Background technique
In overhead transmission lines, power transmission line vibrates with the wind, and the vibration of long time period is easy to cause strand breakage of circuit, sternly
The safe and reliable operation of overhead transmission line is threatened again.Stockbridge damper is the important devices in overhead transmission line, can be played
Reduce the effect of overhead transmission line vibration damage.But overhead transmission line implements remote distance power transmission, during which will necessarily
Across different natural environments.After undergoing long-time exposing to the weather, stockbridge damper can generate various defects, cause stockbridge damper cannot
Play the effect of script.So the defects detection of stockbridge damper is the important work of inspection during the inspection of overhead transmission line
Make.
During the inspection of current overhead transmission line, manual inspection is very time-consuming, and the work of patrol officer exists
Certain safety.In recent years, as the unmanned plane of the fast development of unmanned air vehicle technique and efficient stable is in other field
Using.Domestic many Utilities Electric Co.s start to assist inspection to work using unmanned plane successively.But picture captured by unmanned plane,
Have the characteristics that data volume is big, background is complicated, changeable, power equipment mutually blocks and high resolution relative to stockbridge damper angle.
And if only carry out defects of power equipment in analysis picture by patrol officer, then erroneous detection and erroneous judgement are easily lead to, thus meeting
The cost of maintenance is increased, this does not make a profit from unmanned air vehicle technique is introduced instead.With the development of current depth learning technology, such as
What rapidly, accurately detects that stockbridge damper defect has become a hot topic and asks in the overhead transmission line image of unmanned plane shooting
Topic.
In current object detection field, the applications of convolutional neural networks so that the performance of target detection have one greatly
Leap.Target detection model is divided into two classes: one kind is two stages detector, and prediction is divided into two steps, is respectively completed, this
A kind of Typical Representative is R-CNN, and Fast R-CNN, Faster-RCNN family, they identify that error rate is low, leaks discrimination
It is relatively low, but speed is slower, is not able to satisfy real-time detection scene;Another kind of mode is known as single phase detector, and Typical Representative is
YOLO, SSD, YoloV2, v3 etc., although there are also to be hoisted for detection accuracy quickly for speed.With regard to the power transmission line of unmanned shooting
For the picture of road, the resolution ratio of picture is very high, substantially all 8000 × 6000 level, and stockbridge damper size is average all 255
× 125 is horizontal, finds that omission factor is very high in an experiment using SSD and YOLO, and target to be detected in practical engineering applications
It is more universal to account for the lesser situation of image scaled.And there is no have speed than Faster-RCNN when testing high-resolution picture
Advantage more outstanding.
The detector of current all mainstreams, as Faster R-CNN, SSD, YOLOv2, v3 all rely on one group it is predefined
Anchor box, for a long time it is believed that the use of anchor box is that detector is successfully crucial.Although they achieve huge success,
It is worth noting that, based on the detector of anchor, there are some defects:
1. needs carefully adjust these hyper parameters of the size, aspect ratio and quantity of these anchors in the detector based on anchor.
2. predefined anchor needs are set again in the new detector task with different objects size or aspect ratio
Meter.
3. needing the detector based on anchor point densely to place anchor box over an input image to realize high recall ratio.Greatly
These most anchors are marked as negative sample during the training period.Excessive negative sample exacerbates positive and negative sample in training
Between imbalance.
4. calculating the friendship between all anchors and true value frame during the training period and when than (IOU), excessive anchor significant can also increase
Add calculation amount and EMS memory occupation amount.
But the more complex parameter of network structure is relatively more, thus calculating speed can it is slightly slow, to the occupancy of memory headroom also compared with
It is more, it is limited or to the higher occasion of requirement of real-time to be unable to satisfy some computing resources.In the inspection of stockbridge damper and its defect
Survey in precision and efficiency that there is also a certain distance, put into practical application, there are the problem of with regard to more obvious.
Summary of the invention
For above-mentioned prior art Shortcomings, the present invention provides a kind of vibration damper on power transmission line target detection and defect is known
Other method and device, by characteristic extracting module Jing Guo model reduction and in order to improve the feature enhancing module of feature representation with
And last prediction module, stockbridge damper target detection rapidly and efficiently can be realized in the limited situation of resource.
