CN201810205751A CN108416307B CN 108416307 B CN108416307 B CN 108416307B CN 201810205751 A CN201810205751 A CN 201810205751A CN 108416307 B CN108416307 B CN 108416307B
- Authority
- CN
- China
- Prior art keywords
- pavement
- dimensional
- road surface
- sample
- target
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/182—Network patterns, e.g. roads or rivers
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the invention provides a method, a device and equipment for detecting a pavement crack of an aerial image. The method comprises the following steps: extracting deep high-dimensional features of a road surface area of the aerial photographing road surface image, and obtaining a high-dimensional feature map according to the deep high-dimensional features; based on deep high-dimensional characteristics of the pavement area, screening positive and negative samples of the high-dimensional characteristic diagram to distinguish a pavement crack target from a pavement background; and classifying and positioning the pavement crack target in a coordinate manner to obtain classification information and coordinate information of the pavement crack target. The method and the device can be applied to pavement crack detection of aerial images of high-altitude motion backgrounds and complex scenes, are more suitable for acquiring images by an unmanned aerial vehicle-mounted system compared with various commonly used crack detection algorithms, and have better crack robustness of aerial image detection.
Description
Technical Field
The embodiment of the invention relates to the field of deep learning and pattern recognition, in particular to a method, a device and equipment for detecting an aerial image pavement crack.
Background
At present, one of the main damage forms of the highway pavement is pavement cracks, wherein the highway cracks in China mainly comprise transverse cracks and longitudinal cracks. If the crack can be found at the early stage of crack occurrence and the development condition of the crack can be tracked, the maintenance cost of the pavement is greatly reduced, and the driving safety of the highway is ensured. Therefore, it is very important to regularly investigate and maintain the road surface condition of a road.
The pavement crack detection mode is developed from the initial manual detection mode; with the application of the image processing technology, the vehicle-mounted acquisition device is combined with the image processing technology and applied to pavement crack detection, so that the detection efficiency is greatly improved. In recent years, unmanned aerial vehicle technique obtains rapid development, and the application that combines with it obtains very big abundantly, combines in the road surface crack detection device of unmanned aerial vehicle collection mode, compares in other methods, has fast high efficiency, the visual field is big and the advantage that the storage data volume descends to some extent. But compared with vehicle-mounted collected images, the method has the interference of roadside scenes, vehicles, electric wires, shadows and the like, and the noise is very rich.
The commonly used crack identification method mainly focuses on the applications of threshold segmentation, feature detection, texture analysis, seed growth and the like, and in addition, the method also applies machine learning and fuzzy sets. However, the existing methods are basically developed by detecting on the basis of images of a vehicle-mounted acquisition device, and cannot be applied to aerial images with more interference and noise.
Disclosure of Invention
Aiming at the problems in the prior art, the embodiment of the invention provides a method, a device and equipment for detecting the pavement crack of an aerial image, which combine a series of advantages of an aerial image acquisition mode to enable the crack detection to be more efficient and convenient, and improve the robustness problem of the traditional image processing algorithm.
In a first aspect, an embodiment of the present invention provides an aerial image pavement crack detection method, including:
extracting deep high-dimensional features of a road surface area of the aerial photographing road surface image, and obtaining a high-dimensional feature map according to the deep high-dimensional features;
based on deep high-dimensional characteristics of the pavement area, screening positive and negative samples of the high-dimensional characteristic diagram to distinguish a pavement crack target from a pavement background;
and classifying and positioning the pavement crack target in a coordinate manner to obtain classification information and coordinate information of the pavement crack target.
In a second aspect, an embodiment of the present invention provides an aerial image pavement crack detection apparatus, including:
the high-dimensional characteristic map module is used for extracting deep high-dimensional characteristics of a road surface area of the aerial road surface image and obtaining a high-dimensional characteristic map according to the deep high-dimensional characteristics;
the crack identification module is used for screening positive and negative samples of the high-dimensional characteristic graph based on deep high-dimensional characteristics of the pavement area so as to distinguish a pavement crack target from a pavement background; and
and the classification positioning module is used for classifying and positioning the pavement crack target in a coordinate manner to obtain the classification information and the coordinate information of the pavement crack target.
In a third aspect, an embodiment of the present invention provides an aerial image pavement crack detection apparatus, including:
at least one processor; and
at least one memory communicatively coupled to the processor, wherein:
the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the method for detecting the road surface crack of the aerial image according to the first aspect of the embodiment of the invention and the method according to any optional embodiment thereof.
In a fourth aspect, a non-transitory computer-readable storage medium is provided, which stores computer instructions for executing the method for detecting a road surface crack in an aerial image according to the first aspect of the embodiments of the present invention and the method according to any optional embodiment of the method.
According to the method for detecting the pavement crack of the aerial photography image, provided by the embodiment of the invention, the pavement crack target and the pavement background are preliminarily distinguished by extracting the high-dimensional characteristic map of the pavement area, and the crack target is classified and accurately positioned, so that the crack target of the aerial photography pavement is positioned and classified, and the coordinate information of the crack target is obtained. The method and the device can be applied to pavement crack detection of aerial images of high-altitude motion backgrounds and complex scenes, are more suitable for acquiring images by an unmanned aerial vehicle-mounted system compared with various commonly used crack detection algorithms, and have better crack robustness of aerial image detection.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and those skilled in the art can also obtain other drawings according to the drawings without creative efforts.
FIG. 1 is a flowchart of a method for detecting a pavement crack of an aerial image according to an embodiment of the invention;
FIG. 2 is a schematic diagram of an embodiment of identifying a crack in a road surface by an aerial image according to an embodiment of the invention;
fig. 3 is a schematic frame diagram of an aerial image pavement crack detection device according to an embodiment of the invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some embodiments, but not all embodiments, of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Fig. 1 is a flowchart of an aerial image pavement crack detection method according to an embodiment of the present invention, and the aerial image pavement crack detection method shown in fig. 1 includes:
s100, extracting deep high-dimensional features of a road surface area of the aerial image, and obtaining a high-dimensional feature map according to the deep high-dimensional features;
the embodiment of the invention is suitable for crack detection of an aerial image of a road surface, and the aerial image has the characteristics of high altitude, long distance, shooting work and complex scene, namely the background outside the road surface is complex.
Preferably, step S100 further includes: and roughly dividing the aerial image of the road surface to eliminate the invalid area beside the road surface, and obtaining the road surface area in the aerial image of the road surface.
