CN110335255A - A kind of dam slope crack detection method and bionic gecko based on deep learning is creeped detection device - Google Patents

A kind of dam slope crack detection method and bionic gecko based on deep learning is creeped detection device Download PDF

Info

Publication number
CN110335255A
CN110335255A CN201910514001.0A CN201910514001A CN110335255A CN 110335255 A CN110335255 A CN 110335255A CN 201910514001 A CN201910514001 A CN 201910514001A CN 110335255 A CN110335255 A CN 110335255A
Authority
CN
China
Prior art keywords
crack
dam
bionic
detection
detection device
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.)
Granted
Application number
CN201910514001.0A
Other languages
Chinese (zh)
Other versions
CN110335255B (en
Inventor
邹湘军
雷子毅
唐昀超
黄矿裕
陈明猷
黄钊丰
朱惠贤
张晋豪
徐婉冬
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.)
South China Agricultural University
Zhongkai University of Agriculture and Engineering
Original Assignee
South China Agricultural University
Zhongkai University of Agriculture and Engineering
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 South China Agricultural University, Zhongkai University of Agriculture and Engineering filed Critical South China Agricultural University
Priority to CN201910514001.0A priority Critical patent/CN110335255B/en
Publication of CN110335255A publication Critical patent/CN110335255A/en
Application granted granted Critical
Publication of CN110335255B publication Critical patent/CN110335255B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B62—LAND VEHICLES FOR TRAVELLING OTHERWISE THAN ON RAILS
    • B62D—MOTOR VEHICLES; TRAILERS
    • B62D57/00—Vehicles characterised by having other propulsion or other ground- engaging means than wheels or endless track, alone or in addition to wheels or endless track
    • B62D57/02—Vehicles characterised by having other propulsion or other ground- engaging means than wheels or endless track, alone or in addition to wheels or endless track with ground-engaging propulsion means, e.g. walking members
    • B62D57/024—Vehicles characterised by having other propulsion or other ground- engaging means than wheels or endless track, alone or in addition to wheels or endless track with ground-engaging propulsion means, e.g. walking members specially adapted for moving on inclined or vertical surfaces
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/84—Systems specially adapted for particular applications
    • G01N21/88—Investigating the presence of flaws or contamination
    • G01N21/8851—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0004—Industrial image inspection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/194—Segmentation; Edge detection involving foreground-background segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/60—Analysis of geometric attributes
    • G06T7/62—Analysis of geometric attributes of area, perimeter, diameter or volume
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/80—Analysis of captured images to determine intrinsic or extrinsic camera parameters, i.e. camera calibration
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/84—Systems specially adapted for particular applications
    • G01N21/88—Investigating the presence of flaws or contamination
    • G01N21/8851—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges
    • G01N2021/8887—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges based on image processing techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10004—Still image; Photographic image
    • G06T2207/10012—Stereo images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20228—Disparity calculation for image-based rendering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30108—Industrial image inspection
    • G06T2207/30132—Masonry; Concrete

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Chemical & Material Sciences (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • Immunology (AREA)
  • Transportation (AREA)
  • Biochemistry (AREA)
  • Signal Processing (AREA)
  • Pathology (AREA)
  • Quality & Reliability (AREA)
  • Combustion & Propulsion (AREA)
  • Analytical Chemistry (AREA)
  • Mechanical Engineering (AREA)
  • Geometry (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Length Measuring Devices By Optical Means (AREA)

Abstract

The dam slope crack detection method and bionic gecko that the invention discloses a kind of based on deep learning are creeped detection device.This method trains neural network first, then uses trained u-net neural network, detects to the dam crack image acquired in real time, the crack in accurate segmented image and background sundries.When carrying out Crack Detection to inclination dam, creeped detection device using bionic gecko, which includes fuselage, walking bionic mechanism, video camera clamping device and light-source illuminating system.The present invention can detect inclination dam crack by neural network and binocular vision detection system well, including information such as fracture width, the penetrations of fracture, detection accuracy is high, and enables detection device very stable on inclination dam by walking bionic mechanism and walk and detect.

