The content of the invention
The purpose of the present invention is exactly to solve the above-mentioned problems, it is proposed that a kind of transmission of electricity based on multilayer convolutional neural networks
Circuit automatic recognition system and method, which utilizes the powerful collaboration computation capability of GPU, with reference to multilayer convolution
Parallel feature of the neural network model in training, realizes overhead transmission line visual analysis system rapidly and efficiently, utilizes GPU
Powerful computing capability, its vision system can realize real time target recognitio effect and body to crusing robot in the process of running
Existing real-time analytical capability.
To achieve these goals, the present invention adopts the following technical scheme that:
Overhead transmission line automatic recognition system based on multilayer convolutional neural networks, including:
Equipment picture database:The image or video information of barrier on classification storage transmission line of electricity, and to figure
Picture or video are marked;
Module of target detection:For positioning the target of barrier in outlet in video or image, and provide present frame
Object location information;
Target tracking module:According to the object location information detected, the prior information for being aided with detection target structural is come
Realize the online tracking to target;
Target identification module:All kinds of built-in unit types being used for realization under different scale, background, illumination, angle
Identification.
Further, the image in the equipment picture database or video invitation can be from different directions and different
Reflect the state of equipment under light conditions.
Further, including:
The target of barrier on transmission line of electricity is oriented in video or image, and the object space letter of present frame is provided
Breath;
According to the object location information detected, it is aided with the prior information for detecting target structural, using based on detection
Tracking realizes the online tracking to target;
Using multilayer convolutional neural networks model, realize on the transmission line of electricity under different scale, background, illumination and angle
The identification of device type.
Further, the target of barrier on transmission line of electricity is oriented in video or image, and present frame is provided
During object location information, using the object detection method of multilayer convolutional neural networks model.
Further, it is specially using the object detection method of multilayer convolutional neural networks model:
(1) input picture is divided into the grid of S*S, each grid provides B Bounding Box judgement, each Bounding Box judgement
Five-tuple data, 5 information words are respectively:X, y, w, h, object_prb, wherein x and y are the centre coordinate of Bounding Box, w and h
For the length and width of box, object_prob is that there are the probability P r (Object) of object in Bounding Box;
(2) each small grid provides C in the presence of the object conditional probability Pr to classify (Class | Object), so that
Obtain each class probability Pr (classification)=Pr (Class | Object) * Pr (target) in overall picture in each small grid;
Decision threshold is set, and what it is higher than decision threshold is exactly the target classification that identifies;(3) according in each small grid
Identify target classification, and the corresponding Bounding Box information of each target classification, calculate the whole cut zone of each target
Positional information, so as to be marked in an image or a video;
(4) it is predicted boundary box and the intersection area of object real estate to calculate final IOU.
Further, various deformation and various scales, the attitudes vibration and light that a large amount of tracked targets may occur are selected
Deep learning is carried out according to the sample of situation of change;Assuming that each video frame is independent of one another, according to what is detected and learnt in the past
Object module, carries out each frame picture full figure search to position the region that target is likely to occur.
Further, when realizing the online tracking to target using the tracking based on detection, according to the thing of former frame
Body position information detects whole scanning child windows in present frame, detects one or more position that target is likely to occur
Put;Positional information in present frame with the presence or absence of target and target is provided according to testing result.
Further, when realizing the online tracking to target using the tracking based on detection, it is assumed that adjacent video frames
Between the movement of object be limited, and tracked target is visible, and the movement of target is estimated with this.
Further, the identification of device type on the transmission line of electricity under different scale, background, illumination and angle is realized
When, utilize the feature operator of the method extraction picture of deep learning;Under off-line state, on the basis of existing equipment image data base
On realize the training of multilayer convolutional neural networks, obtain stable grader;The line information that crusing robot is obtained in real time
Contrasted with the information in grader, search for equipment state, realize the automatic identification of barrier on overhead transmission line.
Further, multilayer convolutional neural networks model includes:Input layer, convolutional layer, full articulamentum and output layer;
The convolutional layer that image by pretreatment is sent into convolutional neural networks by input layer carries out convolution;
Carry out down-sampled/pond:All values in Pooling windows are combined, sampled value is used as using maximum;
The projection matrix and threshold value of full articulamentum update optimization by way of stochastic gradient descent;
Output layer, that is, grader, is made of European radial basis function unit, and each RBF units that export calculate input vector x
Euclidean distance between parameter vector c, takes Euclidean distance maximum as final output result.