The technical solution adopted by the present invention are as follows:
A kind of vibration damper on power transmission line target detection and defect identification method, comprising the following steps:
Stockbridge damper picture is obtained by image collecting device;
Stockbridge damper picture is pre-processed to obtain pretreatment picture;
Construction feature extraction module, feature enhancing module and Pixel-level prediction module;
Optimization is attached to characteristic extracting module, feature enhancing module and Pixel-level prediction module, establishes training pattern;
It will be tested in pretreatment picture input training pattern, realize the target detection and knowledge to stockbridge damper in pretreatment picture
Not.
As further technical solution of the present invention are as follows: it is described that stockbridge damper picture is obtained by image collecting device, specifically
Include: the style of shooting combined using unmanned plane and photographic device, the stockbridge damper on transmission line of electricity is shot and is prevented
Shake hammer picture.
As further technical solution of the present invention are as follows: described to be pre-processed to obtain pretreatment figure to stockbridge damper picture
Piece;Specifically include: stockbridge damper picture is cut, rotate and scaling obtain pretreatment picture.
As further technical solution of the present invention are as follows: further include establishing picture database, the picture database is root
Picture database is established according to the mischief rule of stockbridge damper;Wherein the mischief rule of stockbridge damper is divided into that stockbridge damper is normal, stockbridge damper rust
Erosion, stockbridge damper corrosion and breakage;And the mark of target is carried out by labelimg annotation tool;And it will be in picture database
Data set is divided into training set, test set and verifying collection.
As further technical solution of the present invention are as follows: the construction feature extraction module;It specifically includes:
Convolution is separated in conjunction with depth, grouping convolution sum depth separates the concept of channel convolution, proposes grouping module and grouping
Two grouping channel modules and a grouping module are combined, obtain light-type feature extraction network by channel module;Input
Image passes through feature extraction network, constructs multiple dimensioned Fusion Features structure;Multi-scale feature fusion structure and light-type feature
Extract the common construction feature extraction module of network.
As further technical solution of the present invention are as follows: the construction feature enhances module, including channel enhancing module and
Space enhances module, and channel is enhanced module and space enhancing module is combined into a feature enhancing mould by concatenated mode
Block.
As further technical solution of the present invention are as follows: the building Pixel-level prediction module;By feature enhancing module
Enhanced feature passes through two parallel convolution branches, is respectively formed classification function and regression function, and wherein classification function is for pre-
The classification results of each pixel of mapping piece, the parameter that regression function is used to predict each pixel to bounding box.
As further technical solution of the present invention are as follows: described to characteristic extracting module, feature enhancing module and Pixel-level
Prediction module is attached optimization, establishes training pattern;It specifically includes:
Characteristic extracting module, feature enhancing module and prediction module are attached;
The data information for pre-processing picture is input in training pattern, the stockbridge damper bounding box and classification letter of corresponding picture are obtained
Breath;
For predicted value, objective function is set;
Network is input by optimizing to objective function using training data;Realize end-to-end training.
As further technical solution of the present invention are as follows: it is described that training pattern is tested, it realizes to pretreatment picture
The Target detection and identification of middle stockbridge damper, specifically includes:
Solidify using Tensorflow training pattern, and by training pattern, remove extra parameter, passes through the model pair after solidifying
It pre-processes picture and carries out target detection and defect recognition.
The present invention also proposes a kind of vibration damper on power transmission line target detection and defect recognizing device, comprising:
Image acquisition units, for obtaining stockbridge damper picture;
Image pre-processing unit obtains pretreatment picture for being pre-processed to stockbridge damper picture;
Module construction unit is used for construction feature extraction module, feature enhancing module and Pixel-level prediction module;
Training pattern unit, for being attached optimization to characteristic extracting module, feature enhancing module and Pixel-level prediction module,
Establish training pattern;
Recognition unit is detected, is tested for that will pre-process in picture input training pattern, is realized and prevent in pretreatment picture
Shake the Target detection and identification of hammer.
The invention has the benefit that
The present invention provides a kind of vibration damper on power transmission line target detection enhanced based on parameter reduction and feature and defect recognition side
Method, the model compression means of use realize the target detection and defect recognition of stockbridge damper by using a kind of grouping channel convolution;
The present invention can greatly improve the detection speed of high-resolution pictures while not losing precision;By providing a kind of spy
Sign enhancing module, can greatly improve the ability to express of feature;By Pixel-level prediction module, without using preset
Anchor can reduce calculating to a certain degree, and precision does not influence at all;By the combination of three, realizes quickly and accurately prevent
Shake hammer target detection and defect recognition, improve the efficiency of electric inspection process work, stride forward so that electric inspection process works to intelligence.