Description

A kind of dam slope crack detection method and bionic gecko based on deep learning is creeped Detection device
Technical field
The present invention relates to detection field, in particular to a kind of dam slope crack detection method based on deep learning and imitative Raw gecko creeps detection device.
Background technique
Dam is block rivers ditch water flow with raising of water level or the water retaining structure of adjusting flow, if there is crack, meeting Influence the stability on dam inclined-plane and the safety of dam.The detection in dam crack at present, usually professional use crack Sounding instrument and New Instrument for Crack Width periodically go on dam to detect, and then rule of thumb judge the danger etc. in crack by professional Grade.But dam generally has gradient, artificial detection is than relatively hazardous, time-consuming and laborious, it is difficult to be measured in real time, Er Qieren Work detection is easy to appear the phenomenon that failing to judge, judging by accident.Deep learning method has been applied to field of image processing at present, therefore can incite somebody to action Deep learning is used for dam slope Crack Detection, to improve the accuracy of detection.
Summary of the invention
It is an object of the invention to overcome disadvantage existing in the prior art, provide a kind of accuracy height, real-time detection, at This low dam slope crack detection method based on deep learning is to be passed back based on U-net network according to CCD industrial camera The image come, whether real-time detection dam has crack, judges its degree of danger, to judge whether dam needs to repair.
Another object of the present invention is to provide a kind of bionic geckos using above-mentioned dam slope crack detection method to climb Row detection device.
The purpose of the invention is achieved by the following technical solution:
A kind of dam slope crack detection method based on deep learning, includes the following steps:
(1) training neural network: since the image in dam crack includes crack and background sundries, and crack is in the picture Accounting generally there was only 2~6%, therefore can not directly be detected using U-net neural network, need first to improve loss letter Number, then U-net neural network is trained:
The training set and test set of 1-1. preparation neural network: deep learning training needs a large amount of sample, and there are two types of solutions Certainly method: one is obtaining new data, but this method is more troublesome, needs a large amount of cost, and the sample in dam crack Notebook data is fewer；Another kind is enhanced data with existing, i.e., overturns data with existing, translated or rotated, and is created More data out, so that neural network has better extensive effect；The present invention uses second method, first with CCD industry Camera acquires the realtime graphic on dam slope, obtains 1000 or more crack images, and pre-process to the image of acquisition； Then it is used for network training using wherein 60~80% image as training set, in addition 20~40% image is used as test set In network test；
The parameter of 1-2. adjustment neural network: being adjusted by the parameter to U-net neural network, keeps U-net neural Network can adapt to crack image, thus accurate Ground Split crack and background sundries；
1-3. test: the training result of U-net neural network is tested with the test set that 1-1 step obtains, by handing over And it is assessed than MIOU and Dice coefficient；
(2) trained u-net neural network is used, the dam crack image acquired in real time is detected, accurate point Cut the crack in image and background sundries:
2-1. camera calibration: video camera be aligned one have it is many independent can identification point known structure, obtain mark Determine image, the inner parameter and external parameter of video camera is then found out by the built-in function of OpenCV, wherein inner parameter packet Include spin matrix R and translation vector T；
2-2. corrects image: left and right cameras shooting obtains the real-time left images in dam crack；Then basis The spin matrix R and translation vector T that 2-1 step obtains, are corrected real-time left images, the left and right figure after being corrected Picture；
2-3. uses trained u-net neural network, crack and background in the accurate left images divided after correcting Sundries obtains the FRACTURE CHARACTERISTICS information of left images；
2-4. Stereo matching: the FRACTURE CHARACTERISTICS information of left images is matched, and the deviation analyzed between feature regards Difference obtains disparity map；
The depth in 2-5. solution crack: depth map is converted by disparity map, depth depth is as follows:
Wherein, depth indicates the depth information in crack；F indicates normalized focal length；B is between two camera optical centers Distance, also referred to as parallax range；Disp is parallax value；
The width in 2-6. solution crack: width H (A, B) is the maximum width in crack in image, and formula is as follows:
H (A, B)=max [h (A, B), h (B, A)]
