CN108647655A - Low latitude aerial images power line foreign matter detecting method based on light-duty convolutional neural networks - Google Patents
Low latitude aerial images power line foreign matter detecting method based on light-duty convolutional neural networks Download PDFInfo
- Publication number
- CN108647655A CN108647655A CN201810465955.2A CN201810465955A CN108647655A CN 108647655 A CN108647655 A CN 108647655A CN 201810465955 A CN201810465955 A CN 201810465955A CN 108647655 A CN108647655 A CN 108647655A
- Authority
- CN
- China
- Prior art keywords
- power line
- frame
- value
- detection model
- foreign matter
- 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
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
Low latitude aerial images power line foreign matter detecting method based on light-duty convolutional neural networks belongs to computer vision field, has studied a kind of real-time detection method for power line foreign matter in unmanned plane image.Light electric line detection model is built first with convolutional neural networks, the depth characteristic of power line in aerial images is calculated；Then it utilizes convolutional neural networks to build multiple target power line foreign bodies detection model, using the convolutional layer of different length and width, the predicted value of multiscale target is calculated using depth characteristic；The video frame for finally power line detection model being utilized to filter not power line utilizes real-time power line foreign bodies detection in the aerial images of multiple target power line foreign bodies detection model realization low latitude on the video for detecting power line.
Description
Technical field
The present invention is based on depth learning technology, have studied it is a kind of in unmanned plane image power line foreign matter it is real-time
Detection method.Light electric line detection model is built first with convolutional neural networks, power line in aerial images is calculated
Depth characteristic；Then it utilizes convolutional neural networks to build multiple target power line foreign bodies detection model, uses the volume of different length and width
Lamination calculates the predicted value of multiscale target using depth characteristic；Finally power line detection model is utilized to filter no power line
Video frame utilize multiple target power line foreign bodies detection model realization low latitude aerial images on the video for detecting power line
In power line foreign bodies detection in real time.The invention belongs to computer vision fields, and in particular to deep learning, the skills such as target detection
Art.
Background technology
With the development of information technology, high-performance sensor of taking photo by plane is widely used in aeroplane photography.And nothing
It is wide to become a kind of novel foreground more so that the low latitude technology of taking photo by plane has obtained great development for the increasingly maturation of man-machine technology
Wealthy practical technique.Have benefited from this, low latitude aerial images data show magnanimity growth, and show multi-angle, complex background
The features such as, for low latitude aerial images data, realize that the processing of real-time high-efficiency has important research significance and application value.
Natural Disaster Evaluation, communications and transportation, very all various aspects such as urban planning have important application.Due to unmanned plane highly effective and safe
Characteristic, the electric power line inspection in electric system also become its important one of application field.
Power line is the important infrastructure of country, carries the important duty of electric power transport, and electric power line inspection is then
Ensure the important leverage of electric system even running.With the high speed development of China's electric system, remote high-tension circuit is extra-high
The maturation of the technologies such as pressure transmission of electricity, electric power transport show that transmission capacity is big, the remote feature of fed distance, more and more power lines
High mountain is extended to from city, in the geographical environment of the complexity such as open country.Traditional manual inspection mode takes time and effort, by natural ring
Border, the influence of weather constrain the construction of China's electric system.Unmanned plane electric power line inspection has efficiently, and safety is not bullied
The characteristics of time, influence of topography, have become the important way of electric power line inspection.It, can during unmanned plane electric power line inspection
The digital camera that carrying is connect using nobody shoots low latitude power line image, and the basic feelings of power line are contained in these images
Condition.By the processing for power line image of taking photo by plane to low latitude, power line abnormality can be found in time, to rapidly be located
Reason.
Image processing techniques includes compression of images, is divided, enhancing, and parts, the target identification such as description and identification are at image
One of the important application of reason technology.Traditional target identification technology is based on manual features, it is difficult to handle complex background, mass data
Under plurality of target.In recent years, state-of-the-art technology-depth learning technology of artificial intelligence field shows in target detection problems
Go out remarkable performance, such as the depth targets detection framework that SSD (Single Shot MultiBox Detector) is representative,
While high-precision identifies, efficiency is also further promoted.
In view of the increasingly increase of electric power line inspection image, the problems such as conventional target recognition methods effect is limited, the present invention
It is proposed a kind of aerial images power line foreign matter detecting method based on multiple dimensioned convolutional neural networks.First, structure is based on convolution
The light electric line detection model of neural network (Convolutional Neural Network, CNN), and in pre-training data
Learn the multi-level depth characteristic of power line image on collection.The power line foreign bodies detection mould based on convolutional neural networks is built later
Type handles the target of different scale using the convolutional layer of different length and width, obtains the predicted value of multiscale target.Then electricity is utilized
Line of force detection module filters the unrelated images for power line do not occur, and merges multiscale target predicted value.Finally utilize more rulers
Target prediction value is spent, using non-maxima suppression (non maximum suppression) algorithm, retains the higher side of confidence level
Frame realizes the detection of power line abnormal object.
Invention content
The present invention is different from existing power line foreign matter detecting method, using depth learning technology, proposes a kind of based on light
The real-time electric power line foreign matter detecting method of type convolutional neural networks.First, light electric is built using convolutional neural networks (CNN)
Line detection model, for power line simple target, using light electric line detection model, its number of plies is less, can meet single
The requirement of target detection, and effectively reduce training, detection time.The pre-training network on self-built survey power line image data set,
Extract power line depth characteristic.Secondly power line foreign bodies detection model is trained using convolutional neural networks, added in the model
The convolutional layer of different length and width, and calculate predicted value simultaneously in multiple layers.The output of different layers is merged later, to learn more rulers
Spend the depth characteristic of target.Pre-training, and the method for using data augmentation are carried out using self-built power line foreign matter data set, into
Row overturns, cuts at random, color change further increases generalization ability to extended amount of data.Finally, in power line foreign matter
Detection-phase uses the unrelated frame in power line detection model removal video, retains the key frame for including power line, and carry first
The power line bounding box predicted value in key frame is taken out, utilizes power line foreign bodies detection model inspection key frame later, and obtain
The predicted value of each target filters more similar bounding box using non-extreme value restrainable algorithms, retains the higher bounding box of confidence level,
Later using obtained power line bounding box and foreign matter object boundary frame, the quick, accurate of power line image foreign matter of taking photo by plane is realized
Detection.This method main process is as shown in Fig. 1, can be divided into following three steps：Power line based on convolutional neural networks is different
Object target detection model construction, neural network pre-training, power line Anomaly target detection.