Embodiment one
As shown in Figure 1, being a kind of vibration damper on power transmission line target detection proposed by the present invention and defect identification method flow chart.
Referring to Fig.1, a kind of vibration damper on power transmission line target detection and defect identification method, comprising the following steps:
Step 100, stockbridge damper picture is obtained by image collecting device;
Step 200, stockbridge damper picture is pre-processed to obtain pretreatment picture;
Step 300, construction feature extraction module, feature enhancing module and Pixel-level prediction module;
Step 400, optimization is attached to characteristic extracting module, feature enhancing module and Pixel-level prediction module, establishes training
Model;
Step 500, it will be tested in pretreatment picture input training pattern, realize the target to stockbridge damper in pretreatment picture
Detection and identification.
The present invention provides a kind of vibration damper on power transmission line target detection enhanced based on parameter reduction and feature and defect is known
Other method, the model compression means of use realize the target detection and defect of stockbridge damper by using a kind of grouping channel convolution
Identification;The present invention can greatly improve the detection speed of high-resolution pictures while not losing precision;By providing one
Kind feature enhancing module, can greatly improve the ability to express of feature;By Pixel-level prediction module, without using presetting
Anchor, calculating can be reduced to a certain degree, and precision does not influence at all;By the combination of three, realize quickly and accurately
Stockbridge damper target detection and defect recognition improve the efficiency of electric inspection process work, stride forward so that electric inspection process works to intelligence.
In step 100, described that stockbridge damper picture is obtained by image collecting device, specifically include: using unmanned plane and
The style of shooting that photographic device combines shoots the stockbridge damper on transmission line of electricity to obtain stockbridge damper picture.
Specifically, shooting under different background by unmanned plane, the transmission line of electricity picture under different angle, increase the more of sample
Sample, to preferably improve the generalization ability of model.The picture of shooting is screened, the figure for not containing stockbridge damper is deleted
Piece.
In step 200, described that stockbridge damper picture is pre-processed to obtain pretreatment picture;It specifically includes: to shockproof
Hammer picture cut, rotate and scaling obtain pretreatment picture.
Due to very high by the resolution ratio of the stockbridge damper picture of image acquisition device, have can reach 8000 ×
6000, if being directly sent into network not only detects the precision meeting that speed detects slowly but also for Small object this for stockbridge damper very much
It is very low, so needing to cut picture;While in order to enhance the generalization ability of model, need to carry out data increasing to data
By force, comprising some data enhancements such as cutting, overturning/rotation, dimensional variation, mixup for picture.
It is cut specifically, obtaining high-resolution picture to image collecting device, Pruning strategy has weight using piecemeal
Folded cutting;The size of picture is 1300 × 1000 or so after cutting;Then the enhancing for carrying out image data, primarily with respect to
Picture is rotated, and rotation angle is 60 degree, and finally picture is reduced or amplified, completes the expansion of the early period of data set
And the work of enhancing.
It in the embodiment of the present invention, further include establishing picture database, the picture database is the defect according to stockbridge damper
Rule establishes picture database;Wherein the mischief rule of stockbridge damper be divided into stockbridge damper normal, stockbridge damper corrosion, stockbridge damper corrosion and
It is damaged;And the mark of target is carried out by labelimg annotation tool;And by the data set in picture database, it is divided into training
Collection, test set and verifying collection.
According to the mischief rule of stockbridge damper, the detection of stockbridge damper is broadly divided into that stockbridge damper is normal, stockbridge damper corrosion, stockbridge damper
Corrosion and breakage.It is above-mentioned stockbridge damper picture is pre-processed after formed pretreatment picture, by labelimg annotation tool into
The mark of row target.An xml document can be generated after the completion of one picture mark, the inside includes the labeled targets of label target
The upper left corner and bottom right angular coordinate, save the xml document of all pictures;
Then it is individually placed under two files;The name of xml document and the name of picture are to correspond in addition to suffix name
's;Established data set, being divided into training set 33% and test set 33%, there are also verifying 33% 3 data sets of collection cannot repetition
Data.