Wherein, A is the up contour point in crack, and B is the down contour point in crack；H (A, B) be point set A to point set B it is unidirectional away from From h (B, A) is from point set B to point set A one-way distance.The measurement of distance uses the squared-distance of theorem in Euclid space, solves and splits in image The width information of seam, the practical surface distance being just to solve between the upper edge region in crack and lower edge margin.
In step 1-1, the image of acquisition is pre-processed, is the image because acquiring under field conditions (factors), illumination Parameters variation may be bigger, need to carry out uniform illumination processing to the original image of acquisition, to eliminate different illumination counterincision Stitch the influence of detection.
The algorithm of uniform illumination processing is as follows:
F (x, y)=g (x, y)-h (x, y)+a
Wherein, f (x, y) is the image slices vegetarian refreshments by dodging, and g (x, y) is the image slices without dodging Vegetarian refreshments, h (x, y) are the background pixel point for reflecting Luminance Distribution, and a is grayscale shift amount.The effect of grayscale shift amount is to make even light Treated, and image grayscale is distributed in reasonable value range, so that image brilliance mean value is consistent with original brightness, with Solve the problems, such as that some local brightness are big, some local brightness are small.
It is the training set obtained using step 1-1 in step 1-2, u-net network is trained with TensorFlow, using Habit rate is 0.01~0.001, and iteration 2000~3000 times, loss late is iterated to 0.4 hereinafter, completing the training of network model.
In step 1-2, the parameter of adjustment includes setting loss function and adjustment L2 regularization coefficient.
Setting loss function is only to account for very little specific gravity because of crack in the picture, and the specific gravity of complex background is very big, this In the case of easily lead to the over-fitting of neural network model crack region can be improved by Weighted Loss Function Specific gravity enables neural network to focus more on the study of FRACTURE CHARACTERISTICS.The loss function L is as follows:
Wherein, n is expressed as class number (n value 2 indicates complex background and two, crack classification)；W is weight,Respectively indicate the probability of the i-th class classification and the probability of model prediction.
Weight w in loss function will be updated weight to each of image pixel, enable neural network The marginal information in special study crack.Weight w (x) after update, as follows:
Wherein, b indicates deviation, needs artificially to give its parameter；d1Indicate the pixel to crack image coboundary most Closely；d2Indicate the pixel to crack image lower boundary minimum distance.Based on experience, we can be by initial weight w0If It is 10, standard deviation sigma is about 5 pixels, and b is about 1.We obtain each pixel by precalculating weight in loss function Weight, this method compensates for the different frequency of every class pixel in training data, and so that neural network is more focused on study and split The information being sewn in figure.
During training neural network, setting L2 regularization coefficient is 0.0005, to prevent neural network model excessively quasi- It closes.
In step 1-3, Dice is a judgment criteria of crack image segmentation:
Wherein, | G | and | P | the prospect in true value and prediction result is respectively indicated, it is corresponding for the image of crack It is exactly the crack in the crack and segmentation result in expert's uncalibrated image；| G ∩ P | indicate the two overlapping region；0 indicates without weight It is folded, it is completely dissimilar；1 indicates completely overlapped；Numerical value is higher, and expression network segmentation effect is better.
In step 1-3, MIOU is also a judgment criteria of fracture image segmentation, calculate two intersection of sets collection and The ratio between union, the two set are expert's uncalibrated image and predicted value respectively:
Wherein, K is expressed as class number (K value 2 indicates complex background and two, crack classification)；pijIndicate the i-th prediction For the number of jth type objects；pjiIndicate that jth is predicted as the number of the i-th class article；piiIndicate that i-th is predicted as the i-th class article number Mesh.
In step 2-4, the parallax refers to that left and right cameras is observed to obtain difference of the identical feature on x coordinate, depending on Difference disp=xl-xr, wherein xlIt is distance of a certain characteristic point in crack in left video camera figure to optical axis center, xrIt is that the right side is taken the photograph Distance of the camera same characteristic features point to right camera optical axis center.
A kind of bionic gecko is creeped detection device, including fuselage 2-1, walking bionic mechanism 2-2, video camera clamping device 2- 3 and light-source illuminating system 2-4 is detected using the above-mentioned dam slope crack detection method based on deep learning；Walking Bio-mechanism 2-2 is installed on the lower along surrounding of fuselage, for driving bionic gecko detection device of creeping to walk on inclination dam； Video camera clamping device 2-3 is installed on the L-type section sliding rail 2-1-2 of fuselage, for setting up video camera and being adjustable video camera Height apart from detection faces；Light-source illuminating system 2-4 is installed on the top of fuselage, sufficient stable for providing to monitoring region Light source illumination.