(1) the power line detection model structure based on light-duty convolutional neural networks
Research object of the present invention is aerial images, in order to effectively remove the unrelated frame in video, builds one kind first and is based on
The power line detection model of light-duty convolutional neural networks, the prototype network is simple in structure, and the number of plies is less, in effectively extraction power line
It ensure that the real-time of detection on the basis of feature.For power line foreign matter two major classes not-kite, balloon, construct and be based on
The power line foreign bodies detection model of light-duty convolutional neural networks, two model substeps detect, change on the basis of improving real-time
It has been apt to accuracy of detection.
(2) neural network pre-training
For power line detection model, instructed in advance using power line image data set (Powerline Image Dataset)
Practice, for power line foreign bodies detection model, the use of balloon, the kite picture voluntarily collected is source data, is calculated using data augmentation
Method is translated, is cut, is changed colour, and training dataset is used as to extend to 4000.Above-mentioned data set contain different scale,
Illumination condition, the power line of shooting angle, power line foreign matter image, can effectively learn the depth characteristic under different condition.
(3) power line Anomaly target detection
The present invention proposes a kind of multiple power line target detection method.It is regarded first with power line detection model to taking photo by plane
Frequency is detected frame by frame, and gives up the unrelated frame of no electric power line target.For there are the key frame of electric power line target, utilizing electricity
Line of force foreign bodies detection model is further detected, and power line parameters predicted value and the prediction of foreign matter object boundary frame is calculated
Value, to judge whether power line foreign matter target.
Compared with prior art, the present invention having following apparent advantage and advantageous effect：
First, traditional manual features power line target identification method is compared, the present invention utilizes advanced convolutional Neural net
Network builds light electric line detection model and power line foreign bodies detection model, realizes the unrelated filtering frames of power line image, greatly
Improve detection efficiency, and ensure that power line foreign bodies detection real-time using light-duty network.It is experimentally confirmed that using the structure
Unrelated frame can be effectively filtered in unmanned plane image, and detection efficiency is greatly improved.Meanwhile in light electric line foreign bodies detection
Multiple dimensioned convolutional layer is added in model, learns the foreign matter characteristics of image of different scale, is clapped to adapt to unmanned plane different distance
Take the photograph multiple dimensioned situation caused by different target.
Finally, for the power line image after screening, calculated using light electric line model and power line foreign matter model
Power line parameters predicted value and foreign matter object boundary frame predicted value, to judge whether power line foreign matter target.
It is experimentally confirmed that the deep-neural-network based on VGG-16, is learnt, Ke Yi using multiple dimensioned convolutional layer
The mAP (mean average precision, average accuracy of identification) that 74.3% is realized on VOC2007 databases, keeps simultaneously
The detection speed of 59FPS.Therefore, this method is moved in power line Anomaly target detection task, it is efficient, accurate for realizing
Really, real-time electric power line inspection is practical and has important application value.
Description of the drawings：
Aerial images power line foreign matter detecting method flow charts of the Fig. 1 based on light-duty convolutional neural networks
Fig. 2 light electric line detection model Organization Charts
Fig. 3 power line foreign bodies detection model support compositions
Fig. 4 power line foreign bodies detection procedure charts
Specific implementation mode
It is a specific implementing procedure below, but the range that this patent is protected is not limited to this according to foregoing description
Implementing procedure.
Step 1：Power line foreign matter target detection model construction based on convolutional neural networks
Step 1.1：Power line detection model structure based on light-duty convolutional neural networks
Presently, there are deep learning target detection model, application scenarios are wider, often can detect thousands of type objects, such as
YOLO9000 can detect 9418 classifications.Under electric power line inspection scene, targeted species are extremely limited, mainly power line, gas
Ball and kite three classes, in this model, it is only necessary to identify power line, feature is extremely limited, current existing deep learning mesh
It marks detection model and is directed to power line scene too redundancy, and light weight model is effective in this case, light weight model can
To identify limited assortment target, while improving detection speed.