A and Fig. 2 b referring to fig. 2, wherein Fig. 2 a is grouping module flow chart in the light-type network proposed by the present invention;
Fig. 2 b is the grouping channel module flow chart proposed by the present invention;
As shown in Figure 2 a and 2 b, the construction feature extraction module;It specifically includes:
Convolution is separated in conjunction with depth, grouping convolution sum depth separates the concept of channel convolution, proposes grouping module and grouping
Two grouping channel modules and a grouping module are combined, obtain light-type feature extraction network by channel module;Input
Image passes through feature extraction network, constructs multiple dimensioned Fusion Features structure;Multi-scale feature fusion structure and light-type feature
Extract the common construction feature extraction module of network.
After feature extraction network, what reservation network third layer, the 4th layer and layer 5 exported has not input picture
With semantic and resolution ratio feature;Multiple dimensioned Fusion Features structure is constructed, the feature of layer 5 is by being upsampled to and the 4th
Merge therewith after the identical size of layer feature, be equally upsampled to again it is identical as third layer characteristic size after merge therewith, it is false
It is then then M3, M4, M5 by fused character pair if third layer, the 4th layer and layer 5 are expressed as F3, F4, F5;
Multi-scale feature fusion structure and the common construction feature extraction module of light-type feature extraction network;This module is in large size classification number
Pre-training is carried out on ImageNet according to collecting.
Since depth separates the proposition of convolution, depth, which separates convolution, can be decomposed into the convolution of standard one depth
Convolution adds one 1 × 1 convolution;Assuming that input feature vector figure size is M_F × M_F × I, output characteristic pattern size is M_F
× M_F × J, convolution kernel size are D_K × D_K, then the calculation amount of Standard convolution are as follows: D_K × D_K × I × J × M_F × M_F;
Depth separates the calculation amount of convolution: D_K × D_K × I × M_F × M_F+I × J × M_F × M_F;Bringing specific data into can be with
It was found that the calculating time is reduced to original 1/9, and calculating parameter is reduced to original 1/7 under conditions of guaranteeing accuracy rate.It is tying
Close the thought of grouping convolution;It is proposed that a kind of depth separates the concept of channel convolution.
The concept of channel convolution is separated by above-mentioned depth, grouping module and grouping channel module is proposed, wherein dividing
The structure of group module is: for Feature Mapping initial first here we assume that for F0, F0 passes through 3 × 3 depth convolution then
Then one 1 × 1 grouping convolution passes through batch normalization and activation primitive, then connect an identical structure finally obtain a characteristic pattern I
Be assumed to be F1, then F1 is added with F0;Finally output characteristic pattern is set as F2.This just constructs a grouping module.I
3 × 3 depth Convolution DW, grouping Convolution is GC, and batch normalization indicates that BN, activation primitive indicate ReLu, so
The structure specific descriptions of grouping module are: F0-> DW-> GC-> BN, ReLu6-> DW-> GC-> BN, ReLu6-> F1->
F2(F1+F0);Then building grouping channel module, only need to be added in the grouping module mentioned a grouping channel convolution and
Second 1 × 1 grouping convolution is changed into 1 × 1 convolution;It here can be grouping channel Convolution at GCWC;So point
The specific structure of group channel module is: F0-> DW-> GC-> BN, ReLu6-> DW-> 1 × 1 convolution-BN, ReLu6->
F1-> GCWC-> F2(F1+F0).
By two modules of step as above, carried out using two grouping channel modules and a grouping module ways of carrying out
It stacks, proposes a kind of new light-type feature extraction network.
Three layers of spy that multiple dimensioned structure preserves in real time specifically by feature extraction network in characteristic extracting module
Sign figure is set as third layer (F3), the 4th layer (F4), layer 5 (F5);The Feature Mapping of layer 5 is done 2 times and is upsampled to and the 4th
Element addition is pressed after the identical size of layer feature, equally do again 2 times be upsampled to it is identical as third layer characteristic size after by member
Element is added, then is then M3, M4, M5 by fused character pair, corresponding and corresponding to F3, F4, F5 has identical ruler
It is very little;In order to start iteration, only 1 × 1 convolutional layer need to be added on F5 to generate low resolution figure M5;Finally, in each conjunction
And figure on add 3 × 3 convolution and generate final Feature Mapping, this is the aliasing effect in order to reduce up-sampling;This
A final Feature Mapping collection is known as { F3, F4, F5 }, corresponds respectively to { M3, M4, M5 }.The port number of Feature Mapping is herein
256。
It is the feature enhancing module schematic diagram proposed by the present invention referring to Fig. 3;
As shown in figure 3, construction feature enhances module, including channel enhancing module and space enhance module, and channel is enhanced module
A feature enhancing module is combined by concatenated mode with space enhancing module.