The fuselage 2-1 include hood 2-1-1, four L-type section sliding rail 2-1-2, two pick up the ears to detain 2-1-3.In hood Empty part is detection zone；Four L-type section sliding rails are vertically installed at hood top；Two buttons of picking up the ears are separately mounted in hood Portion two sides, for connecting cable.There are one large square hole and eight small rectangular openings at the top of hood, are video camera clamping device 2- 3 provide the adjustable range from hood top surface lower part to top, while testing staff being facilitated to screw holding screw.
The walking bionic mechanism 2-2 includes four bionic gecko feet, is separately mounted under fuselage along surrounding, each bionical Gecko foot includes gecko foot main body 2-2-2, gecko foot limbs 2-2-3 and gecko foot toe 2-2-1.On gecko foot limbs 2-2-3 It is provided with steering engine, can be turned to.Bionic gecko foot is multifilament shape surface texture, and a row is provided below in each of which toe Above liner is provided with one or more bristle on each liner, and the top of every bristle is provided with one or more young pilose antler Fine and soft small brushes.These small brushes form very big adsorption capacity, and bionic gecko is allowed to creep detection device in inclination dam Upper smoothly detection and walking.When bionic gecko creeps detection device when detecting, bionic gecko foot, which can be firmly attached to, to incline On oblique dam, it will not glide because of gravity and dam gradient.When bionic gecko creeps detection device walking, gecko is shunk Part bristle on foot toe, reduces the tilt angle of bristle, to reduce frictional force of the detection device on dam, then passes through The realization of gecko foot limbs is walked on dam.
The video camera clamping device 2-3 includes cover 2-3-1, sliding block 2-3-2, fixture 2-3-4；The four of cover 2-3-1 There is sliding block 2-3-2 at angle, for moving on the L-type section sliding rail 2-1-2 of fuselage, has scale on the sliding rail of L-type section, for calculating Distance of the camera optical center to detection faces；It is fixed by holding screw, video camera clamping device can be fixed on to distance detection In the certain height of face, also height of the adjustable video camera apart from detection faces；Camera shooting machine clamp is provided with above cover 2-3-1 2-3-4, for clamping monocular, binocular or multi-lens camera 2-3-3, the angle adjustable of camera optical axis and detection faces.
The light-source illuminating system 2-4 include four light sources, each light source towards fuselage axial line inclination 5~100, for for Detection zone provides sufficient uniform illumination.
The bionic gecko is creeped detection device, further includes drawing descending mechanism 1, is creeped detection device for pulling bionic gecko Main body 2, the detection device main body 2 include fuselage 2-1, walking bionic mechanism 2-2, video camera clamping device 2-3, light source illumination System 2-4.Detection device main body 2 is stablized in lateral detection process without gliding in a certain height of inclination dam, and drop is drawn Mechanism 1 can make in the lifting of detection device main body 2 to the arbitrary height of inclination dam.Descending mechanism 1 is drawn to be mounted on the upper of inclination dam Edge, and detection device main body 2 is pulled by cable.
The drawing descending mechanism 1 includes long track 1-1, fixed pulley 1-2, slide 1-3, cable 1-4；Long track 1-1 is mounted on The upper edge of dam is tilted, slide 1-3 is mounted on long track 1-1 and can move on long track；Two fixed pulley 1-2 are mounted on On slide 1-3, the rotational angular velocity of the two is consistent；One end of cable 1-4 be fixed on the wheel trench bottom of fixed pulley 1-2 and by Fixed pulley pulls, the other end bolt of cable 1-4 fuselage 2-1 pick up the ears detain 2-1-3.In use, being pulled by rotation fixed pulley Cable, so that detection device Soil stability is on the arbitrary height of inclination dam surface.Can also be installed on slide host computer, The electromechanics component such as motor, air pump, solenoid valve, to simplify the structure of detection device main body.
The present invention has the following advantages that compared with prior art and effect:
(1) present invention can detect inclination dam crack, packet by neural network and binocular vision detection system well The information such as fracture width, the penetration of fracture are included, detection accuracy is high.
(2) of the invention highly-safe, testing staff can be allowed far from dangerous inclination dam metope, operated in safe place The present apparatus can tilt dam metope to full wafer and carry out crack detection；Cost of labor can be reduced, detection efficiency, structure letter are improved It is single, conveniently guide and support.
(3) detection range of the invention is wide, can switch traveling wheel group by way of lifting, carries out full wafer and tilts dam The detection of metope；And video camera mesh number, away from detection faces height it is adjustable, be adapted to multiple types camera chain.