Increase income deep learning frame Caffe realization of the light-duty convolutional neural networks proposed by the present invention based on mainstream, this step
Rapid concrete structure diagram is shown in attached drawing 2.Input power line aerial images carry out convolution by 6 convolutional layers, and convolution kernel size is 3 ×
3, first four convolutional layer is all using batch normalization (Batch Normalization), it is by the input to subsequent activation function
It is normalized, it is in standardized normal distribution (mean value 1, standard deviation 0) to make batch, so that numerical value is more stablized, and is returning in batches
Linear amending unit (Rectified Linear Unit, ReLU) will be used to be used as activation primitive after one change, make model convergence speed
Degree is faster.Maximum value pond (Max Pooling) operation is carried out after the 4th convolutional layer, to reduce characteristic dimension, is reduced and is calculated
Amount.In the 5th convolutional layer, we use one 3 × 3 convolution kernel as class prediction module, and the port number of output is 6, often
A channel corresponds to the confidence level of an anchor frame.In the 6th convolutional layer, we predict side using one 3 × 3 convolution kernel
Boundary's frame.For each prediction block, generic is determined according to the class prediction value being calculated, and filter out and belong to the pre- of background
Survey frame.Then the lower prediction block of threshold value is filtered out with 0.5 confidence threshold value, and it is pre- to retain higher first 200 of confidence level
Survey frame.Non-maxima suppression NMS algorithms are finally used, the prediction block that threshold value is more than 0.7 is filtered out, finally obtains prediction result
Step 1.2：Power line foreign bodies detection model construction based on convolutional neural networks
The convolutional neural networks concrete structure diagram that this step proposes is shown in attached drawing 3.Input power line foreign matter aerial images pass through
10 convolutional layers carry out convolution, and convolution kernel size is 3 × 3, and preceding 6 convolutional layers are different for extracting power line as major network
Object target signature, and pondization operation is added after the 2nd, 4,6 convolutional layer.7th, 8 layer, respectively there are one 3 × 3 convolution kernel, is all adopted
With batch normalize (Batch Normalization), the input of subsequent activation function will be normalized in it, make batch be in
Standardized normal distribution (mean value 1, standard deviation 0), so that numerical value is more stablized, and will use after batch normalizes and linearly repair
Positive unit (Rectified Linear Unit, ReLU) is used as activation primitive, makes model convergence rate faster.It is each after 7,8 layers
The maximum pond layer that one span of addition is 2, the length and width of input feature vector are halved.7th, 8,9,10 convolutional layer, all as prediction
Module, each module contain the convolutional layer there are two 3 × 3, are respectively intended to carry out class prediction and bounding box prediction, in this way, we
Remain the predicted value of different scale between multilayer.Later, we are translated into two-dimensional array, and the first dimension is number of samples, the
Two dimension is port number, and all outputs are stitched together in second dimension, realize the merging of multi-scale prediction value.For each
Prediction block determines generic according to the class prediction value being calculated, and filters out the prediction block for belonging to background.Then with
0.5 confidence threshold value filters out the lower prediction block of threshold value, and retains higher preceding 200 prediction blocks of confidence level.Finally adopt
With non-maxima suppression NMS algorithms, the prediction block that threshold value is more than 0.7 is filtered out, prediction result is finally obtained.
Step 2：Neural network pre-training
The present invention is used trains light electric line detection model using power line image data set, uses power line foreign matter number
According to collection training power line foreign matter model, power line image is first fed into power line detection model, filters unrelated frame, it later will packet
Key frame containing power line is sent into power line foreign bodies detection model, to realize real-time electric power line foreign matter target detection.
Step 2.1：Target detection model pre-training
Step 2.1.1：Build pre-training data set
The pre-training stage selects power line image data set (Powerline Image Dataset) training electric power line target
Detection model, including 2000 power line Aerial Images and 2000 background Aerial Images.Power line Aerial Images are derived from difference
The different regions in season, picture size are 512 × 512.Power line foreign matter data set is selected to train power line foreign bodies detection model,
Including balloon, two class Aerial Images of kite, have 1000, cover different angle, area, background respectively.
2.1.2 model pre-training
In power line foreign matter scene, frame can appear in any one position of picture, and have arbitrary size.For
Simplified search process, power line foreign matter model use default boundary frame namely anchor frame (anchor box), and are to search with anchor frame
Suo Qidian.The setting of anchor frame includes two aspects of scale and length-width ratio.For inputting size w × h, for given size s ∈
(0,1) will generate the bounding box that size is ws × hs；For given ratio r>0, will generate size isBoundary
Frame.S takes 0.1,0.25,0.5, r to take 0.5,1,2 in the present invention.For the pixel of each input, anchor frame is given tacit consent to its center sampling
5.In the training process, it is first determined the actual value (ground truth) in training data is matched with which anchor frame, therewith
Bounding box corresponding to corresponding anchor frame is predicted.For each real goal in photo, hands over and compare with it
The maximum anchor frame matching of (Intersection over Union, IoU) value.It is the probability value for describing frame distance to hand over and compare, such as
Shown in formula (1)：
Wherein α is prediction result, and ξ is real border value, and big friendship is simultaneously more much like than two frames of expression, small friendship and ratio
Indicate that two each frames are dissimilar.Anchor frame is not matched for remaining, if the IoU of some actual value is more than threshold value 0.5, then anchor frame
Also it will be matched with this actual value.
In power line detects mould model and power line foreign bodies detection model, L (x, c, l, g) represents loss function, definition
For the weighted sum of site error (locatization loss, loc) and confidence level error (confidence loss, conf),
As shown in formula (2), x is the training image of input, and c is classification confidence level predicted value, and l is that anchor frame corresponds to bounding box predicted value, g
For the location parameter of actual value, N is the positive sample quantity of anchor frame, and α is the adjusting ratio of foreground loss function and background loss function
Example, takes 1 herein.
Lloc(x, l, g) is the loss function of bounding box prediction, as shown in formula (3).Wherein cx, cy are in bounding box
Heart coordinate, w, h are boundary frame width and height, anchor frame position d=(dcx, dcy, dw, dh) indicate, corresponding bounding box b=(bcx,
bcy, bw, bh), That is conversion value of the bounding box relative to anchor frame is calculated according to formula (4) (5) (6) (7)
Go out.The predicted value of the m parameter of bounding box is corresponded to for i-th of anchor frame.Pos represents positive sample set, and i indicates anchor frame serial number, j tables
Show actual value serial number.WhenWhen indicate i-th of anchor frame and the matching of j-th actual value, and the classification of actual value is k, whenWhen indicate mismatch.For site error, using Smooth L1 functions.
Lconf(x, c) represents the loss function of class prediction, and wherein x indicates that the image of input, Neg represent negative sample collection
It closes, o indicates the anchor frame serial number for being derived from positive sample, indicates that the anchor frame serial number for being derived from negative sample, t indicate actual value serial number.It is used for
Illustrate matching status, indicates to work asWhen indicate o-th of anchor frame and the matching of t-th actual value, and the classification of actual value is p,
WhenWhen indicate mismatch.As shown in formula (8)：
Then, loss function is minimized to be trained.Using stochastic gradient descent (SGD) method, above-mentioned cost letter is minimized
Number, prediction is all calculated for the characteristic pattern result of multiple and different scale convolutional layers, and merges different layers prediction output.Pre-training needs
All size of data are normalized, therefore original image is reset to 512 × 512 pixels for pre-training by the present invention.Study
Rate is the most important parameter of stochastic gradient descent method, determines the speed of right value update.Momentum parameter and weights decay factor
Trained adaptivity can be improved.By Germicidal efficacy, learning rate is set as 10 by the present invention-3, momentum parameter is set as 0,99,
Weights decay factor is set as acquiescence 0.0005. stochastic gradient descents learning process and is accelerated by NVIDIA TITAN XP equipment,
60000 iteration are carried out altogether.