Beneficial effect of the present invention:
1. the present invention uses deep learning algorithm, the desirable features to reflection target can be learnt automatically, without manually setting
Meter, without redesign, once trains, performs everywhere, only need model tuning and multiple off-line training.Empirical tests are trained
Model is high for image recognition accuracy and robustness is good.
Service ability powerful 2.GPU solves the real time problems of large-scale data training and test.
Embodiment
The present invention is further illustrated with embodiment below in conjunction with the accompanying drawings.
Feature operator is time-consuming and laborious asks for the extraction of the methods of for traditional PAST, LOG, HOG, SIFT, SURF, HARRIS
Topic, convolutional neural networks (CNN) are mainly used to identify the X-Y scheme of displacement, scaling and other forms distortion consistency.Due to
The feature detection layer of CNN is learnt by training data, so when using CNN, avoids explicit feature extraction, and hidden
Learnt from training data likes;Furthermore since the neuron weights on same Feature Mapping face are identical, so network can
With collateral learning, this is also that convolutional network is connected with each other a big advantage of network relative to neuron.
Convolutional neural networks have solely with the special construction that its local weight is shared in terms of speech recognition and image procossing
Special superiority, it, which is laid out closer to actual biological neural network, weights, shares the complexity for reducing network, particularly
The image of more dimensional input vectors can directly input network this feature and avoid data reconstruction in feature extraction and assorting process
Complexity.
The present invention proposes a kind of overhead transmission line automatic recognition system based on multilayer convolutional neural networks, including:
Equipment picture database:The image or video information of barrier on classification storage transmission line of electricity, and to figure
Picture or video are marked;The shooting of the method for traditional artificial inspection and helicopter routing inspection has a large amount of grounded-lines, gold utensil, insulation
Son, shaft tower picture, these picture classifications are preserved and picture is marked with labelimg softwares.Due to barrier thousand on line
Poor ten thousand are not, and dimensional variation is very big, and illumination effect clearly, is especially subject to the more figures of a justice, the more things of a figure, a thing polymorphic, different
The similar influence for waiting multiple factors of thing, the picture in equipment picture database are required to from different directions and different light feelings
Condition reflects the state of equipment.
Module of target detection:For positioning the target of barrier in outlet in video or image, and provide present frame
Object location information;
Target tracking module:According to the object location information detected, the prior information for being aided with detection target structural is come
Realize the online tracking to target;
Target identification module:All kinds of built-in unit types being used for realization under different scale, background, illumination, angle
Identification.
The present invention proposes a kind of overhead transmission line automatic identification new method based on multilayer convolutional neural networks, and exploitation is simultaneously
Extensive multi-class multi-modal transmission line of electricity built-in unit image data base is improved, by pattern-recognition, deep learning and GPU simultaneously
Row Computational frame, according to human brain visual cortex visual signal conduction path and treatment mechanism, establishes the calculating of multilayer convolutional neural networks
Model, is obtained for transmission line of electricity built-in unit and is analyzed with gold utensil characteristics of image, identified, finally obtain one it is multi-modal
Multi-class image analysis system.Utilize the powerful collaboration computation capability of GPU, combining target image multilayer convolutional Neural net
Parallel feature of the network model in training, realizes depth convolutional neural networks training rapidly and efficiently, obtains crusing robot and regard
The real time target recognitio effect of feel system on-line operation.
Specific implementation as shown in Figure 1, including:
The first step, equipment picture database is established according to existing information.Traditional artificial inspection and the method for helicopter routing inspection
Shooting has a large amount of grounded-lines, gold utensil, insulator, shaft tower picture, these picture classifications are preserved and use labelimg softwares to figure
Piece is marked.Since barrier is multifarious on line, dimensional variation is very big, and illumination effect clearly, is especially subject to one
The influence of Yi Duotu, the more things of a figure, the multiple factors such as a thing is polymorphic, foreign matter is similar, picture database are required to never Tongfang
Position and the state of different light conditions reflection equipment.
Second step, extracts feature operator.Using the feature operator of the method extraction picture of deep learning, deep learning can be with
Automatic study, without engineer, without redesign, is once trained, performed everywhere to the desirable features of reflection target.Only
Model tuning and multiple off-line training are needed, model can map extraction from low-level image features such as edge, textures to high level by level
A series of character representation of different levels such as abstract semantics feature.The two-stage that Fig. 2 gives multilayer convolutional neural networks model is special
Levy extraction process.
3rd step, training grader.Under off-line state, convolutional Neural is realized on the basis of existing equipment image data base
The training of network, obtains stable grader.
4th step, identification.Information in line information and grader that crusing robot is obtained in real time is contrasted, and is searched
Rope equipment state.Use identifier of the algorithm for the multiple dimensioned multilayer convolutional neural networks based on GPU/CUDA Computational frames.