(4) testing result stability of the invention is high, enables detection device very stable by walking bionic mechanism It walks and detects in the inclination enterprising every trade of dam.
Detailed description of the invention
Fig. 1 is operation schematic diagram of the invention.
Fig. 2 is drawing descending mechanism schematic diagram of the invention.
Fig. 3 is detection device main body main view of the invention.
Fig. 4 is detection device main body top view of the invention.
Fig. 5 is walking bionic structural scheme of mechanism of the invention.
Fig. 6 is fuselage top view of the invention.
Fig. 7 is the structure chart of u-net network of the present invention.
In figure, 1, drawing descending mechanism；1-1, long track；1-2, fixed pulley；1-3, slide；1-4, cable；
2, subject detection device；
2-1, fuselage；2-1-1, hood；2-1-2, L-type section sliding rail；2-1-3, it picks up the ears to detain；
2-2, walking bionic mechanism；2-2-1, gecko foot toe；2-2-2, gecko foot main body；2-2-3, gecko foot limbs；
2-3, video camera clamping device；2-3-1, cover；2-3-2, sliding block；2-3-3, video camera；2-3-4, fixture；
2-4, light-source illuminating system.
Specific embodiment
The present invention will be further described with reference to the accompanying drawing, but embodiments of the present invention are not limited thereto.? In description of the invention, it is to be understood that if having "upper", "lower", "left", "right", "horizontal", "vertical" etc. instruction direction or The term of positional relationship, then they is are based on the orientation or positional relationship shown in the drawings, be merely for convenience of the description present invention or Simplify description, rather than the device or element of indication or suggestion meaning must have a particular orientation, with specific orientation construction And operation, therefore the terms describing the positional relationship in the drawings are only for illustration, should not be understood as the limitation to this patent.
Embodiment
As shown in fig. 7, being trained first to neural network.First prepare the training set and test set of neural network, then With TensorFlow training U-net network, initial learning rate is adjusted to 0.001, is dropped loss late iteration 5000 times or so To 0.4 hereinafter, obtained MIOU and Dice coefficient is respectively 0.89 He again by MIOU and Dice test model accuracy rate 0.91, illustrate that segmentation effect is good.
Using trained u-net neural network, the dam crack image acquired in real time is detected, accurate segmentation figure Crack and background sundries as in, and the depth and width in crack are calculated in real time, feed back to testing staff.
Bionic gecko is creeped detection device, as shown in Fig. 1~Fig. 6, including fuselage 2-1, walking bionic mechanism 2-2, camera shooting Machine clamping device 2-3 and light-source illuminating system 2-4, be using the above-mentioned dam slope crack detection method based on deep learning into Row detection；Walking bionic mechanism 2-2 is installed on the lower along surrounding of fuselage, for driving bionic gecko detection device of creeping tilting It walks on dam；Video camera clamping device 2-3 is installed on the L-type section sliding rail 2-1-2 of fuselage, for setting up video camera and energy Adjust height of the video camera apart from detection faces；Light-source illuminating system 2-4 is installed on the top of fuselage, for providing to monitoring region Sufficient stable light source illumination.
When detecting, include the following steps:
(1) the long track 1-1 for drawing descending mechanism 1 is mounted on to the upper edge of inclination dam metope, is pulled and is detected by cable 1-4 Apparatus main body 2 makes detection device main body 2 not glide.The industrial camera 2-3-3 that will test apparatus main body 2 is adjusted to distance inspection The proper height in survey face；
(2) detection device main body 2 acquires dam image in inclination dam metope walking, industrial camera 2-3-3 in real time, when Detect on dam that starting walking bionic mechanism 2-2 will test device and be fixed on inclination dam metope when having crackle, The information such as specific width and the depth in crack are detected, then real-time Transmission telegram in reply brain；
(3) when detection device main body 2 needs to be moved to the other end of inclination dam metope, its gecko foot toe is first shunk Bristle above 2-2-1, to reduce stickiness, the steering engine on the gecko foot limbs 2-2-3 in walking bionic mechanism 2-2 turns to 90 °, Then it walks a distance, then steering engine is turned back again by gecko foot limbs 2-2-3, can detecte entire inclination to reach Dam；
(4) during detection device main body 2 traverses inclination dam metope, light-source illuminating system 2-4 starting, for detection Region provides uniformly sufficient light source；Video camera 2-3-3 also starts work, detection inclination dam wall cracking.
It although an embodiment of the present invention has been shown and described, for the ordinary skill in the art, can be with Understand and a variety of variations, modification, replacement are carried out to these embodiments without departing from the principles and spirit of the present invention and become Type, the scope of the present invention is by appended claims and its equivalent limits.