By considering the potential relationship in channel and the context space relationship of feature in feature, a kind of channel enhancing is invented
Module and space enhance module, and channel enhancing module and space enhancing module is allowed to be combined into a feature by concatenated mode
Enhance module.By this module the M3 exported in step S4, the feature of tri- sizes of M4, M5 is enhanced respectively.
Assuming that enhanced characteristic pattern is respectively AM3, AM4, AM5;Enhanced feature AM3, AM4, AM5 can pass through two
Parallel convolution branch, single convolution branch are made of 4 convolutional layers.One of branch is used to classify, another is with back and forth
Return.It is classification that wherein classification branch includes one, Liang Ge branch again, the other is centrad predicted branches, classification branch is finally defeated
The vector of a W × H × C dimension out, come predict this feature figure each pixel classification results.And centrad predicted branches are defeated
The vector of H × W × 1 out, to predict the "center" (as distance of the position to affiliated object centers) of a position.In
Heart degree predicted branches are mainly used to the low quality bounding box for inhibiting to detect, without introducing any hyper parameter.Equally, branch is returned
The vector of W × H × 4 can be finally exported, to predict bounding box relevant parameter that each pixel predicts.Wherein W represents picture
Width, H represents the height of picture, and C is the classification number of classification.
In the present embodiment, the relationship of channel and spatial context in Feature Mapping is fully taken into account, utilizes a channel
Enhancing module and space enhance the mode of block coupled in series to strengthen the ability to express of Feature Mapping, do further to Feature Mapping
Enhancing.
Constructing a space first enhances module, is related to us using a kind of sky in order to capture the global space of characteristic pattern
Between enhance module, space enhancing module is defined as follows with formula:
;
It is assumed that entire characteristic pattern is usedIt indicates, whereinWithIndicate the single vector in Feature Mapping;
It can be understood as a vector of the value composition of each channel same position in characteristic pattern, the size of single vector is C × 1 × 1;
Wherein
;
Here;
WithIt is the weight matrix that can learn in the network training stage.
Specific implementation step be: Feature Mapping X first pass through respectively one 1 × 1 convolution generateWith, thenWithIt is generated respectively by shape conversion and transposition operationWithTwo squares
Battle array;Last both the above matrix by matrix multiple operation generate having a size ofMatrix;Then reconvert atThe matrix of shape;Operation is completed;
After the completion of operation, one is generatedThe matrix of shape;Then pass through a ReLu activation primitive;Extremely
This is completed in the formula of space characteristics enhancing module definitionAll operationss.
Will lead to unilateral judgement due to relying solely on spatial relationship, further by spatial relation characteristics mapping with
It is as follows that primitive character maps X fusion:;Here an attended operation by channel direction is only used, that is,
[,] enhances original feature with spatial relationship;In this way, output feature not only possesses global space relationship abundant, protect simultaneously
High-level semantics feature is stayed;It finally obtains。
It is the operation that channel enhancing is defined by following formula in building channel enhancing module:
;
The insideThe characteristic pattern of input is represented,WithRespectively represent i-th and j-th channel in X
Feature Mapping.
It is defined as foloows:
;
GAP(therein) operation is that the overall situation being added on each channel be averaged pond, mainly for the overall situation of capture Feature Mapping
Relationship;When the Feature Mapping for having C channel is by global average Chi Huahou, size becomes C × 1 × 1. and is equivalent to each channel guarantor
A real number is deposited, we can be interpreted as this real number one descriptor in this channel;Specific implementation step be:
Feature Mapping X first is by a global draw pond GAP(), size becomes C × 1 × 1;Pass through one 1 × 1 convolution respectively
It generatesWith, thenWithOne is generated by vector multiplication crossMatrix;It has operated
At;
After the completion of operation, one is generatedThe matrix of shape;Then pass through a softmax activation primitive;So far complete
In formula at channel characteristics enhancing module definitionAll operationss.