Specifically, the aerial image of the road surface comprises a road surface area and a roadside invalid area outside the road surface area, the roadside invalid area outside the road surface area in the aerial image of the road surface is removed, and the rest is the road surface area; on the basis, high-dimensional feature extraction is carried out on the pavement area, and a high-dimensional feature map is obtained according to the extracted deep high-dimensional features. The high-dimensional features are: and the overall semantic information image characteristics of the crack target are different from the local information characteristics of the shallow edge.
S200, screening positive and negative samples of the high-dimensional characteristic graph based on deep high-dimensional characteristics of the pavement area to distinguish a pavement crack target from a pavement background;
the embodiment of the invention screens positive and negative samples, namely, preliminarily identifies the pavement crack target, and distinguishes the pavement crack target from a pavement background, wherein the positive sample is the crack target, and the negative sample is the pavement background.
S300, classifying and coordinate positioning the pavement crack target to obtain classification information and coordinate information of the pavement crack target.
After the preliminary positioning is performed in step S200 to obtain the crack target and the road surface background, step S300 classifies based on the preliminarily positioned crack target, accurately positions the position of the crack target, and finally obtains detailed classification information and coordinate position information of the crack target. The classification information refers to type classification of fracture targets, such as transverse fractures or longitudinal fractures; and the coordinate positioning is the accurate position coordinate of the quantitative crack target.
According to the method for detecting the pavement crack of the aerial photography image, provided by the embodiment of the invention, the pavement crack target and the pavement background are preliminarily distinguished by extracting the high-dimensional characteristic map of the pavement area, and the crack target is classified and accurately positioned, so that the crack target of the aerial photography pavement is positioned and classified, and the coordinate information of the crack target is obtained. The method and the device can be applied to pavement crack detection of aerial images of high-altitude motion backgrounds and complex scenes, are more suitable for acquiring images by an unmanned aerial vehicle-mounted system compared with various commonly used crack detection algorithms, and have better crack robustness of aerial image detection.
Based on the above embodiment, the classifying and coordinate positioning the pavement crack target to obtain the classification information and the coordinate information of the pavement crack target, and then further includes:
and S400, calculating the length of the pavement crack target according to the classification information and the coordinate information of the pavement crack target.
Based on the detailed classification information and coordinate position information of the crack target, step S400 further calculates the length of the pavement crack target, so that the following information can be finally obtained according to the aerial pavement image in the embodiment of the present invention: locating the fracture target, detailed classification of the fracture target, accurate coordinate location and quantitative length data.
The embodiment of the invention overcomes the image processing difficulty caused by an unmanned aerial vehicle acquisition mode, can be applied to the detection of high-altitude motion backgrounds and complex-scene pavement cracks, has stronger robustness compared with various commonly-used crack detection algorithms, and obtains better crack identification effect of aerial images. The method is applied to the aerial crack detection, can provide more prominent images of crack targets for observers, can perform quantitative analysis on the crack length, and provides reference basis for follow-up road maintenance.
In an optional embodiment, in step S100, the extracting deep high-dimensional features of the road surface region of the aerial image, and obtaining a high-dimensional feature map according to the deep high-dimensional features specifically include:
s100.1, constructing a feature extraction network by using a convolutional neural network, and adding a road rough segmentation layer based on a K-means clustering algorithm in the feature extraction network;
s100.2, screening and eliminating the roadside invalid area of the aerial road surface image by utilizing the road rough segmentation layer to obtain the road surface area of the aerial road surface image;
and S100.3, combining the low-latitude features of the road surface area into high-latitude features by using the feature extraction network to obtain a high-dimensional feature map.
In step S100.1, the embodiment of the present invention performs learning training based on a convolutional neural network, and constructs a feature extraction network by analyzing a mature network structure and combining with a complex scene small target detection task condition, where the feature extraction network is used to extract high-dimensional features of a road surface area. Furthermore, the embodiment of the invention combines a K-means algorithm, and adds a road rough segmentation layer based on the K-means clustering algorithm in the feature extraction network, wherein the road rough segmentation layer meets the characteristic of high recall rate, and can remove the invalid area beside the road.
Step S100.2, screening and eliminating the roadside invalid area of the aerial road surface image by utilizing the road rough segmentation layer to obtain the road surface area of the aerial road surface image, and specifically processing as follows:
in order to ensure that all road surface areas can be extracted to high-dimensional features by the feature extraction network, the input aerial photographing road surface image is roughly segmented by using a road rough segmentation layer based on a K-means clustering algorithm, and invalid areas beside the road are removed.
For a given sample set, calculating the distance between each clustering center and each data element by using initial K clustering centers according to the distance between samples, distributing K data elements nearest to each clustering center in each iteration to form K clusters, then recalculating the distributed clustering centers, and distributing all data elements in an iteration manner until all clusters are not changed, dividing the sample set into fixed K clusters, and enabling points in the clusters to be connected together as closely as possible, so that the distance between the clusters is as large as possible.
Assuming the clustering number K and the maximum iteration number N, the input sample set data is as follows:
D＝{x1,x2…xm} (1)
wherein x isiTo input a sample, i is 0, 2, …, m.
Randomly selected k samples from dataset D as centroid vector are:
μ＝{μ1,μ2…μk} (2)
for each iteration process of the N number of iterations, the cluster partitioning is initialized to:
and calculates a calculation sample x for each input samplei(i-1, 2 … m) and respective centroid vectors μj(j ═ 1,2 … k), expressed as:
will have the minimum distance d among all samplesijSample x ofiDivision into corresponding classes λiIn (2), the sample cluster update rule is:
the update rule of the sample particles is as follows:
repeating the steps (4), (5) and (6) until N iterations are completed or the iteration process cluster is not updated, and obtaining the last K segmentation clusters as follows:
C＝{C1,C2…Ck} (7)
in general, the minimum squared error E of the target is updated to the minimum, which is expressed as:
where β is a distance type of the cost function, a manhattan distance is used when β is 1, and a euclidean distance is used when β is 2.
According to the cluster updating rule, after the image regions with similar features are divided into the same cluster after passing through the K-means rough dividing layer, the road surface regions generally have similar feature distribution, the roadside region features are distributed in a discrete manner, and the image rough dividing is carried out according to the cluster with K being 2, so that the features similar to the road surface regions and the features dissimilar to the road surface regions can be divided into two categories, and the road surface region dividing effect with high recall rate is obtained.