This needs to modify to detection module, substitutes the full articulamentum of detection module using convolutional layer in output layer, thing is used in combination
Labeled data and accurate object classification labeled data training objective detection model are surveyed in physical examination.
Overhead transmission line automatic identification new method based on multilayer convolutional neural networks is integrated in overhead transmission line and regards
Feel automatic identification module, overhead transmission line vision automatic recognition module belongs to non-contact vision image processing module.The mould
Block integrates following three types submodule:Module of target detection, target tracking module and target identification module.Three classes submodule is mutual
Linking mutually support, functionally complementary effect.Wherein the function of module of target detection is to be positioned in video image
Go out the target of barrier such as insulator chain on specific line, and the object location information of present frame is provided for target tracking module.Adopt
With the detector that algorithm is the multiple dimensioned multilayer convolutional neural networks based on GPU/CUDA Computational frames.Target tracking module
Function is the object location information provided using module of target detection, is aided with the prior information of the detection a small amount of structuring of target
To realize the long-term online tracking to target.Algorithm is used as the tracker (tracking by detection) based on detection.
Detection module can detect whole scanning child windows in present frame according to the object location information of former frame in operation, inspection
Measure one or more position that target is likely to occur.And provided in present frame and whether deposited according to tracking result and testing result
In information such as the positions of target and target.In target identification module:Automatic identification module be required to realize different scale,
All kinds of built-in unit type identifications under background, illumination, angle.Algorithm is used as based on the multiple dimensioned of GPU/CUDA Computational frames
Multilayer convolutional neural networks identifier.
A module of target detection
The functions of modules is that the target of barrier such as insulator chain on specific line is oriented in video image, and is target
Tracking module provides the object location information of present frame.Algorithm is used as the multiple dimensioned multilayer volume based on GPU/CUDA Computational frames
The detector of product neutral net.Compare other mainstream detection methods, it is mainly characterized in that:
(1) simple homing method is used, input picture just can obtain the position of all objects in image by once deducing
Put and its generic and corresponding fiducial probability.And other mainstream detection methods such as rcnn/fast rcnn/faster rcnn
Testing result is divided into two parts to solve:Object classification (classification problem), object space, that is, bounding box (regression problem).
Therefore this method is faster than mainstream depth convolution learning network, and the precision than other real-time systems is high.
(2) global prediction is carried out to picture, rather than the local easily erroneous judgement backgrounds of R-CNN, background erroneous judgement reduces half.
(3) study be target object more general characterization, than DPM method (deformable parts models) and
R-CNN models have the prediction accuracy of higher.
(4) quick identification positioning, precision is changed with the time.
The step of algorithm of target detection, is described as follows:
(1) input picture is divided into the grid (grid cell) of S*S, each grid provides B Bounding Box
(bounding box) is adjudicated, and each Bounding Box adjudicates five-tuple data, and 5 information words are respectively (x, y, w, h, object_
Prb), wherein x and y is the centre coordinate of box, w and the length and width that h is box, and object_prob is that there are the probability of object in box
Pr(Object)。
(2) each small grid provides C in the presence of the object conditional probability Pr to classify (Class | Object), so that
It can obtain each class probability Pr (classification)=Pr (Class | Object) * Pr (mesh in overall picture in each small grid
Mark),
Suitable decision threshold is set, and what it is higher than decision threshold is exactly the target classification that identifies.If in the unit not
There are target, then confidence score should be zero.
(3) the identification target classification in each small grid, and corresponding Bounding Box information, can calculate each
The whole cut zone positional information of target, coordinate, length and width etc., so as to be marked in an image or a video.
(4) it is that prediction bounding box and object are true to calculate final IOU (intersection over union)
The intersection area (in units of pixel, [0,1] section is normalized to the elemental area of real estate) in region.
IOU refers to check and evaluation function;Bounding Box refers to object space;It is calculated in real estate, that is, step 3
The higher part of each bounding box intermediate values accumulation regions.So as to predict Bounding Box and thing according to final IOU
The area of body real estate intersection.The convolutional network model of algorithm of target detection is as follows:
(1) model derives from general multilayer convolutional neural networks CNN.Entirety is by 24 convolutional layers and 2 full articulamentum groups
Into as shown in Figure 3.Wherein convolutional layer is used for extracting characteristics of image, and full articulamentum is used for prognostic chart picture position and classification is general
Rate value.
(2) last layer has 7x7=49 output, each output is 30 dimensions, and wherein 30=20 (classification)+5x2 (is returned
Bounding Box, that is, bounding box).20 be predtermined category number.It can be set according to database actual conditions.