Claims (10)

1. a kind of dam slope crack detection method based on deep learning, it is characterised in that include the following steps:
(1) training neural network:
The training set and test set of 1-1. preparation neural network: the realtime graphic on dam slope first is acquired simultaneously with CCD industrial camera It is pre-processed, then image is enhanced, including overturning, translation or rotation, obtain 1000 or more crack images；Again It is used for network training using wherein 60~80% image as training set, in addition 20~40% image to be used for net as test set Network test；
The parameter of 1-2. adjustment neural network: it is adjusted by the parameter to U-net neural network, makes U-net neural network It can adapt to crack image, thus accurate Ground Split crack and background sundries；
1-3. test: the training result of U-net neural network is tested with the test set that 1-1 step obtains, by handing over and comparing MIOU and Dice coefficient is assessed；
(2) trained u-net neural network is used, the dam crack image acquired in real time is detected, accurate segmentation figure Crack and background sundries as in:
2-1. camera calibration: video camera be aligned one have it is many independent can identification point known structure, obtain calibration maps Then picture finds out the inner parameter and external parameter of video camera by the built-in function of OpenCV, wherein inner parameter includes rotation Torque battle array R and translation vector T；
2-2. corrects image: left and right cameras shooting obtains the real-time left images in dam crack；Then it is walked according to 2-1 Suddenly the spin matrix R and translation vector T obtained, is corrected real-time left images, the left images after being corrected；
2-3. uses trained u-net neural network, accurately the crack in the left images after segmentation correction and background sundries, Obtain the FRACTURE CHARACTERISTICS information of left images；
2-4. Stereo matching: the FRACTURE CHARACTERISTICS information of left images is matched, and is analyzed deviation, that is, parallax between feature, is obtained To disparity map；
The depth in 2-5. solution crack: depth map is converted by disparity map, depth depth is as follows:
Wherein, depth indicates the depth information in crack；F indicates normalized focal length；B be between two camera optical centers away from From also referred to as parallax range；Disp is parallax value；
The width in 2-6. solution crack: width H (A, B) is the maximum width in crack in image, and formula is as follows:
H (A, B)=max [h (A, B), h (B, A)]
Wherein, A is the up contour point in crack, and B is the down contour point in crack；H (A, B) is one-way distance of the point set A to point set B, h (B, A) is from point set B to point set A one-way distance.
2. the dam slope crack detection method according to claim 1 based on deep learning, it is characterised in that: step 1- The training set obtained using step 1-1 in 2, with TensorFlow training u-net network, use learning rate for 0.01~ 0.001, iteration 2000~3000 times, loss late is iterated to 0.4 hereinafter, completing the training of network model；The parameter packet of adjustment Include setting loss function and adjustment L2 regularization coefficient；The loss function L is as follows:
Wherein, n is expressed as class number, and n value 2 indicates complex background and two, crack classification；W is weight, Respectively indicate the probability of the i-th class classification and the probability of model prediction；
Weight w in loss function will be updated weight to each of image pixel, enable neural network special Learn the marginal information in crack；Weight w (x) after update, as follows:
Wherein, b indicates deviation；d1Indicate the pixel to crack image coboundary minimum distance；d2Indicate the pixel to splitting Stitch the minimum distance of image lower boundary.
3. the dam slope crack detection method according to claim 1 based on deep learning, it is characterised in that: step 2- In 4, the parallax refers to that left and right cameras is observed to obtain difference of the identical feature on x coordinate, parallax value disp=xl-xr, Wherein xlIt is distance of a certain characteristic point in crack in left video camera figure to optical axis center, xrIt is that right video camera same characteristic features point arrives The distance at right camera optical axis center.