S100.3, combining the low-latitude features of the road surface area into high-latitude features by using the feature extraction network to obtain a high-dimensional feature map, and specifically processing the high-dimensional feature map as follows:
and performing high-dimensional feature extraction on the road surface area by using a feature extraction network constructed by the convolutional neural network. The method comprises the steps of forming neurons according to levels, training and updating weights and bias in the neurons according to a back propagation algorithm, combining low-dimensional features into high-dimensional features by training according to the local and overall relation of input data, and obtaining the spatial correlation among different features of different dimensions. Local connection and weight sharing are realized by sharing convolution kernel parameters of the same level, prior knowledge is introduced into the convolution neural network, the difficulty of network training is greatly reduced, and the convolution neural network has good applicability to image data.
In the hierarchy of convolutional neural networks, there are mainly four basic hierarchies: a convolution layer; a pooling layer; a fully-connected layer; and activating the layer.
The convolutional layer is a filter that can train learning parameters. There are two main types of classification of convolution kernels by filling: the convolution kernel with unfilled edge and the convolution kernel with the edge filled with 0 pixel points according to half of the size of the convolution kernel have the advantage of preventing the image from being seriously reduced in size after the multi-layer convolution operation and ensuring that the feature diagram is not influenced by the convolution operation. The sizes of convolution kernels of the same level are consistent, one convolution kernel is responsible for extracting one image shape feature, and the sizes of the convolution kernels between different levels are not limited in size.
Assuming that a two-dimensional feature image F is operated by a two-dimensional convolution H, the expression is:
wherein, FlayerIs a two-dimensional feature map of layer, Flayer+1Is a (layer +1) layer two-dimensional characteristic diagram, i and j are image coordinate points corresponding to the convolution center, and m and n are the length and width dimensions of the two-dimensional convolution respectively.
After the feature map passes through the convolution layer, the size of the next layer of feature map is as follows:
heightlayer+1＝(heightlayer-m+2*padding)/stride+1 (10)
widthlayer+1＝(widthlayer-n+2*padding)/stride+1 (11)
wherein, heightlayerIs height of two-dimensional feature map of layerlayer+1Is the height, width of the (layer +1) layer two-dimensional characteristic diagramlayerIs the width of the layer two-dimensional characteristic diagramlayer+1The width of the (layer +1) layer two-dimensional feature map, padding the edge filling size of the feature map, and stride the convolution calculation step size.
The pooling layer is used as a transition layer between adjacent convolution layers, can effectively compress the number of data and network parameters, and is helpful for preventing the network from generating an overfitting phenomenon. The pooling operation can be divided into two categories: and performing maximum pooling and average pooling, wherein the former is to divide the feature map into corresponding areas according to the size of the pooling blocks, the data of the maximum feature in each area is selected as the parameter of the pooled feature map, and the latter is to perform average processing on the features in each corresponding area and select the average value as the parameter of the pooled feature map. Maximum pooling is applied to extract features of image objects and average pooling is applied to average image background features.
The fully-connected layer is one of the original forms of the neural network, high-dimensional features output by the previous network can be connected into a slender feature vector through the fully-connected layer, and the slender feature vector is simultaneously mapped into a linearly separable space, so that the output layer can realize classification of feature targets, and the output size of the output layer is the type of data classification.
The activation layer is mainly used for adding a nonlinear factor to a linear model, removing redundancy in data, and reserving and mapping the features extracted by the convolutional layer so as to ensure that the model is better fitted. To avoid gradient disappearance and neuron misfiring and make the loss curve converge faster, an ELU activation function is used, whose expression is:
and x is parameter information input into the active layer, alpha is a preset parameter of the active layer, and meanwhile, the Dropout and Batchnormalization methods are adopted to avoid the over-fitting risk and reduce the fitting condition of the deep network to data.
Dropout inactivates a plurality of proportional neurons at random in the current layer, forward propagation calculates loss of a predicted value and a label value, backward propagation updates parameters of the network, and the process is repeated continuously. The inactivation is random during each iteration, so that the parameters are iteratively updated by different network structures, and the purpose of avoiding overfitting can be achieved on one hand; the complex co-adaptation relation among the neurons is reduced, the weight updating is not influenced by the neurons with the internal connection, the model is screened to show more generalized characteristics, and the network is forced to learn more robust characteristics.
Batch Normalization is a method of normalizing data distribution. During deep network training, because the data distribution of the input layer is not fixedly distributed, the distribution of the current layer also changes randomly, so that a training model needs to be adapted to new data distribution continuously, the learning rate needs to be set very small, and the requirement on parameter initialization is very high. The Batchnormalization is to solve the problem, before activating the function, the data is distributed to a fixed interval through normalization, and the model training does not need to adapt to different distributions each time, so that the parameters are better updated by gradients.
In the training stage, calculating the mini-batch sample mean value as follows:
wherein m is the number of the mini-batch samples.
Calculating the standard deviation of the mini-batch sample according to the sample mean as follows:
wherein, for the preset parameter, X is each sample in the data set, and the sample distribution after normalization is:
the sample distribution for the alternative batch sample normalization is:
wherein gamma is a learning parameter of the sample mean,
learning parameters are sample standard deviations.
The gradient updating algorithm adopts a self-adaptive learning rate method RMSprop with faster convergence, and updates the gradient of the current position and the learning rate by using the gradient of the previous moment, wherein the gradient updating formula is as follows:
wherein, E [ g2]t-1Mean of the squares of the gradients at the preceding time, E [ g ]2]tIs the average of the squares of the gradients at the current time,
is the square of the gradient of the current time position, η is the learning rate, θtFor the weighting parameter of the previous moment, θt+1Is a current time weighting parameter.
In addition, aiming at the convolution of the small-size characteristic diagram, the small convolution kernel combination is used for replacing a large convolution kernel, and the n multiplied by n convolution kernel combination is replaced by the n multiplied by 1 and 1 multiplied by n convolution kernel combination at the back end of the network, so that the number of network parameters can be effectively reduced, and the network convergence speed is accelerated.
Based on the above embodiment, in step S200, the screening of positive and negative samples is performed on the high-dimensional feature map based on the deep high-dimensional features of the road surface region to distinguish the road surface crack target from the road surface background, which specifically includes:
s200.1, traversing the high-dimensional feature map by using an anchor sliding window based on the deep high-dimensional features of the road surface region to obtain a candidate sample frame with a preset area scale and a preset aspect ratio, wherein the candidate sample frame is a candidate sample region;
wherein, the preset area scale can be 3 area scales; preferably, the predetermined area scale is 1282、2562、5122. The preset aspect ratio may be a 3-medium aspect ratio; preferably, the preset aspect ratio is 1:1, 1:3, 3: 1.