(3) pre-training.Use preceding 20 convolutional layers of the 1000 class data training networks of ImageNet of Google and 1
Average ponds layer and 1 full articulamentum.Training image resolution ratio needs to be adjusted to 224x224.
(4) formal training.The preceding 20 convolutional layer network parameters obtained with pre-training are come 20 convolution before initialization model
The network parameter of layer, then carries out model training with VOC20 classes labeled data.To improve the precision of images, in training detection model
When, input image resolution is adjusted to 448x448.
(5) final test.Tested using the line epigraph database of point good class, obtain result.
In the training process, the loss function of model carrys out Optimized model parameter using side and error, i.e. network exports
Side and error of S*S* (B*5+C) dimensional vectors with corresponding S*S* (B*5+C) dimensional vector of true picture.Function is defined as follows:
Wherein, loss represents model loss function, coordError denotation coordination mean square errors, and iouError represents detection
Evaluation function error, classError presentation class errors.
Last whole loss function is as described below:
Wherein, denotation coordination mean square error in first square frame;IOU errors are represented in second square frame, wherein first
It is divided into the boxde confidence predictions containing object, Part II is pre- for the boxde confidence without object
Survey;Presentation class error in 3rd square frame.
B target tracking modules
For long-time tracks, the problem of key, is:When target is reappeared in camera fields of view, system
It should be able to be detected again, and starts to track again.But during tracking for a long time, tracked target will be inevitable
Generation change in shape, illumination condition change, dimensional variation, situations such as blocking.Traditional track algorithm, front end are needed with detection
Module cooperates, and after tracked target is detected, begins to enter tracking module, and hereafter, detection module would not
During getting involved in tracking.But this method have one it is fatal the defects of, i.e., when tracked target there are change in shape or blocks
When, tracking is just easy to failure.Therefore, for long-time track, or tracked target there are in the case of change in shape with
Track, it is necessary to using detection method come instead of tracking (tracking by detection) or using on-line study method come
Into line trace (tracking by learning).Wherein although the tracking of on-line study can improve in some cases
Tracking effect, but algorithm effect and the efficiency of on-line study are closely related.In other words, if the performance of on-line study is because of hardware
Or there is bottleneck in software, then tracking will fail.
Another yet unresolved issue of on-line study is tracking drift (drifting).Vision based on on-line study
Tracking is generally required is converted into two classification problems on target and background by tracking problem, that is, passes through the learning process of discriminate
Come into line trace.But when long period low deformation occurs in target and more violent illumination is blocked, discriminative model can be by
Gradually learn noise section and it is non-targeted itself, the result is that tracking box occur drift (from tracking target become track background).Phase
Than under, the method based on detection is simple and reliable, and computing cost is less.But it needs an offline learning process.Examining
, it is necessary to select the sample of substantial amounts of tracked target to be learnt and be trained before survey.This also means that training sample will
Cover various deformation and various scales, the situation of attitudes vibration and illumination variation that tracked target may occur.In other words, it is sharp
Achieving the purpose that to track for a long time with the method for detection, the selection for training sample is most important, otherwise, the robust of tracking
Property is just difficult to ensure that.This problem can be resolved by using deep learning to carry out training under large-scale line.
Consider factors above, the present invention is using the track algorithm based on detection.Operating mechanism is by detection module
Handled parallel between the complementary interference of tracking module.First, tracking module assumes the movement of object between adjacent video frames
It is limited, and tracked target is visible, and the movement of target is estimated with this.If target disappears in camera fields of view,
Tracking will be caused to fail.Detection module assumes that each depending on frame is independent of each other, and is arrived according to conventional detection and study
Object module, full figure search is carried out to each frame picture to position the region that target is likely to occur.
The priori of C track algorithms
Research finds, when such as directly carrying out vision tracking on line using conventional target track algorithm, algorithm often because of
Character representation is single, background information is under-utilized and causes tracking robustness poor.In fact, in the target following of view-based access control model
In algorithm, the priori of target is most important, and target is because itself result and electrical characteristic are often with the outer of structuring on line
See.As can learn to obtain reliable structural target priori visual information by sample statistics, it is possible to carry out target efficient
Represent, by making full use of these prioris to improve tracking accuracy in target following.Therefore, understand solid on transmission line of electricity
Determining barrier and its architectural characteristic helps to build priori.During power Transmission, aerial high voltage power line passes through gold
Belong to component to be fixed on shaft tower, shaft tower annex becomes the major obstacle thing that inspection robot completes patrol task.For being studied
110KV/220KV voltage class, wire gauge be LGJ-185 transmission line of electricity for, its component distribution situation such as Fig. 3 institutes
Show, critical piece is as follows with priori:
Shaft tower:For supporting overhead line conductor and aerial earth wire, and make between conducting wire, have between conducting wire and aerial earth wire
Enough safe distances.Mainly there are two kinds of straight line pole and strain rod tower, wherein straight line pole is used for fixing straightway power transmission line
Road, and strain rod tower is used for the fixed part of path with turning or wire-connection.The support performance evidence of shaft tower has determined its structure
Shape is multistage fork type, and more cross linear condensates are projected as in video camera imaging plane.Therefore the detection for shaft tower
It is considered that the detection algorithm based on straight line or line segment geometric primitive priori can play preferable auxiliaring effect.