The detailed pre-training process of power line target detection model is as follows, whereinFor initial power line Boundary Prediction
Value and class prediction value, c1, l1For final power line Boundary Prediction value and class prediction value,Represent power line detection model
Network parameter, u ∈ (0,15) are the serial number of parameter iteration.
1) power line image data set is read in, and power line detection model is initialized
2) it is calculated using power line detection network, output boundary predicted valueWith class prediction value
3) willWithEntrance loss function, and loss function is exported and is summed, namely the output of two loss functions is closed
And obtain loss output valve
4) basisNetwork is detected using SGD training power lines, undated parameter is
5) basisNetwork is detected using SGD training power lines, undated parameter is
6) it repeats above-mentioned 2-5 and walks 15 acquisition power line detection model pre-training final argument β1, c1, l1。
The detailed pre-training process of power line foreign bodies detection model is as follows, whereinFor initial power line Boundary Prediction
Value and class prediction value, c2, l2For final power line Boundary Prediction value and class prediction value,Represent power line detection model
Network parameter, u ∈ (0,15) are the serial number of parameter iteration.
1) power line foreign matter image data set is read in, and power line detection model is initialized
2) it is calculated using power line detection network, output boundary predicted valueWith class prediction value
3) willWithEntrance loss function, and loss function is exported and is summed, namely the output of two loss functions is closed
And obtain loss output valve
4) basisNetwork is detected using SGD training power lines, undated parameter is
5) basisNetwork is detected using SGD training power lines, undated parameter is
6) it repeats above-mentioned 2-5 and walks 15 acquisition power line detection model pre-training final argument β2, c2, l2
Step 3：Power line foreign matter identifies
In aerial images, there are a large amount of unrelated frames, such as take off landing and the power line periphery flight course of unmanned plane,
They do not include electric power line target, reduce the recognition efficiency of power line foreign matter image, so we use power line first
Detection model is detected, and the frame there is no power line is not processed, for there are the frames of power line to carry out next step
Power line foreign bodies detection, to improve whole detection speed.
Step 3.1 power line visual inspection is surveyed
First by video frame input power line target detection model, and output boundary frame predicted value and class prediction value, so
Non-maxima suppression algorithm is used afterwards, is retained the higher bounding box of confidence level, is finally drawn frame.
Step 3.1.1 power lines target category and Boundary Prediction
By an image x to be detectediIt is sent into power line target detection model, and exports predicted boundary value c1And classification
Predicted value l1, since each pixel can generate several anchor frames, so we can predict a large amount of similar bezel, clusters.
Step 3.1.2 power lines target category and Boundary Prediction result optimizing
For calculated a large amount of similar bezel, clusters in step 3.1.1, we will inhibit redundancy using non-maxima suppression
Frame, sorted according to confidence level by institute is framed, and choose the highest frame of confidence level, it is framed to traverse remaining institute later, if
It is more than threshold value 0.8 with the IoU values of the highest frame of present score, we are just deleted, and constantly repeat the above process, final to retain
The higher frame of confidence level.Finally, in non-maxima suppression treated frame set, the frame that confidence level is more than 0.6 is drawn
For final frame.
3.2 power line foreign matter target detections
By key frame input power line foreign matter target detection model comprising power line of the step 3.1 after processed, calculate
Foreign matter object boundary predicted value and class prediction value, for there are the key frame of foreign matter, drawing its foreign matter bounding box, and judge with
Whether power line bounding box overlaps, and finally draws the bounding box of coincidence.
Step 3.2.1 power line foreign matter target categories and Boundary Prediction
By an image x to be detectediIt is sent into power line foreign matter target detection model, and exports predicted boundary value c2With
Class prediction value l2, each pixel can generate several anchor frames, so we can predict a large amount of similar bezel, clusters.
Step 3.2.2 power lines target category and Boundary Prediction result optimizing
For calculated a large amount of similar bezel, clusters in step 3.2.2, we will inhibit redundancy using non-maxima suppression
Frame, and retain the frame that confidence level is more than 0.6.Then by power line foreign matter frame predicted value and power line frame predicted value
Comparison, and the frame that IoU is 0 is deleted, finally remaining frame is drawn.
Step 3.3：Evaluation
The present invention, which is used, evaluates boundary prediction result based on the average error criterion of degree absolutely.Mean absolute error is
MAE, formula are as follows：
ei=| fi-yi| (9)
Wherein, fiIndicate predicted value, yiIndicate actual value yi, ei, it is absolute error.
Claims (5)
1. the low latitude aerial images power line foreign matter detecting method based on light-duty convolutional neural networks, it is characterised in that：
First, light electric line detection model is built using convolutional neural networks, it is pre- on self-built survey power line image data set
Training network, extracts power line depth characteristic；Secondly power line foreign bodies detection model is trained using convolutional neural networks, in the mould
The convolutional layer of different length and width is added in type, and calculates predicted value simultaneously in multiple layers；The output of different layers is merged later, to
Learn the depth characteristic of multiscale target；Pre-training is carried out using self-built power line foreign matter data set, and uses data augmentation
Method, overturn, cut at random, color change, to extended amount of data, further increasing generalization ability；Finally, in electricity
The line of force foreign bodies detection stage uses the unrelated frame in power line detection model removal video, retains the pass for including power line first
Key frame, and the power line bounding box predicted value in key frame is extracted, utilize power line foreign bodies detection model inspection crucial later
Frame, and the predicted value of each target is obtained, more similar bounding box is filtered using non-extreme value restrainable algorithms, it is higher to retain confidence level
Bounding box, later using obtained power line bounding box and foreign matter object boundary frame, realization is taken photo by plane power line image foreign matter
Detection.