Specifically, the positive and negative samples are divided according to the following criteria:
dividing 2 candidate sample regions with the IOU of the candidate sample frame and any calibration sample frame larger than a first preset threshold value and the IOU of the candidate sample frame and the IOU of the remaining calibration sample frames largest into positive samples;
and dividing the candidate sample frame and the candidate sample frame of which the IOU of the calibration sample frame except the positive sample is smaller than a second preset threshold value into negative samples.
Wherein the content of the first and second substances,
IOU ═ (candidate sample frame ═ calibration sample frame)/(candidate sample frame ═ calibration sample frame).
Preferably, the first preset threshold is 0.7; the second preset threshold is 0.3.
S200.2, training the candidate sample frame by utilizing a classification loss function, a positioning loss function and a multitask loss function of the regional nomination network of the feature extraction network to screen positive and negative samples to obtain a pavement crack target sample and a pavement background sample, wherein the positive sample is the pavement crack target sample, and the negative sample is the pavement background sample.
Specifically, the classification loss function of the area nomination network is as follows:
lcls(p)＝-(1-pu)γlogpu；
wherein p isuThe probability that the calibration sample frame is a positive sample and a negative sample is shown, and gamma is a Focal local training parameter;
the positioning loss function of the area nomination network is as follows:
wherein v isiIs the coordinate information of the candidate sample block,
predicting regression correction parameters of sample frames for foreground and background, u being the u-th prediction sample frame, tiIn order to calibrate the coordinate information of the sample frame,
the loss function is:
the multitask loss function of the regional nomination network is as follows:
wherein n isclsFor all sample numbers, nregFor positive sample number, μ is normalized to 0.2, piTo predict the probability of being a fracture target,
are discrete tag values.
Based on the above features, the specific processing of step S200.1 is:
firstly, traversing the whole feature map by using an anchor sliding window with a fixed size according to a high-dimensional feature map extracted from the last layer of convolution layer of the feature extraction network, wherein the center of each sliding window generates different local receptive fields according to different levels, and the expression is as follows:
sizel-1＝stride×(sizel-1)+sizeconv-2×padding (19)
wherein, sizel-1Is a size of the upper layer of receptive fieldlFor feature output size, stride is the convolution move step size and padding is the convolution fill size.
Each sliding window center corresponds to 9 candidate sample frames, and the candidate sample regions are selected by using the anchor sample frames with the 3 aspect ratios and the 3 areas in the receptive field region corresponding to the sliding window center. The preferable area scale and aspect ratio in 3 are selected in combination with the slender characteristic of the crack target, and 3 area scales are 1282、2562、5122And the aspect ratio of 3 types is 1:1, 1:3 and 3: 1.
Then mapping the obtained candidate region coordinates into a 1024-dimensional feature vector corresponding to the number of output channels of the convolutional layer, and respectively calculating two loss functions: classification loss functions and localization loss functions. Positive and negative samples which can be used for training in a candidate sample frame generated by an anchor mechanism need to be selected firstly, an IOU is utilized to calculate the overlapping rate of the candidate sample frame and an adjacent training data calibration frame, and a sample used for training is selected, wherein the expression is as follows:
IOU ═ to (candidate sample frame andcalibrationsample frame)/(candidate sample frame ═ calibration sample frame) (20)
The specific basis for positive and negative sample division is as follows: dividing the 2 candidate sample frames with the largest IOU of the candidate sample frame and any calibration sample frame >0.7 and the candidate sample frames and the rest calibration sample frames into training positive samples; and dividing the candidate sample frame and the candidate sample frame of the calibration sample frame except the positive sample, wherein the IOU of the candidate sample frame and the IOU of the calibration sample frame are less than 0.3, into training negative samples.
Based on the above features, the specific processing of step S200.2 is:
the method comprises the steps of distinguishing two types of samples, namely a crack target and a background area, by using a classification loss function of a regional nomination network, and adopting calibration sample frame actual samples corresponding to the divided positive and negative samples, wherein the sample of the calibration target is defined as 1, and the background area is defined as 0. Each calibration sample box corresponds to a discrete probability distribution, which can be expressed as:
p＝(p0,p1) (21)
wherein p is0Is the probability that it is a background region, p1Is the probability that it is the target region of the fracture.
Secondly, the proportion of positive and negative samples divided by the anchor mechanism cannot guarantee balance, so that the problem of class imbalance is caused, most negative samples are areas which are simple and easy to judge as negative samples, and the loss function training is influenced due to the large number of the simple samples, so that the loss function cannot converge to a good result. Therefore, by combining the Focal local idea, the classification Loss function of the area nomination network has the expression:
lcls(p)＝-(1-pu)γlogpu(22)
wherein p isuTo calibrate the probability that the sample frame is a positive or negative sample, γ is the Focal local training parameter and the initial value is 5.
Finally, the positions of candidate sample frames divided by the receptive field are corrected by utilizing the positioning loss in combination with training positive and negative samples, each training positive and negative sample corresponds to a group of coordinate frame size coordinate vectors and is calculated by utilizing the divided positive and negative sample coordinates and sample coordinate mapping calibrated by a data set, and the mapping relation between the frames is as follows:
wherein, f (P)x,Py,Pw,Ph) For the divided positive and negative sample frame coordinate information,
sample frame coordinate information, G (G), is corrected for the framex,Gy,Gw,Gh) In order to calibrate the coordinate information of the sample frame, each coordinate information comprises 4 pieces of quantitative coordinate information, namely a horizontal and vertical coordinate of the upper left corner and the width and height of the sample frame.
Referring to a frame regression correction mode in fast-RCNN, a series of translations and scale scaling are utilized to convert the frame regression correction mode into two groups of 4 feature vectors respectively.
The regression correction parameters of the calibration sample frame are as follows:
the regression correction parameters for the predicted sample box are:
the loss function of the bounding box correction uses L which is more robust to discrete points1A loss function, expressed as:
wherein the content of the first and second substances,
the loss function expression is:
the multitask loss function of the area nomination network is
Wherein n isclsFor all sample numbers, nregFor positive sample number, μ is normalized to 0.2, piTo predict the probability of being a fracture target,
is a discrete mark value, and the expression is:
the method comprises the steps of screening positive and negative sample regions required by training through a multitask loss function of a region nomination network, preliminarily classifying a crack target and a road background, and combining a calibration sample frame to realize preliminary regression positioning of the crack target sample frame.