Conductor spacer:Used in double divisions or more division overhead transmission lines, play fixed linking multiple fission conductor, shockproof effect
Fruit, usually has yi word pattern and two kinds cross, conductor spacer is projected as single straight line or cross video camera imaging plane.Cause
This for conductor spacer detection it is considered that the detection algorithm based on straight line or line segment geometric primitive priori can play compared with
Good auxiliaring effect.
Stockbridge damper:Use it to absorb wind shake energy, with reach mitigate vibratory overhead line effect, be typically distributed across parallel to
Shaft tower, 1~3 is installed according to shockproof requirement respectively in shaft tower both sides.Stockbridge damper length about 50cm, height about 15cm, with shaft tower
Distance is projected as positive circular or side dumbbell in 70~100cm or so, stockbridge damper video camera imaging plane between wire clamp
Shape.Therefore detection for stockbridge damper, can be by positive circular unit detection or side along grounded-line two-sided search.
Suspension clamp:For conducting wire is fixed on the suspension insulator of straight line pole, or aerial earth wire is suspended on
On the aerial earth wire stent of straight line pole.Suspension clamp is crescent, and length is about 20cm.Therefore the inspection for suspension clamp
Survey, can be carried out by side along grounded-line two-sided search arc geometric primitive.
Strain clamp:For conducting wire or aerial earth wire are fixed on strain insulator string, anchorage effect is mainly played, mainly
For strain rod tower.Strain clamp has bolt type, compression-type, wedge type three classes, and wherein bolt type strain clamp is to borrow U-shaped screw
Pressure at right angle and wire clamp waveform wire casing caused by friction effect fix conducting wire, this be in ultra-high-tension power transmission line most
For a kind of common strain clamp, the structure of other two kinds of wire clamps is relatively easy, only considers bolt type wire clamp in the design.It is this
Wire clamp uses shoe structure, its length is about 60cm.Strain clamp is generally front circle in the projection of video camera imaging plane
Or the therefore detection for suspension clamp of side dumb-bell shape, can by side along grounded-line two-sided search arc geometric primitive come into
OK.
Splicing fitting:For continuing for conducting wire and continuing for aerial earth wire, with strain rod tower, generally using aluminium hydraulic pressed connecting pipe
Or multiple aluminium parallel groove clamp connections, its total length are no more than 50cm.
Drain wire jumper:For connecting the lead of load shaft tower (strain insulator, corner and terminal shaft tower) both sides conducting wire, due to installation
During difference, its sagging radian temporarily considers that the maximum sagging radian of wire jumper is there are a great difference in detecting system design
1m.It can be detected by side along search long arc shape geometric primitive below grounded-line.
D target identification modules
Module of target detection quickly can detect whether that there are examining object in real time to input picture or video flowing.Mainly
Reason is that model is solved object detection as regression problem, while whole detection network pipeline organization is simple.
On the premise of ensureing Detection accuracy, the detection speed of 30 frames can be reached more than, while background false drop rate is low.However, identification
Object space accuracy difference and identification object accuracy rate difference are the general shortcomings of detection module.Therefore, it is necessary to add target identification
Module is further implemented in all kinds of the built-in unit types and defect recognition under different scale, background, illumination, angle, under
The object representation analysis module of one step provides type information.Algorithm is used to be multiple dimensioned more based on GPU/CUDA Computational frames
The identifier of layer convolutional neural networks, this needs to modify to detection module, and detection mould is substituted using convolutional layer in output layer
The full articulamentum of block, is used in combination object detection labeled data and accurate object classification labeled data training objective detection mould
Type.
Although above-mentioned be described the embodiment of the present invention with reference to attached drawing, model not is protected to the present invention
The limitation enclosed, those skilled in the art should understand that, on the basis of technical scheme, those skilled in the art are not
Need to make the creative labor the various modifications that can be made or deformation still within protection scope of the present invention.