2. detection method according to claim 1, it is characterised in that：
Step 1：Power line foreign matter target detection model construction based on convolutional neural networks
Step 1.1：Power line detection model structure based on light-duty convolutional neural networks
Input power line aerial images carry out convolution by 6 convolutional layers, and convolution kernel size is 3 × 3, and first four convolutional layer is all
It is normalized using batch, the input of subsequent activation function will be normalized in it, and it is in standardized normal distribution to make batch, in batch
Linear amending unit will be used as activation primitive after normalization, and make model convergence rate faster；It is carried out most after the 4th convolutional layer
Big value pondization operation reduces calculation amount to reduce characteristic dimension；In the 5th convolutional layer, one 3 × 3 convolution kernel is used
As class prediction module, the port number of output is 6, and each channel corresponds to the confidence level of an anchor frame；In the 6th convolutional layer
In, using one 3 × 3 convolution kernel come predicted boundary frame；It is true according to the class prediction value being calculated for each prediction block
Determine generic, and filters out the prediction block for belonging to background；Then the lower prediction of threshold value is filtered out with 0.5 confidence threshold value
Frame, and retain higher preceding 200 prediction blocks of confidence level；Non-maxima suppression NMS algorithms are finally used, threshold value is filtered out and is more than
0.7 prediction block, finally obtains prediction result；
Step 1.2：Power line foreign bodies detection model construction based on convolutional neural networks
Input power line foreign matter aerial images carry out convolution by 10 convolutional layers, and convolution kernel size is 3 × 3, preceding 6 convolution
Layer is used as major network, and pondization operation is added for extracting power line foreign matter target signature, and after the 2nd, 4,6 convolutional layer；The
7,8 layers there are one 3 × 3 convolution kernels respectively, are all normalized using batch, it will carry out normalizing to the input of subsequent activation function
Change, it is in standardized normal distribution to make batch so that numerical value is more stablized, and linear amending unit conduct will be used after batch normalizes
Activation primitive makes model convergence rate faster；The maximum pond layer that a span is 2 is respectively added after 7,8 layers, by input feature vector
Length and width halve；7th, 8,9,10 convolutional layer is all used as prediction module, each module to contain the convolutional layer there are two 3 × 3, respectively
For carrying out class prediction and bounding box prediction, in this way, remaining the predicted value of different scale between multilayer；Later, it is converted
For two-dimensional array, the first dimension is number of samples, and the second dimension is port number, and all outputs are stitched together in second dimension,
Realize the merging of multi-scale prediction value；For each prediction block, generic is determined according to the class prediction value being calculated, and
Filter out the prediction block for belonging to background；Then the lower prediction block of threshold value is filtered out with 0.5 confidence threshold value, and retains confidence
Spend higher preceding 200 prediction blocks；Non-maxima suppression NMS algorithms are finally used, the prediction block that threshold value is more than 0.7 is filtered out,
Finally obtain prediction result.
3. detection method according to claim 1, it is characterised in that：
Step 2：Neural network pre-training
Using using power line image data set to train light electric line detection model, electricity is trained using power line foreign matter data set
Power line image is first fed into power line detection model, filters unrelated frame, will include power line later by line of force foreign matter model
Key frame is sent into power line foreign bodies detection model, to realize real-time electric power line foreign matter target detection；
Step 2.1：Target detection model pre-training
Step 2.1.1：Build pre-training data set
The pre-training stage selects power line image data set to train power line target detection model, the figure including multiple power lines are taken photo by plane
Picture and multiple background Aerial Images；Power line Aerial Images are derived from the different regions of Various Seasonal, and picture size is 512 × 512；
Power line foreign matter data set is selected to train power line foreign bodies detection model, including balloon, two class Aerial Images of kite, cover difference
Angle, area, background；
2.1.2 model pre-training
In power line foreign matter scene, frame can appear in any one position of picture, and have arbitrary size；Power line
Foreign matter model uses default boundary frame namely anchor frame, and using anchor frame as search starting point；The setting of anchor frame includes scale and length-width ratio
Two aspects；For inputting size w × h, for given size s ∈ (0,1), the bounding box that size is ws × hs will be generated；
For given ratio r>0, will generate size isBounding box；Middle s takes 0.1,0.25,0.5, r to take 0.5,1,2；It is right
In the pixel of each input, anchor frame 5 is given tacit consent to its center sampling；In the training process, it is first determined true in training data
Real value is matched with which anchor frame, and bounding box corresponding to corresponding anchor frame is predicted；For the true mesh of each of photo
Mark, hands over it and anchor frame more maximum than IoU values matches；It is the probability value for describing frame distance to hand over and compare, as shown in formula (1)：
Wherein α is prediction result, and ξ is real border value, and big friendship is simultaneously more much like than two frames of expression, and small friendship and ratio indicate
Two each frames are dissimilar；Anchor frame is not matched for remaining, if the IoU of some actual value is more than threshold value 0.5, then anchor frame also will
It is matched with this actual value；
In power line detects mould model and power line foreign bodies detection model, L (x, c, l, g) represents loss function, is defined as position
The weighted sum for setting error and confidence level error, as shown in formula (2), x is the training image of input, and c predicts for classification confidence level
Value, l are that anchor frame corresponds to bounding box predicted value, and g is the location parameter of actual value, and N is the positive sample quantity of anchor frame, and α damages for foreground
The adjusting ratio for losing function and background loss function, takes 1 herein；
Lloc(x, l, g) is the loss function of bounding box prediction, as shown in formula (3)；Wherein cx, cy are that the center of bounding box is sat
Mark, w, h are boundary frame width and height, anchor frame position d=(dcx, dcy, dw, dh) indicate, corresponding bounding box b=(bcx, bcy, bw,
bh),That is conversion value of the bounding box relative to anchor frame is calculated according to formula (4) (5) (6) (7)；The predicted value of the m parameter of bounding box is corresponded to for i-th of anchor frame；Pos represents positive sample set, and i indicates that anchor frame serial number, j indicate
Actual value serial number；WhenWhen indicate i-th of anchor frame and the matching of j-th actual value, and the classification of actual value is k, when
When indicate mismatch；For site error, using Smooth L1 functions；
Lconf(x, c) represents the loss function of class prediction, and wherein x indicates that the image of input, Neg represent negative sample set, o tables
Show the anchor frame serial number for being derived from positive sample, indicates that the anchor frame serial number for being derived from negative sample, t indicate actual value serial number；For illustrating
With state, indicate to work asWhen indicate o-th of anchor frame and the matching of t-th actual value, and the classification of actual value is p, whenWhen indicate mismatch；As shown in formula (8)：
Then, loss function is minimized to be trained；Using stochastic gradient descent method, above-mentioned cost function is minimized, for more
The characteristic pattern result of a different scale convolutional layer all calculates prediction, and merges different layers prediction output；Pre-training needs will own
Size of data normalizes, therefore original image is reset to 512 × 512 pixels for pre-training；Learning rate is set as 10-3,
Momentum parameter is set as 0,99, and weights decay factor is set as acquiescence 0.0005. stochastic gradient descent learning processes and passes through
NVIDIA TITAN XP equipment accelerates, and carries out 60000 times or more iteration altogether.