The enhanced Feature Mapping of channel characteristics and primitive character are mapped into X convergence strategy: being first converted into X-shape, then utilize formulaIt carries out merging to the end。
The space enhancing module that the channel of building enhances module and building is connected.Rule of connecting is by space characteristics
Enhancing block coupled in series enhances module in channel characteristics, so exported in channel enhancing moduleIt can enhance as space characteristics
The input of module;Finally described in output space characteristics enhancing module.Then M3, M4, M5 characteristic pattern are increased respectively
By force.Obtain AM3, AM4, AM.
Fig. 4 is the Pixel-level prediction module schematic diagram proposed by the present invention;
As shown in figure 4, in the embodiment of the present invention, the building Pixel-level prediction module;By the Enhanced feature of feature enhancing module
By two parallel convolution branches, it is respectively formed classification function and regression function, wherein classification function is for predicted pictures
The classification results of each pixel, the parameter that regression function is used to predict each pixel to bounding box.
By enhanced feature AM3, AM4, AM5 can pass through two parallel convolution branches, and single convolution branch is by 4
Convolutional layer composition.One of branch is used to classify, another is used to return;Wherein classification branch includes one, Liang Ge branch again
It is classification, the other is centrad predicted branches, classification branch finally exports a W × H × C dimension matrix, and (C here is indicated
Classification number), come predict this feature figure each pixel classification results;And centrad predicted branches export H × W × 1
Matrix, with predict a position "center" (as distance of the position to affiliated object centers).Centrad predicted branches
It is mainly used to the low quality bounding box for inhibiting to detect, without introducing any hyper parameter;Equally, W can finally be exported by returning branch
The matrix of × H × 4, to predict bounding box relevant parameter that each pixel predicts;Wherein W represents the width of picture, and H represents figure
The height of piece.
When being tested using Pixel-level prediction module, it can directly obtain each in the Feature Mapping that each input is come in
The classification score of a point, is set as here;When test, the centrad that centrad predicted branches are exported is divided with corresponding classification
Number is multiplied, and calculates final score (for being ranked up to the bounding box detected);Returning branch simultaneously can be pre- for each pixel
Survey one=(l, t, r, b) parameter, four parameters therein respectively indicate the distance of four frames of pixel (x, y);It is logical
Cross following formula carry out conversion last rectangle frame can be obtained:
;
;
Wherein (x, y) indicates the coordinate of pixel,Indicate the top left co-ordinate for the rectangle frame finally predicted, similarlyIndicate lower-left angular coordinate;Rectangle prediction result to the end is obtained by the two coordinates.
In step 400, it is described characteristic extracting module, feature enhancing module and Pixel-level prediction module are attached it is excellent
Change, establishes training pattern;It specifically includes:
Step 401, characteristic extracting module, feature enhancing module and prediction module are attached;
Step 402, the data information for pre-processing picture is input in training pattern, obtains the stockbridge damper bounding box of corresponding picture
And classification information;
Step 403, objective function is set for predicted value;
Step 404, network is input by optimizing to objective function using training data, realize end-to-end training.
Wherein, the training data that will pre-process picture is input into the neural network model by building, finally obtains about this
The information of the bounding box and classification of image deflects stockbridge damper and stockbridge damper;Specifically:
It sets objectives function for predicted value, during backpropagation, minimizes loss function;By training classification and return damage
Lose function is defined as:
,
Classification Loss is represented,It represents and returns loss,Indicate the quantity of positive sample,1 is taken herein;It is
A target function, when> 0, function takes 1, otherwise takes 0;For predicting the loss function of centrad, using formula:
;
This loss function, which is added to do together in the loss of classification and recurrence, to be optimized, and realizes that entirety is trained end to end.
In step 500, described that training pattern is tested, realize the target detection to stockbridge damper in pretreatment picture
And identification, it specifically includes: using Tensorflow training pattern, and training pattern being solidified, remove extra parameter, by solid
Model after change carries out target detection and defect recognition to pretreatment picture.
In the present embodiment, Tensorflow training pattern is utilized.Using official's mold curing tool after training, mould
Type solidification, removes extra parameter.It is encapsulated under windows10 and uses later.The realization high-resolution power transmission line of efficiently and accurately
Road stockbridge damper target detection and defect recognition.