Based on the above embodiment, in step S300, the classifying and coordinate positioning the pavement crack target to obtain the classification information and the coordinate information of the pavement crack target specifically includes:
s300.1, regulating positive and negative samples screened by the regional nomination network to a feature map with uniform size by utilizing the ROI pooling layer of the feature extraction network, and performing classified output to obtain classified information comprising transverse cracks, longitudinal cracks and a road background;
specifically, when a sample frame which is generated by training classification and frame regression by a regional nomination network is combined with a convolutional neural network, an ROI pooling layer of a spatial pyramid pooling structure is utilized to averagely divide a feature map into grids with the size of n multiplied by n, maximum pooling is adopted for the feature map part in a small grid, and only the maximum extraction is reserved.
The output feature map of the last layer of the convolution layer of the feature extraction network and the original image frame coordinates corresponding to the candidate sample frame output by the regional nomination network pass through the ROI pooling layer, the feature sizes are unified and are connected to the full connection layer, wherein the output size of the full connection layer is the number of categories of the classification network plus the background classification, namely 3 kinds of output classification network parameters are included, including transverse cracks, longitudinal cracks and pavement backgrounds. The classification in this step is finer than the classification in step S200 of the preliminary classification of the crack target and the road surface background.
And S300.2, classifying the concrete crack types of the positive sample by using a classification loss function, performing frame regression, and correcting the coordinate information of a crack target frame.
Specifically, a classification loss function is used to classify specific crack classes of the positive sample candidate sample frames, each candidate sample frame corresponds to a discrete probability distribution, and the expression is as follows:
P＝(P0,P1,P3) (36)
wherein, P0Is the probability that it is a background region, P1Is the probability that it is a transverse crack region, P2Is the probability that it is a longitudinal fracture zone.
In the classification of the crack target, the number of transverse crack samples and longitudinal crack samples is not balanced, so a classification Loss function is designed by using Focal local, and the expression is as follows:
Lcls(P)＝-(1-PU)γlogPU(37)
wherein, PUThe probability that the candidate sample box is a positive or negative sample is set.
And (4) performing coordinate positioning on the classified positive sample candidate sample frame by using a loss function of the frame regression network, and adopting the regression correction parameter T ═ T (T) of the corresponding calibration sample frame in the S2.1x,Ty,Tw,Th) And (V) a regression correction parameter of the prediction sample framex,Vy,Vw,Vh) And correcting, wherein the expression is as follows:
for this fracture classification network, the multi-tasking loss function expression that combines the classification loss function and the localization loss function is:
wherein N isclsFor all sample numbers, NregIs the number of positive samples, PiTo predict the probability of being a fracture target,
is a discrete mark value, and the expression is:
classifying the crack targets preliminarily identified by the regional nomination network through a multitask loss function, specifically dividing the types of the crack targets, completing secondary regression positioning on the crack target sample frame by combining a calibration sample frame, and correcting the coordinate information of the crack target frame.
Based on the above embodiment, in step S400, calculating the length of the pavement crack target according to the classification information and the coordinate information of the pavement crack target specifically includes:
s400.1, according to the classification information and the coordinate information of the pavement crack target, utilizing morphology to hit the single-pixel framework of the hitless transformation extraction crack target, and calculating the pixel length of the crack target; the fracture target comprises a transverse fracture and a longitudinal fracture;
s400.2, converting the pixel length of the crack target according to the pixel coordinate of the aerial image of the road surface and the road length of the actual road surface to obtain the length of the crack target.
Specifically, in step S400.1, fracture segments fractured in the fracture detection process are connected through morphological expansion operation, and the fracture is completely expanded by refining the expanded fracture by using a morphological processing hit-miss conversion method, so that redundant edge pixel information is removed under the condition that the geometric scale of the fracture is not changed, and meanwhile, the problem of multi-pixel width generated by the fracture detection target is removed, and a fracture skeleton composed of single-pixel width can accurately reflect the detailed characteristics of the fracture.
Aiming at a crack detection target, respectively using structural elements in multiple directions for iterative processing, wherein the expression is as follows:
where f is the input structural image, s is a number of suitable structural elements, and c is the number of iterations.
Assume that the sequence of structural elements s is defined as:
{s}＝{s1,s2,·····sn} (42)
then there is:
and (3) performing iterative operation on all result elements by using the methods of the formulas (3.21) to (3.23), and repeating the operation on each structural element in sequence if the result is not converged until the result is not changed.
Secondly, discrete points are fitted to be a skeleton curve by using a Huber weight function of a least square fitting method, and the distance sum of each discrete point to the fitted skeleton curve is minimum according to a standard least square principle. But the robustness is not good for outliers far away from the skeleton curve, and weight threshold processing needs to be set.
The Huber weight function is:
where τ represents a distance threshold, which is the distance of the adjacent curves.
And selecting a proper threshold value to constrain the image crack recognition effect and the distance between adjacent curves, including noise interference and crack targets.
When the point-to-curve distance is less than or equal to a threshold τ, a weight is given as 1, and when the point-to-curve distance is greater than the threshold τ, the weight function is equal to the inverse of the distance multiplied by the threshold τ, with values decreasing with increasing distance. Preferably, τ is 3.
And finally, in the step S400.2, the actual length of the crack target can be obtained by combining the pixel length of the crack target obtained by calculation in the step S400.1 with parameters such as the flying height and the flying speed of the unmanned aerial vehicle and converting the proportion of the image pixel to the actual length.
The embodiment of the invention can be combined with display equipment to display relevant information, including road length, crack number, crack type, information of crack length of each section and the like.
In summary, in the method for detecting the pavement cracks of the deep-learning aerial images, which is provided by the embodiment of the invention, the invalid regions beside the road are removed by using the K-means road rough segmentation layer with high recall rate; combining the complex scene small target detection task condition, providing a feature extraction network structure, and extracting high-dimensional features of a road surface area; providing a region nomination structure of a feature extraction network meeting the requirement of the detection of the pavement crack of the aerial image to generate a candidate sample frame, and primarily screening positive and negative samples of a crack target and a pavement background; the classification and positioning structure is used for detecting and classifying the crack target and regressing the secondary frame, the length data of the crack target is calculated by extracting a single-pixel skeleton of the crack from the detection result area, and the method has good crack robustness of aerial image detection.