4. detection method according to claim 1, it is characterised in that：
The detailed pre-training process of power line target detection model is as follows, whereinFor initial power line Boundary Prediction value and
Class prediction value, c1, l1For final power line Boundary Prediction value and class prediction value,Represent power line detection model network
Parameter, u ∈ (0,15) are the serial number of parameter iteration；
1) power line image data set is read in, and power line detection model is initialized
2) it is calculated using power line detection network, output boundary predicted valueWith class prediction value
3) willWithEntrance loss function, and loss function is exported and is summed, namely the output of two loss functions is merged,
Obtain loss output valve
4) basisNetwork is detected using SGD training power lines, undated parameter is
5) basisNetwork is detected using SGD training power lines, undated parameter is
6) it repeats above-mentioned 2-5 and walks 15 acquisition power line detection model pre-training final argument β1, c1, l1；
The detailed pre-training process of power line foreign bodies detection model is as follows, whereinFor initial power line Boundary Prediction value and
Class prediction value, c2, l2For final power line Boundary Prediction value and class prediction value,Represent power line detection model network
Parameter, u ∈ (0,15) are the serial number of parameter iteration；
1) power line foreign matter image data set is read in, and power line detection model is initialized
2) it is calculated using power line detection network, output boundary predicted valueWith class prediction value
3) willWithEntrance loss function, and loss function is exported and is summed, namely the output of two loss functions is merged,
Obtain loss output valve
4) basisNetwork is detected using SGD training power lines, undated parameter is
5) basisNetwork is detected using SGD training power lines, undated parameter is
6) it repeats above-mentioned 2-5 and walks 15 acquisition power line detection model pre-training final argument β2, c2, l2。
5. detection method according to claim 1, it is characterised in that：
Step 3：Power line foreign matter identifies
It is detected first using power line detection model, the frame there is no power line is not processed, for there are electric power
The frame of line carries out the power line foreign bodies detection of next step, to improve whole detection speed；
Step 3.1 power line visual inspection is surveyed
First by video frame input power line target detection model, and output boundary frame predicted value and class prediction value, then make
With non-maxima suppression algorithm, retains the higher bounding box of confidence level, finally draw frame；
Step 3.1.1 power lines target category and Boundary Prediction
By an image x to be detectediIt is sent into power line target detection model, and exports predicted boundary value c1And class prediction
Value l1, since each pixel can generate several anchor frames, so a large amount of similar bezel, clusters can be predicted；
Step 3.1.2 power lines target category and Boundary Prediction result optimizing
For calculated a large amount of similar bezel, clusters in step 3.1.1, the frame of redundancy will be inhibited using non-maxima suppression, by institute
Framed to sort according to confidence level, and choose the highest frame of confidence level, it is framed to traverse remaining institute later, and if present score
The IoU values of highest frame are more than threshold value 0.8, are just deleted, constantly repeat the above process, and final reservation confidence level is higher
Frame；Finally, in non-maxima suppression treated frame set, it is final frame to draw the frame that confidence level is more than 0.6；
3.2 power line foreign matter target detections
By key frame input power line foreign matter target detection model comprising power line of the step 3.1 after processed, foreign matter is calculated
Object boundary predicted value and class prediction value, for there are the key frame of foreign matter, drawing its foreign matter bounding box, and judgement and electric power
Whether line boundary frame overlaps, and finally draws the bounding box of coincidence；
Step 3.2.1 power line foreign matter target categories and Boundary Prediction
By an image x to be detectediIt is sent into power line foreign matter target detection model, and exports predicted boundary value c2And classification
Predicted value l2, each pixel can generate several anchor frames, so a large amount of similar bezel, clusters can be predicted；
Step 3.2.2 power lines target category and Boundary Prediction result optimizing
For calculated a large amount of similar bezel, clusters in step 3.2.2, the frame of redundancy will be inhibited using non-maxima suppression, and protect
Reserve the frame that confidence level is more than 0.6；Then power line foreign matter frame predicted value and power line frame predicted value are compared, and deleted
Except the frame that IoU is 0, finally remaining frame is drawn；
Step 3.3：Evaluation
Boundary prediction result is evaluated using based on averagely degree error criterion absolutely；Mean absolute error is MAE, and formula is such as
Under：
ei=| fi-yi| (9)
Wherein, fiIndicate predicted value, yiIndicate actual value yi, ei, it is absolute error.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810465955.2A CN108647655B (en) | 2018-05-16 | 2018-05-16 | Low-altitude aerial image power line foreign matter detection method based on light convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810465955.2A CN108647655B (en) | 2018-05-16 | 2018-05-16 | Low-altitude aerial image power line foreign matter detection method based on light convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108647655A true CN108647655A (en) | 2018-10-12 |