The detection device 4. a kind of bionic gecko is creeped, it is characterised in that: including fuselage, walking bionic mechanism, video camera clamping machine Structure and light-source illuminating system；Walking bionic mechanism is installed on the lower along surrounding of fuselage, detects dress for driving bionic gecko to creep It sets and walks on inclination dam；Video camera clamping device is installed on the L-type section sliding rail of fuselage, for setting up video camera and energy Adjust height of the video camera apart from detection faces；Light-source illuminating system is installed on the top of fuselage, fills for providing to monitoring region Stable light source illumination enough.
The detection device 5. bionic gecko according to claim 4 is creeped, it is characterised in that: the fuselage includes hood, four L-type section sliding rail, two pick up the ears to detain；Hood hollow space is detection zone；Four L-type section sliding rails are vertically installed at hood Top；Two buttons of picking up the ears are separately mounted to hood on both sides of the middle, for connecting cable.
The detection device 6. bionic gecko according to claim 4 is creeped, it is characterised in that: the walking bionic mechanism includes Four bionic gecko feet, are separately mounted under fuselage along surrounding；Each bionic gecko foot includes gecko foot main body, gecko foot limbs With gecko foot toe；It is provided with steering engine on gecko foot limbs, can be turned to；Bionic gecko foot is multifilament shape surface knot The liner of a row or more is provided below in structure, each of which toe, and one or more bristle, every bristle are provided on each liner Top be provided with one or more hairy small brushes.
The detection device 7. bionic gecko according to claim 4 is creeped, it is characterised in that: the video camera clamping device packet Include cover, sliding block, fixture；There is sliding block in the quadrangle of cover, for moving on the L-type section sliding rail of fuselage, on the sliding rail of L-type section There is scale, the distance for calculating camera optical center to detection faces；It is fixed by holding screw, it can be by video camera clamping device It is fixed in detection faces certain height, also height of the adjustable video camera apart from detection faces；It is provided with above cover Machine clamp is imaged, for clamping monocular, binocular or multi-lens camera, the angle adjustable of camera optical axis and detection faces.
The detection device 8. bionic gecko according to claim 4 is creeped, it is characterised in that: the light-source illuminating system includes Four light sources, each light source tilt 5~10 ° towards fuselage axial line, for providing sufficient uniform illumination for detection zone.
The detection device 9. bionic gecko according to claim 4 is creeped, it is characterised in that: including drawing descending mechanism, for drawing It drags bionic gecko to creep detection device main body, the detection device main body includes fuselage, walking bionic mechanism, video camera clamping machine Structure, light-source illuminating system.
The detection device 10. bionic gecko according to claim 9 is creeped, it is characterised in that: the drawing descending mechanism includes length Track, fixed pulley, slide, cable；Long track installation is mounted on long track in the upper edge of inclination dam, slide and can be in long rail It is moved on road；Two fixed pulleys are mounted on slide, and the rotational angular velocity of the two is consistent；One end of cable, which is fixed on, determines cunning The wheel trench bottom of wheel is simultaneously pulled by fixed pulley, and the other end bolt of cable picks up the ears to detain in fuselage.
CN201910514001.0A 2019-06-14 2019-06-14 Dam slope crack detection method and bionic gecko crawling detection device Active CN110335255B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910514001.0A CN110335255B (en) 2019-06-14 2019-06-14 Dam slope crack detection method and bionic gecko crawling detection device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910514001.0A CN110335255B (en) 2019-06-14 2019-06-14 Dam slope crack detection method and bionic gecko crawling detection device