Fig. 2 is a schematic diagram of an embodiment of identifying a pavement crack in an aerial image according to the present invention, please refer to fig. 2, where the embodiment of the present invention samples an aerial image as an input, removes an invalid area beside a road in the aerial image by using a K-means road rough segmentation layer with a high recall rate according to step S100, and designs a feature extraction network structure in combination with a complex scene small target detection task condition to extract a high-dimensional feature of the pavement area; according to the step S200, providing a region nomination structure of a feature extraction network meeting the requirement of the detection of the pavement crack of the aerial image, generating a candidate sample frame for the feature map of the last layer of the network in the step 2, and primarily screening positive and negative samples of a crack target and a pavement background; according to the step S300, classifying and secondary frame regression are carried out on the crack targets by utilizing the classification positioning structure; according to the step S400, extracting a crack single-pixel framework from the detection result area by using the hit-miss transformation, and further calculating the crack length according to the image coordinate and the actual pavement size ratio; and finally, outputting the crack image and related information including the road length, the crack number, the crack type and the crack length of each section for display by display equipment.
For the aerial image, the embodiment of the invention can overcome the image processing difficulty caused by an unmanned aerial vehicle acquisition mode, can be applied to the detection of high-altitude motion background and complex-scene pavement cracks, has better robustness compared with various commonly used crack detection algorithms, and obtains better aerial image crack identification effect. The method is applied to the aerial crack detection, can automatically detect, classify and position crack targets, can quantitatively analyze the length of the crack, and provides a reference basis for subsequent road maintenance.
The embodiment of the invention carries out operation processing on the road surface image based on deep learning, mainly comprising three steps, wherein one step is to extract a deep high-dimensional characteristic map of a road surface area, the other step is to generate effective positive and negative samples for screening and training and preliminarily distinguish a road surface crack target and a road surface background, and the third step is to classify the crack target, position specific coordinates and further calculate the skeleton length of the crack target. In each processing step, different algorithm networks are used for processing and analyzing image data, different super parameters are set for carrying out primary distinguishing, detailed classification and coordinate positioning on a pavement crack target and a pavement background, and accurate classification positioning and quantitative skeleton calculation on the pavement crack are achieved automatically.
The algorithm selected by the embodiment of the invention is particularly suitable for identifying the pavement cracks of the aerial images with large data volume and much interference noise, overcomes the defects of the prior art and has good beneficial effects.
The embodiment of the invention also provides a device for detecting the pavement crack of the aerial image, which comprises:
the high-dimensional characteristic map module is used for extracting deep high-dimensional characteristics of a road surface area of the aerial road surface image and obtaining a high-dimensional characteristic map according to the deep high-dimensional characteristics;
the crack identification module is used for screening positive and negative samples of the high-dimensional characteristic graph based on deep high-dimensional characteristics of the pavement area so as to distinguish a pavement crack target from a pavement background; and
and the classification positioning module is used for classifying and positioning the pavement crack target in a coordinate manner to obtain the classification information and the coordinate information of the pavement crack target.
The device of the embodiment of the invention can be used for executing the technical scheme of the embodiment of the method for detecting the pavement crack of the aerial image shown in fig. 1, the implementation principle and the technical effect are similar, and the detailed description is omitted here.
Fig. 3 is a schematic frame diagram of an aerial image pavement crack detection device according to an embodiment of the invention. Referring to fig. 3, an embodiment of the invention provides an aerial image pavement crack detection apparatus, including: a processor (processor)310, a communication Interface (communication Interface)320, a memory (memory)330 and a bus 340, wherein the processor 310, the communication Interface 320 and the memory 330 complete communication with each other through the bus 340. The processor 310 may call logic instructions in the memory 330 to perform methods comprising: extracting deep high-dimensional features of a road surface area of the aerial photographing road surface image, and obtaining a high-dimensional feature map according to the deep high-dimensional features; based on deep high-dimensional characteristics of the pavement area, screening positive and negative samples of the high-dimensional characteristic diagram to distinguish a pavement crack target from a pavement background; and classifying and positioning the pavement crack target in a coordinate manner to obtain classification information and coordinate information of the pavement crack target.
An embodiment of the present invention discloses a computer program product, which includes a computer program stored on a non-transitory computer readable storage medium, the computer program including program instructions, when the program instructions are executed by a computer, the computer can execute the methods provided by the above method embodiments, for example, the method includes: extracting deep high-dimensional features of a road surface area of the aerial photographing road surface image, and obtaining a high-dimensional feature map according to the deep high-dimensional features; based on deep high-dimensional characteristics of the pavement area, screening positive and negative samples of the high-dimensional characteristic diagram to distinguish a pavement crack target from a pavement background; and classifying and positioning the pavement crack target in a coordinate manner to obtain classification information and coordinate information of the pavement crack target.
Embodiments of the present invention provide a non-transitory computer-readable storage medium, which stores computer instructions, where the computer instructions cause the computer to perform the methods provided by the above method embodiments, for example, the methods include: extracting deep high-dimensional features of a road surface area of the aerial photographing road surface image, and obtaining a high-dimensional feature map according to the deep high-dimensional features; based on deep high-dimensional characteristics of the pavement area, screening positive and negative samples of the high-dimensional characteristic diagram to distinguish a pavement crack target from a pavement background; and classifying and positioning the pavement crack target in a coordinate manner to obtain classification information and coordinate information of the pavement crack target.
Those of ordinary skill in the art will understand that: the implementation of the above-described apparatus embodiments or method embodiments is merely illustrative, wherein the processor and the memory may or may not be physically separate components, i.e. may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.
Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and certainly can also be implemented by hardware. With this understanding in mind, the above technical solutions may be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as a usb disk, a removable hard disk, a ROM/RAM, a magnetic disk, an optical disk, etc., and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute the method according to the embodiments or some parts of the embodiments.
Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.