| CN108647655B CN108647655B (en) | 2022-07-12 |
Family
ID=63755957
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810465955.2A Active CN108647655B (en) | 2018-05-16 | 2018-05-16 | Low-altitude aerial image power line foreign matter detection method based on light convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108647655B (en) |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109389322A (en) * | 2018-10-30 | 2019-02-26 | 福州大学 | The disconnected broken lot recognition methods of grounded-line based on target detection and long memory models in short-term |
| CN109472214A (en) * | 2018-10-17 | 2019-03-15 | 福州大学 | One kind is taken photo by plane foreign matter image real-time detection method based on deep learning |
| CN109726741A (en) * | 2018-12-06 | 2019-05-07 | 江苏科技大学 | A kind of detection method and device of multiple target object |
| CN109902730A (en) * | 2019-02-21 | 2019-06-18 | 国网山东省电力公司临沂供电公司 | Broken strand of power transmission line detection method based on deep learning |
| CN110033453A (en) * | 2019-04-18 | 2019-07-19 | 国网山西省电力公司电力科学研究院 | Based on the power transmission and transformation line insulator Aerial Images fault detection method for improving YOLOv3 |
| CN110033451A (en) * | 2019-04-17 | 2019-07-19 | 国网山西省电力公司电力科学研究院 | A kind of power components defect inspection method based on SSD framework |
| CN110070530A (en) * | 2019-04-19 | 2019-07-30 | 山东大学 | A kind of powerline ice-covering detection method based on deep neural network |
| CN110163081A (en) * | 2019-04-02 | 2019-08-23 | 宜通世纪物联网研究院(广州)有限公司 | SSD-based real-time regional intrusion detection method, system and storage medium |
| CN110175524A (en) * | 2019-04-26 | 2019-08-27 | 南京航空航天大学 | A kind of quick vehicle checking method of accurately taking photo by plane based on lightweight depth convolutional network |
| CN110232370A (en) * | 2019-06-21 | 2019-09-13 | 华北电力大学（保定） | A kind of transmission line of electricity Aerial Images fitting detection method for improving SSD model |
| CN110796186A (en) * | 2019-10-22 | 2020-02-14 | 华中科技大学无锡研究院 | Dry and wet garbage identification and classification method based on improved YOLOv3 network |
| CN111753606A (en) * | 2019-07-04 | 2020-10-09 | 杭州海康威视数字技术股份有限公司 | Intelligent model upgrading method and device |
| CN112364878A (en) * | 2020-09-25 | 2021-02-12 | 江苏师范大学 | Power line classification method based on deep learning under complex background |
| CN113670929A (en) * | 2021-07-05 | 2021-11-19 | 国网宁夏电力有限公司电力科学研究院 | Power transmission line foreign matter detection method and device, storage medium and terminal equipment |
| CN113723181A (en) * | 2021-07-20 | 2021-11-30 | 深圳大学 | Unmanned aerial vehicle aerial photography target detection method and device |
| CN114001833A (en) * | 2021-11-25 | 2022-02-01 | 三门核电有限公司 | Equipment infrared imaging diagnosis method and system based on target detection |
| CN114723678A (en) * | 2022-03-21 | 2022-07-08 | 盛视科技股份有限公司 | High-voltage wire foreign matter detection method and detection system based on video image |
| CN115311584A (en) * | 2022-08-15 | 2022-11-08 | 贵州电网有限责任公司 | Unmanned aerial vehicle high-voltage power grid video inspection floating hanging method based on deep learning |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106909924A (en) * | 2017-02-18 | 2017-06-30 | 北京工业大学 | A kind of remote sensing image method for quickly retrieving based on depth conspicuousness |
| CN106971152A (en) * | 2017-03-16 | 2017-07-21 | 天津大学 | A kind of method of Bird's Nest in detection transmission line of electricity based on Aerial Images |
| CN107392901A (en) * | 2017-07-24 | 2017-11-24 | 国网山东省电力公司信息通信公司 | A kind of method for transmission line part intelligence automatic identification |
| CN107423760A (en) * | 2017-07-21 | 2017-12-01 | 西安电子科技大学 | Based on pre-segmentation and the deep learning object detection method returned |
| CN107563412A (en) * | 2017-08-09 | 2018-01-09 | 浙江大学 | A kind of infrared image power equipment real-time detection method based on deep learning |
-
2018
- 2018-05-16 CN CN201810465955.2A patent/CN108647655B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106909924A (en) * | 2017-02-18 | 2017-06-30 | 北京工业大学 | A kind of remote sensing image method for quickly retrieving based on depth conspicuousness |
| CN106971152A (en) * | 2017-03-16 | 2017-07-21 | 天津大学 | A kind of method of Bird's Nest in detection transmission line of electricity based on Aerial Images |
| CN107423760A (en) * | 2017-07-21 | 2017-12-01 | 西安电子科技大学 | Based on pre-segmentation and the deep learning object detection method returned |
| CN107392901A (en) * | 2017-07-24 | 2017-11-24 | 国网山东省电力公司信息通信公司 | A kind of method for transmission line part intelligence automatic identification |
| CN107563412A (en) * | 2017-08-09 | 2018-01-09 | 浙江大学 | A kind of infrared image power equipment real-time detection method based on deep learning |
Non-Patent Citations (1)
| Title |
|---|
| WEI LIU等: "SSD: Single Shot MultiBox Detector", 《SPRINGER》 * |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109472214A (en) * | 2018-10-17 | 2019-03-15 | 福州大学 | One kind is taken photo by plane foreign matter image real-time detection method based on deep learning |
| CN109389322A (en) * | 2018-10-30 | 2019-02-26 | 福州大学 | The disconnected broken lot recognition methods of grounded-line based on target detection and long memory models in short-term |
| CN109726741A (en) * | 2018-12-06 | 2019-05-07 | 江苏科技大学 | A kind of detection method and device of multiple target object |
| CN109726741B (en) * | 2018-12-06 | 2023-05-30 | 江苏科技大学 | Method and device for detecting multiple target objects |
| CN109902730A (en) * | 2019-02-21 | 2019-06-18 | 国网山东省电力公司临沂供电公司 | Broken strand of power transmission line detection method based on deep learning |