Publications (2)

Publication Number Publication Date
CN110335255A true CN110335255A (en) 2019-10-15
CN110335255B CN110335255B (en) 2023-04-18

Family

ID=68142065

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910514001.0A Active CN110335255B (en) 2019-06-14 2019-06-14 Dam slope crack detection method and bionic gecko crawling detection device

Country Status (1)

Country Link
CN (1) CN110335255B (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111105411A (en) * 2019-12-30 2020-05-05 创新奇智(青岛)科技有限公司 Magnetic shoe surface defect detection method
CN111294514A (en) * 2020-02-11 2020-06-16 天津大学 Data set acquisition system for image deep learning
CN111751380A (en) * 2020-07-08 2020-10-09 中国水利水电科学研究院 Concrete dam crack inspection method based on light and small unmanned aerial vehicle
CN115082450A (en) * 2022-07-26 2022-09-20 北京中科慧眼科技有限公司 Pavement crack detection method and system based on deep learning network
CN116309447A (en) * 2023-03-17 2023-06-23 水利部交通运输部国家能源局南京水利科学研究院 Dam slope crack detection method based on deep learning

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106548182A (en) * 2016-11-02 2017-03-29 武汉理工大学 Based on deep learning and the causal analytic pavement distress survey method and device of master
WO2018000731A1 (en) * 2016-06-28 2018-01-04 华南理工大学 Method for automatically detecting curved surface defect and device thereof

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018000731A1 (en) * 2016-06-28 2018-01-04 华南理工大学 Method for automatically detecting curved surface defect and device thereof
CN106548182A (en) * 2016-11-02 2017-03-29 武汉理工大学 Based on deep learning and the causal analytic pavement distress survey method and device of master

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
TANG Y C ET.AL: "Binocular vision measurement and its application in full-field convex deformation of concrete-filled steel tubular columns" *
鲍玉龙: "水库大坝安全监测中的物联网技术应用研究" *

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111105411A (en) * 2019-12-30 2020-05-05 创新奇智(青岛)科技有限公司 Magnetic shoe surface defect detection method
CN111105411B (en) * 2019-12-30 2023-06-23 创新奇智(青岛)科技有限公司 Magnetic shoe surface defect detection method
CN111294514A (en) * 2020-02-11 2020-06-16 天津大学 Data set acquisition system for image deep learning
CN111751380A (en) * 2020-07-08 2020-10-09 中国水利水电科学研究院 Concrete dam crack inspection method based on light and small unmanned aerial vehicle
CN111751380B (en) * 2020-07-08 2021-08-31 中国水利水电科学研究院 Concrete dam crack inspection method based on light and small unmanned aerial vehicle
CN115082450A (en) * 2022-07-26 2022-09-20 北京中科慧眼科技有限公司 Pavement crack detection method and system based on deep learning network
CN116309447A (en) * 2023-03-17 2023-06-23 水利部交通运输部国家能源局南京水利科学研究院 Dam slope crack detection method based on deep learning
CN116309447B (en) * 2023-03-17 2024-01-05 水利部交通运输部国家能源局南京水利科学研究院 Dam slope crack detection method based on deep learning

Also Published As

Publication number Publication date
CN110335255B (en) 2023-04-18

Similar Documents

Publication Publication Date Title
CN110335255A (en) A kind of dam slope crack detection method and bionic gecko based on deep learning is creeped detection device
CN110378909B (en) Single wood segmentation method for laser point cloud based on Faster R-CNN
CN104554341B (en) The system and method for detecting rail smooth degree
Lines et al. An automatic image-based system for estimating the mass of free-swimming fish
CN105447853B (en) Flight instruments, flight control system and method
CN106340044B (en) Join automatic calibration method and caliberating device outside video camera
CN106679567A (en) Contact net and strut geometric parameter detecting measuring system based on binocular stereoscopic vision
CN109460753A (en) A method of detection over-water floats
CN103714343B (en) Under laser line generator lighting condition, the pavement image of twin-line array collected by camera splices and homogenizing method
CN109063575A (en) A kind of intelligent grass-removing based on monocular vision is independently orderly mowed method
CN116879308A (en) Industrial machine vision system image processing method
CN110687904A (en) Visual navigation routing inspection and obstacle avoidance method for inspection robot
CN105652576B (en) The adjusting device and method of adjust automatically projector position
CN101839692A (en) Method for measuring three-dimensional position and stance of object with single camera
CN108174111B (en) Crusing robot target image grasping means
CN101999972A (en) Stereoscopic vision based auxiliary walking device for blindmen and auxiliary method thereof
CN108209926A (en) Human Height measuring system based on depth image
AU2015335554A1 (en) Method and device for selecting detection area, and elasticity detection system
CN109740584B (en) Automatic parking parking space detection method based on deep learning
CN106846352A (en) A kind of edge of a knife image acquisition method and device for camera lens parsing power test
CN108324369A (en) Method for registering and Use of Neuronavigation equipment in art based on face
CN110220481A (en) Hand-held visual detection equipment and its position and posture detection method
CN107358610A (en) Crop high flux Phenotypic examination method and system
JP2019152924A (en) Self-position identification system, vehicle, and processing device
CN105391998B (en) Automatic detection method and apparatus for resolution of low-light night vision device

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