Claims (7)
1. An aerial image pavement crack detection method is characterized by comprising the following steps:
extracting deep high-dimensional features of a road surface area of the aerial photographing road surface image, and obtaining a high-dimensional feature map according to the deep high-dimensional features;
based on deep high-dimensional characteristics of the pavement area, screening positive and negative samples of the high-dimensional characteristic diagram to distinguish a pavement crack target from a pavement background;
classifying and coordinate positioning the pavement crack target to obtain classification information and coordinate information of the pavement crack target;
the method comprises the following steps of extracting deep high-dimensional features of a road surface region of an aerial image, and obtaining a high-dimensional feature map according to the deep high-dimensional features, wherein the method specifically comprises the following steps:
constructing a feature extraction network by using a convolutional neural network, and adding a road rough segmentation layer based on a K-means clustering algorithm in the feature extraction network;
screening and eliminating roadside invalid areas of the aerial road surface image by using the road rough segmentation layer to obtain a road surface area of the aerial road surface image;
combining the low-dimensional features of the road surface area into high-dimensional features by using the feature extraction network to obtain a high-dimensional feature map;
the method comprises the following steps of screening positive and negative samples of the high-dimensional characteristic diagram based on deep high-dimensional characteristics of the pavement area to distinguish a pavement crack target from a pavement background, and specifically comprises the following steps:
traversing the high-dimensional feature map by using an anchor sliding window based on the deep high-dimensional features of the road surface area to obtain a candidate sample frame with a preset area scale and a preset aspect ratio;
training the candidate sample frame by using a classification loss function, a positioning loss function and a multitask loss function of a regional nomination network of the feature extraction network to screen positive and negative samples to obtain a pavement crack target sample and a pavement background sample, wherein the positive sample is the pavement crack target sample, and the negative sample is the pavement background sample;
the classifying and coordinate positioning of the pavement crack target to obtain the classification information and the coordinate information of the pavement crack target specifically comprises the following steps:
regulating positive and negative samples screened by the regional nomination network to a feature map with uniform size by utilizing the ROI pooling layer of the feature extraction network, and performing classified output to obtain classified information including transverse cracks, longitudinal cracks and a road background;
and classifying the concrete crack types of the positive sample by using a classification loss function, performing frame regression, and correcting the coordinate information of a crack target frame.
2. The method of claim 1, wherein the classifying and coordinate locating the pavement crack target, obtaining classification information and coordinate information of the pavement crack target, and then further comprising:
and calculating the length of the pavement crack target according to the classification information and the coordinate information of the pavement crack target.
3. The method of claim 1, wherein the positive and negative samples are divided according to:
dividing 2 candidate sample regions with the IOU of the candidate sample frame and any calibration sample frame larger than a first preset threshold value and the IOU of the candidate sample frame and the IOU of the remaining calibration sample frames largest into positive samples;
dividing the candidate sample frame and the candidate sample frame of which the IOU of the calibration sample frame except the positive sample is smaller than a second preset threshold value into negative samples;
wherein the content of the first and second substances,
IOU ═ (candidate sample frame ═ calibration sample frame)/(candidate sample frame ═ calibration sample frame).
4. The method according to claim 2, wherein the calculating the length of the pavement crack target according to the classification information and the coordinate information of the pavement crack target specifically comprises:
according to the classification information and the coordinate information of the pavement crack target, a single-pixel framework of the crack target is extracted by morphological hit-miss transformation, and the pixel length of the crack target is calculated; the fracture target comprises a transverse fracture and a longitudinal fracture;
and converting the pixel length of the crack target according to the pixel coordinates of the aerial image of the road surface and the road length of the actual road surface to obtain the length of the crack target.
5. An aerial image pavement crack detection device, characterized by, includes:
the high-dimensional characteristic map module is used for extracting deep high-dimensional characteristics of a road surface area of the aerial road surface image and obtaining a high-dimensional characteristic map according to the deep high-dimensional characteristics;
the crack identification module is used for screening positive and negative samples of the high-dimensional characteristic graph based on deep high-dimensional characteristics of the pavement area so as to distinguish a pavement crack target from a pavement background; and
the classification positioning module is used for classifying and positioning the pavement crack target in a coordinate mode to obtain classification information and coordinate information of the pavement crack target;
wherein the high-dimensional feature map module is further configured to,
constructing a feature extraction network by using a convolutional neural network, and adding a road rough segmentation layer based on a K-means clustering algorithm in the feature extraction network;
screening and eliminating roadside invalid areas of the aerial road surface image by using the road rough segmentation layer to obtain a road surface area of the aerial road surface image;
combining the low-dimensional features of the road surface area into high-dimensional features by using the feature extraction network to obtain a high-dimensional feature map;
the classification positioning module is also used for positioning the classification positioning module,
traversing the high-dimensional feature map by using an anchor sliding window based on the deep high-dimensional features of the road surface area to obtain a candidate sample frame with a preset area scale and a preset aspect ratio;
training the candidate sample frame by using a classification loss function, a positioning loss function and a multitask loss function of a regional nomination network of the feature extraction network to screen positive and negative samples to obtain a pavement crack target sample and a pavement background sample, wherein the positive sample is the pavement crack target sample, and the negative sample is the pavement background sample;
the classification positioning module is also used for positioning the classification positioning module,
regulating positive and negative samples screened by the regional nomination network to a feature map with uniform size by utilizing the ROI pooling layer of the feature extraction network, and performing classified output to obtain classified information including transverse cracks, longitudinal cracks and a road background;
and classifying the concrete crack types of the positive sample by using a classification loss function, performing frame regression, and correcting the coordinate information of a crack target frame.
6. An aerial image pavement crack detection device, comprising:
at least one processor; and
at least one memory communicatively coupled to the processor, wherein:
the memory stores program instructions executable by the processor, the processor invoking the program instructions to perform the method of any of claims 1 to 4.