| CN109902730B (en) * | 2019-02-21 | 2021-06-01 | 国网山东省电力公司临沂供电公司 | Power transmission line broken strand detection method based on deep learning |
| CN110163081A (en) * | 2019-04-02 | 2019-08-23 | 宜通世纪物联网研究院(广州)有限公司 | SSD-based real-time regional intrusion detection method, system and storage medium |
| CN110033451A (en) * | 2019-04-17 | 2019-07-19 | 国网山西省电力公司电力科学研究院 | A kind of power components defect inspection method based on SSD framework |
| CN110033453A (en) * | 2019-04-18 | 2019-07-19 | 国网山西省电力公司电力科学研究院 | Based on the power transmission and transformation line insulator Aerial Images fault detection method for improving YOLOv3 |
| CN110033453B (en) * | 2019-04-18 | 2023-02-24 | 国网山西省电力公司电力科学研究院 | Power transmission and transformation line insulator aerial image fault detection method based on improved YOLOv3 |
| CN110070530A (en) * | 2019-04-19 | 2019-07-30 | 山东大学 | A kind of powerline ice-covering detection method based on deep neural network |
| CN110175524A (en) * | 2019-04-26 | 2019-08-27 | 南京航空航天大学 | A kind of quick vehicle checking method of accurately taking photo by plane based on lightweight depth convolutional network |
| CN110232370A (en) * | 2019-06-21 | 2019-09-13 | 华北电力大学（保定） | A kind of transmission line of electricity Aerial Images fitting detection method for improving SSD model |
| CN111753606A (en) * | 2019-07-04 | 2020-10-09 | 杭州海康威视数字技术股份有限公司 | Intelligent model upgrading method and device |
| WO2021000956A1 (en) * | 2019-07-04 | 2021-01-07 | 杭州海康威视数字技术股份有限公司 | Method and apparatus for upgrading intelligent model |
| CN110796186A (en) * | 2019-10-22 | 2020-02-14 | 华中科技大学无锡研究院 | Dry and wet garbage identification and classification method based on improved YOLOv3 network |
| CN112364878A (en) * | 2020-09-25 | 2021-02-12 | 江苏师范大学 | Power line classification method based on deep learning under complex background |
| CN113670929A (en) * | 2021-07-05 | 2021-11-19 | 国网宁夏电力有限公司电力科学研究院 | Power transmission line foreign matter detection method and device, storage medium and terminal equipment |
| CN113723181A (en) * | 2021-07-20 | 2021-11-30 | 深圳大学 | Unmanned aerial vehicle aerial photography target detection method and device |
| CN113723181B (en) * | 2021-07-20 | 2023-10-20 | 深圳大学 | Unmanned aerial vehicle aerial photographing target detection method and device |
| CN114001833A (en) * | 2021-11-25 | 2022-02-01 | 三门核电有限公司 | Equipment infrared imaging diagnosis method and system based on target detection |
| CN114723678A (en) * | 2022-03-21 | 2022-07-08 | 盛视科技股份有限公司 | High-voltage wire foreign matter detection method and detection system based on video image |
| CN115311584A (en) * | 2022-08-15 | 2022-11-08 | 贵州电网有限责任公司 | Unmanned aerial vehicle high-voltage power grid video inspection floating hanging method based on deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108647655B (en) | 2022-07-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108647655A (en) | Low latitude aerial images power line foreign matter detecting method based on light-duty convolutional neural networks | |
| CN110348376B (en) | Pedestrian real-time detection method based on neural network | |
| CN107016405A (en) | A kind of insect image classification method based on classification prediction convolutional neural networks | |
| CN108830188A (en) | Vehicle checking method based on deep learning | |
| CN113065558A (en) | Lightweight small target detection method combined with attention mechanism | |
| CN109800735A (en) | Accurate detection and segmentation method for ship target | |
| CN112446388A (en) | Multi-category vegetable seedling identification method and system based on lightweight two-stage detection model | |
| CN108427920A (en) | A kind of land and sea border defense object detection method based on deep learning | |
| CN108764308A (en) | A kind of recognition methods again of the pedestrian based on convolution loop network | |
| CN108037770A (en) | Unmanned plane power transmission line polling system and method based on artificial intelligence | |
| CN106446930A (en) | Deep convolutional neural network-based robot working scene identification method | |
| CN105825502B (en) | A kind of Weakly supervised method for analyzing image of the dictionary study based on conspicuousness guidance | |
| CN109029363A (en) | A kind of target ranging method based on deep learning | |
| CN109766873B (en) | Pedestrian re-identification method based on hybrid deformable convolution | |
| CN112464911A (en) | Improved YOLOv 3-tiny-based traffic sign detection and identification method | |
| CN113111727B (en) | Feature alignment-based method for detecting rotating target in remote sensing scene | |
| CN110189304A (en) | Remote sensing image target on-line quick detection method based on artificial intelligence | |
| CN110321862B (en) | Pedestrian re-identification method based on compact ternary loss | |
| CN107967474A (en) | A kind of sea-surface target conspicuousness detection method based on convolutional neural networks | |
| Alidoost et al. | Knowledge based 3D building model recognition using convolutional neural networks from LiDAR and aerial imageries | |
| Doi et al. | The effect of focal loss in semantic segmentation of high resolution aerial image | |
| CN108764018A (en) | A kind of multitask vehicle based on convolutional neural networks recognition methods and device again | |
| CN114842208A (en) | Power grid harmful bird species target detection method based on deep learning | |
| CN109583499B (en) | Power transmission line background object classification system based on unsupervised SDAE network | |
| CN111753682A (en) | Hoisting area dynamic monitoring method based on target detection algorithm |
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 |