7. A non-transitory computer-readable storage medium storing computer instructions that cause a computer to perform the method of any one of claims 1 to 4.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810205751.5A CN108416307B (en) | 2018-03-13 | 2018-03-13 | Method, device and equipment for detecting pavement cracks of aerial images |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810205751.5A CN108416307B (en) | 2018-03-13 | 2018-03-13 | Method, device and equipment for detecting pavement cracks of aerial images |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108416307A CN108416307A (en) | 2018-08-17 |
| CN108416307B true CN108416307B (en) | 2020-08-14 |
Family
ID=63131115
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810205751.5A Active CN108416307B (en) | 2018-03-13 | 2018-03-13 | Method, device and equipment for detecting pavement cracks of aerial images |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108416307B (en) |
Families Citing this family (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109165658B (en) * | 2018-08-28 | 2021-08-13 | 哈尔滨工业大学（威海） | Strong negative sample underwater target detection method based on fast-RCNN |
| CN109447047A (en) * | 2018-12-27 | 2019-03-08 | 山东大学 | The pavement detection system and detection method of UAV system |
| CN109814551A (en) * | 2019-01-04 | 2019-05-28 | 丰疆智慧农业股份有限公司 | Cereal handles automated driving system, automatic Pilot method and automatic identifying method |
| CN109902738B (en) * | 2019-02-25 | 2021-07-20 | 深圳市商汤科技有限公司 | Network module, distribution method and device, electronic equipment and storage medium |
| CN109934352B (en) * | 2019-03-06 | 2021-08-31 | 北京深度奇点科技有限公司 | Automatic evolution method of intelligent model |
| CN110059751A (en) * | 2019-04-19 | 2019-07-26 | 南京链和科技有限公司 | A kind of tire code and tire condition recognition methods based on machine learning |
| CN110197477A (en) * | 2019-05-07 | 2019-09-03 | 北京邮电大学 | The method, apparatus and system of pavement crack detection |
| CN110120041A (en) * | 2019-05-14 | 2019-08-13 | 郑州大学 | Pavement crack image detecting method |
| CN110222727A (en) * | 2019-05-15 | 2019-09-10 | 广东电网有限责任公司电力调度控制中心 | A kind of short-term load forecasting method and device based on deep neural network |
| CN110197483A (en) * | 2019-05-31 | 2019-09-03 | 南京深地智能建造技术研究院有限公司 | Deep basal pit crack detection method based on vision signal |
| CN110334630A (en) * | 2019-06-27 | 2019-10-15 | 和远智能科技股份有限公司 | A method of equipment rhegma automatic identification is realized based on depth learning technology |
| CN110598751B (en) * | 2019-08-14 | 2022-06-07 | 安徽师范大学 | Anchor point generation method based on geometric attributes |
| CN111091552A (en) * | 2019-12-12 | 2020-05-01 | 哈尔滨市科佳通用机电股份有限公司 | Method for identifying closing fault image of angle cock handle of railway wagon |
| CN111310558B (en) * | 2019-12-28 | 2023-11-21 | 北京工业大学 | Intelligent pavement disease extraction method based on deep learning and image processing method |
| CN111612747B (en) * | 2020-04-30 | 2023-10-20 | 湖北煌朝智能自动化装备有限公司 | Rapid detection method and detection system for product surface cracks |
| CN111797687A (en) * | 2020-06-02 | 2020-10-20 | 上海市城市建设设计研究总院（集团）有限公司 | Road damage condition extraction method based on unmanned aerial vehicle aerial photography |
| CN111709931B (en) * | 2020-06-16 | 2023-05-23 | 四川赛康智能科技股份有限公司 | Automatic acquisition method and system for strain clamp defect detection and identification report |
| CN111723750B (en) * | 2020-06-23 | 2023-05-30 | 安阳工学院 | Road crack detection method and device for automatic driving vehicle |
| CN112132097B (en) * | 2020-09-30 | 2024-06-14 | 珠海阔朗科技有限公司 | Intelligent pavement crack identification system and method |
| CN112767322B (en) * | 2021-01-05 | 2023-06-13 | 成都圭目机器人有限公司 | Airport cement pavement FOD risk assessment method and device |
| CN113092502A (en) * | 2021-04-14 | 2021-07-09 | 成都理工大学 | Unmanned aerial vehicle pavement damage detection method and system |
| CN113723181B (en) * | 2021-07-20 | 2023-10-20 | 深圳大学 | Unmanned aerial vehicle aerial photographing target detection method and device |
| CN113838078B (en) * | 2021-09-06 | 2023-06-30 | 中国矿业大学(北京) | Method and device for identifying and extracting coal mining subsidence ground cracks and storage medium |
| CN114202733A (en) * | 2022-02-18 | 2022-03-18 | 青岛海信网络科技股份有限公司 | Video-based traffic fault detection method and device |
| CN115830007B (en) * | 2023-02-06 | 2023-06-09 | 山东省滨州公路工程有限公司 | Pavement void detection method and system based on image recognition |
| CN116596930B (en) * | 2023-07-18 | 2023-09-22 | 吉林大学 | Semi-supervised multitasking real image crack detection system and method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9235902B2 (en) * | 2011-08-04 | 2016-01-12 | University Of Southern California | Image-based crack quantification |
| CN106651872A (en) * | 2016-11-23 | 2017-05-10 | 北京理工大学 | Prewitt operator-based pavement crack recognition method and system |
-
2018
- 2018-03-13 CN CN201810205751.5A patent/CN108416307B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9235902B2 (en) * | 2011-08-04 | 2016-01-12 | University Of Southern California | Image-based crack quantification |
| CN106651872A (en) * | 2016-11-23 | 2017-05-10 | 北京理工大学 | Prewitt operator-based pavement crack recognition method and system |
Non-Patent Citations (1)
| Title |
|---|
| 航拍图像的路面裂缝识别;王博 等;《光学学报》;20170831;第37卷(第8期);1-7 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108416307A (en) | 2018-08-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108416307B (en) | Method, device and equipment for detecting pavement cracks of aerial images | |
| Shen et al. | Detection of stored-grain insects using deep learning | |
| CN103984953B (en) | Semantic segmentation method based on multiple features fusion Yu the street view image of Boosting decision forests | |
| CN111353512B (en) | Obstacle classification method, obstacle classification device, storage medium and computer equipment | |
| CN108876797B (en) | Image segmentation system and method based on Spiking-SOM neural network clustering | |
| CN109447998B (en) | Automatic segmentation method based on PCANet deep learning model | |
| CN109584209B (en) | Vascular wall plaque recognition apparatus, system, method, and storage medium | |
| CN111402227B (en) | Bridge crack detection method | |
| Alidoost et al. | A CNN-based approach for automatic building detection and recognition of roof types using a single aerial image | |
| CN105160355B (en) | A kind of method for detecting change of remote sensing image based on region correlation and vision word | |
| Wang et al. | Combined use of FCN and Harris corner detection for counting wheat ears in field conditions | |
| Shen et al. | Biomimetic vision for zoom object detection based on improved vertical grid number YOLO algorithm | |
| WO2019136591A1 (en) | Salient object detection method and system for weak supervision-based spatio-temporal cascade neural network | |
| JP7054278B1 (en) | Edge identification method based on deep learning | |
| CN114758288A (en) | Power distribution network engineering safety control detection method and device | |
| CN109685045A (en) | A kind of Moving Targets Based on Video Streams tracking and system | |
| CN109886159B (en) | Face detection method under non-limited condition | |
| CN112101189B (en) | SAR image target detection method and test platform based on attention mechanism | |
| CN111126459A (en) | Method and device for identifying fine granularity of vehicle | |
| CN110991444A (en) | Complex scene-oriented license plate recognition method and device | |
| CN105809716A (en) | Superpixel and three-dimensional self-organizing background subtraction algorithm-combined foreground extraction method | |
| CN111738164A (en) | Pedestrian detection method based on deep learning | |
| CN109741358B (en) | Superpixel segmentation method based on adaptive hypergraph learning | |
| CN117437382B (en) | Updating method and system for data center component | |
| CN111339950B (en) | Remote sensing image target detection method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |