CN109961070A - The method of mist body concentration is distinguished in a kind of power transmission line intelligent image monitoring - Google Patents
The method of mist body concentration is distinguished in a kind of power transmission line intelligent image monitoring Download PDFInfo
- Publication number
- CN109961070A CN109961070A CN201910222698.4A CN201910222698A CN109961070A CN 109961070 A CN109961070 A CN 109961070A CN 201910222698 A CN201910222698 A CN 201910222698A CN 109961070 A CN109961070 A CN 109961070A
- Authority
- CN
- China
- Prior art keywords
- image
- value
- mist
- mxm
- rxr
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/56—Extraction of image or video features relating to colour
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of methods that mist body concentration is distinguished in power transmission line intelligent image monitoring, it uses combination features method, include step: extracting several features in picture about mist, and the several characteristic synthetics for extracting each image constitute a two dimensional character figure at a characteristic parameter table；It is trained and models using the characteristic parameter figure of the deep learning convolutional neural networks to great amount of samples image；Unknown images are carried out to judge whether there is mist with the model that training obtains, and distinguish the concentration scale of mist: clear, middle mist and thick fog.The present invention rises by the several profound and irrelevant characteristic synthetics of provider as image features, then one non1inear classifying algorithm model is trained to great amount of samples image using convolutional neural networks, solves the problems, such as whether detection image has mist accuracy rate lower.
Description
Technical field
The present invention relates to power monitoring field, and the side of mist body concentration is distinguished in espespecially a kind of power transmission line intelligent image monitoring
Method.
Background technique
Transmission line of electricity distribution in China's is wide, and route is long, with a varied topography.Traditional artificial inspection is only leaned on, will lead to heavy workload, is imitated
Rate is low, and detection effect is bad.
And computer depth learning technology is utilized, intelligent image can be carried out to the transmission line of electricity taken and periphery and known
Not, whether have construction machinery, the object of the potential risks such as Overheight Vehicles, floating material exists, to substantially increase line if analyzing
Road routing inspection efficiency and quality.
Currently, the sharply development of industrial technology brings serious influence to weather, and extreme weather such as mist, haze (below will
Mist, haze are referred to as mist) etc. take place frequently, and transmission line of electricity is distributed in mountain area mostly, it is easy to have a mist to be centered around transmission line of electricity week
It enclosing, this can cause the image quality decrease (the problems such as such as decline of contrast, clarity, color change) taken, so as to cause
Picture material can not identify.
Present industry is very more to the algorithm research of image defogging, is all excellent with the dark of doctor's He Kaiming proposition mostly
It is improved based on first defogging algorithm.But to whether thering is the algorithm research of this Image Feature Detection of mist relatively to lack
It is weary.The basis for whether having mist to be only subsequent defogging in picture material is correctly detected in fact.For example, if given one is secondary
Image is inputted as system, but may have mist by erroneous detection without mist or have mist but missing inspection, such a false judgment in itself
Certainly leading to subsequent processing is also mistake.It can be seen that whether correctly identify in image has the even entire intelligence of mist
Change, the prerequisite of automated image monitoring system successful implementation.So before carrying out Target detection and identification to image, to figure
It is become more and more important as doing the detection of mist and distinguishing the concentration of mist and corresponding defogging is taken to pre-process.
Classify at present to a part research of mist figure detection with machine learning to realize.Specifically, first right
Then image zooming-out feature, such as gray scale, gradient and entropy etc. come using common classifications algorithms such as naive Bayesian, SVM to big
These characteristic values in amount sample are learnt and are modeled.The model finally learnt with this has the picture newly inputted
Mist or fogless classification, thus detection whether reaching mist figure.
Such as a kind of entitled method of the mist grade based on image recognition transmission line of electricity of publication number CN103903008A with
And the patent of system, multiple training using acquisition transmission line of electricity in fine day, mist, mist, dense fog, thick fog, strong thick fog weather are schemed
Picture；Extract respectively fine day, mist, mist, dense fog, thick fog, the corresponding image class of strong thick fog feature, main includes the maximum of image
The features such as gray-scale intensity, contrast, saturation degree；It is instructed the feature of described image class as the input data of naive Bayesian
Practice, obtains mist grade recognizer model；Acquire the images to be recognized of transmission line of electricity；It is corresponding to extract the images to be recognized
Feature；Trained algorithm model identifies the corresponding feature of the images to be recognized before use, obtains mist
Grade recognition result；
Having mist grade separation with machine learning method progress image, there are two significant challenges, and first is Feature Selection.No
Whether being only required to feature itself really can express image and have mist, also to guarantee low correlation between feature.And above-identified patent
Correlation is relatively high between the image brightness properties and contrast metric.Second is when samples pictures are very more or true
In the real world between object when Relationship Comparison complexity, it is simple using the conventional machines learning algorithms such as Bayes or SVM be not enough into
The effective modeling of row, it is not high so as to cause classification accuracy.
It in addition is that deeper relevant with mist time in image is innovatively excavated by many experiments there are also some researchs
Feature, such as dark, tone otherness etc.., but be all individually separately research, generally, exactly piece image is calculated
Then some characteristic value comes whether simple zones partial image has mist using preset threshold value.
Summary of the invention
To solve the above problems, present invention is primarily aimed at provide a kind of method of combination features detection, specifically
It says, first excavates several profound and not strong correlation features relevant with mist, and integrate and constitute two dimensional character parameter
Figure, is then trained using characteristic parameter figure of the depth convolutional neural networks InceptionV3 to great amount of samples image, is obtained
One image classification algorithms model.It is then based on it, judges whether clear new image or middle mist or thick fog.
To achieve the above object, the present invention provides mist body concentration is distinguished in a kind of power transmission line intelligent image monitoring
Method uses combination features method, and it includes steps:
1) several features in picture about mist are extracted, and several characteristic synthetics that each image is extracted are at a spy
Parameter list is levied, a two dimensional character figure is formed；
2) it is trained using the characteristic parameter figure of the deep learning convolutional neural networks algorithm to great amount of samples image
And modeling；
3) unknown images are carried out to judge whether there is mist with the model obtained, and distinguishes the concentration scale of mist: clear, middle mist
And thick fog.
Wherein preferably, not having correlation between several features.Several features include dark feature, maximum
Contrast metric, color decay characteristics and tone otherness feature.
Wherein preferably, the dark characteristic formula is as follows:
Wherein: IC(y) be value of the image on tri- channels RGB on y-coordinate point, before plus min be take in RGB it is minimum
That value.
Ωr(x) be by image centered on x coordinate, the rectangular image block of the wide a height of rxr of x；Outmost min is to ask
Minimum value inside whole image block, finds, the seldom image of mist, pixel value very little at least one channel through many experiments,
Close to 0.
If a sub-picture is divided into mxm rxr image block, will acquire a length by formula above is
The dark feature vector of mxm.
Wherein preferably, the maximum-contrast characteristic formula is expressed as follows:
Wherein, Ωs(y) it is small template centered on y, such as 3x3 or 5x5 etc.Above in formula, radical sign part is
The quadratic sum of other pixel values and y value difference in first seeking template, then obtain mean value divided by number of pixels, finally evolution again；It is right respectively
Image block Ωr(x) each position calculates such a value, then asks one of maximum value, i.e. maximum-contrast with max
Characteristic value；Wherein mist is bigger in image, and the value is also bigger；
Similarly, it if a sub-picture to be divided into the image block of mxm rxr, can be calculated by formula above
The maximum-contrast feature vector that a length is mxm out.
Wherein preferably, described its formula of color decay characteristics is as follows:
A (x)=Iυ(x)-Is(x)
Image is first transformed into HSV space from rgb space first, wherein Iυ(x) maximum brightness in some image block is represented
It is worth (i.e. V value), and Is(x) full color saturation value in the image block: i.e. S value is then represented, the difference between them just characterizes should
The color decay characteristics of image block, mist is bigger in image, and brightness will increase, and color saturation just sharply declines, and corresponds to
Difference, i.e. color decay characteristics value will be bigger；
Piece image is divided into the image block of mxm rxr, can calculate a length according to formula above is mxm's
Feature vector.
Wherein preferably, the tone otherness characteristic formula is as follows:
Wherein, IsiIt (x) is semi-inverted image, value is equal to max [Ic(x), 255-Ic(x)], c characterizes R, G, channel B, so
Original image and semi-inverted image are transformed into HSV from the domain RGB afterwards, and take the channel H, is obtainedAnd Ih(x), finally ask poor
And thoroughly deserve tone otherness characteristic value；For rxr image block, the absolute value that each pixel corresponds to difference is summed
And average tone difference characteristic value is obtained divided by rxr.Image mist is bigger, and tone difference characteristic value is relatively small；
Piece image is divided into mxm rxr image block, then calculates a tone difference characteristic for rxr image block
Value, such entire image will calculate the feature vector that a length is mxm.
Wherein preferably, the convolutional neural networks are using depth convolutional neural networks as classification method when modeling
Inception-V3 network structure has four parallel routes in Inception block, and preceding three-line is distinguished using window size
It is 1 × 1,3 × 3 and 5 × 5 convolutional layer to extract the information under different spaces size.Wherein intermediate two routes can be to input
1 × 1 convolution is first done to reduce input channel number, to reduce model complexity.Article 4 route then uses 3 × 3 maximum pond layers,
1 × 1 convolutional layer is followed by change port number.Four routes, which all employ suitable filling, to be come so that input and output height and width one
It causes.The output of every route is linked in the dimension of channel finally, and is input in next layer.
The beneficial effects of the invention are that by above-mentioned technical proposal, it can be achieved that the image with mist is accurately identified, to figure
As the concentration classification accuracy with mist may be up to 98% or so.
Detailed description of the invention
The Inception V3 block structure schematic diagram that Fig. 1 one specific embodiment of the present invention uses；
Fig. 2 present invention clear, middle mist of one specific embodiment and the template image of thick fog；
The dark feature flow chart of Fig. 3 specific embodiment of the invention；
The maximum-contrast feature flow chart of Fig. 4 specific embodiment of the invention；
The color decay characteristics flow chart of Fig. 5 specific embodiment of the invention；
The tone difference feature stream journey figure of Fig. 6 specific embodiment of the invention.
Specific embodiment
Below with reference to the embodiments and with reference to the accompanying drawing technical solution of the present invention is described in further detail.
Based on the above issues, combination features detect whether mist in power transmission line intelligent image monitoring of the invention
Method is to carry out the classification of clear, middle mist and thick fog to image using the convolutional neural networks of deep learning.
Usually, classical deep learning does not need to carry out image feature extraction, but uses CNN automatically right
Image carries out feature extraction and study, and classifies in the result that last full articulamentum extracts preceding features and learns.But
Inventor through experiments, it was found that, there is the relevant profound feature of mist to be difficult directly to extract by CNN with image, this will lead to point
Class accuracy is relatively low.Therefore the present invention proposes a new scheme by largely analysis experiment: first having extracted four manually
Profound and not strong correlation feature, and it is comprehensive at two dimensional image characteristic parameter figure.Then using convolutional neural networks to big
The corresponding image features figure of amount sample image trains an image classification algorithms model.After tested, the model is to image
Whether there is the Detection accuracy of mist to reach more than 98%, has reached practical commercially available level.
The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention..
In the specific embodiment of the invention, feature extraction, each feature packet first are carried out to each width sample image manually
Feature containing dark, maximum-contrast feature, color decay characteristics and tone otherness feature etc. do not have correlation mutually
Feature.
Respectively process is as follows for its:
1) dark feature
Dark characteristic formula of the invention is as follows:
Wherein: IC(y) be value of the image on tri- channels RGB on y-coordinate point, before plus min be take in RGB it is minimum
That value.
Ωr(X) be by image centered on x coordinate, the rectangular image block of the wide a height of rxr of x；Outmost min is to ask
Minimum value inside whole image block, finds, the seldom image of mist, pixel value very little at least one channel through many experiments,
Close to 0.
If a sub-picture is divided into mxm rxr image block, will acquire a length by formula above is
The dark feature vector of mxm.
It is shown in Figure 3, it is the dark feature flow chart of the specific embodiment of the invention.It implements step:
S11 divides the image into the image block that rxr size is mxm；
Each pixel on S12 poll mxm image block obtains tri- channel minimum values of its R, G, B by comparing, generates
Mxm minimum value；
Then S13 is compared this mxm minimum value again, obtain a minimum value.Here it is mxm image blocks to deserved
Dark characteristic value；
S14 is final, and entire image will obtain the feature vector of rxr dark characteristic value composition.
2) maximum-contrast feature
Maximum-contrast characteristic formula is expressed as follows:
Wherein, Ωs(y) it is small template centered on y, such as 3x3 or 5x5 etc.Above in formula, radical sign part is
The quadratic sum of other pixel values and y value difference in first seeking template, then obtain mean value divided by number of pixels, finally evolution again；It is right respectively
Image block Ωr(x) each position calculates such a value, then asks one of maximum value, i.e. maximum-contrast with max
Characteristic value；Wherein mist is bigger in image, and the value is also bigger；
Similarly, it if a sub-picture to be divided into the image block of mxm rxr, can be calculated by formula above
The maximum-contrast feature vector that a length is mxm out.
As shown in figure 4, being the maximum-contrast feature flow chart of the specific embodiment of the invention.
Implementation step:
Image is transformed into gray scale domain from the domain RGB by S21；
It is the image block of mxm that gray image is divided into rxr size by S22；
S23 calculates periphery 3x3-1 or 5x5-1 adjacent pixel and center pixel centered on pixel each on mxm image block
The quadratic sum of the difference of gray value, then radical sign is opened again divided by number of pixels, obtain contrast metric value；
Then S24 is compared this mxm contrast metric value, acquire a maximum value.Here it is mxm image blocks pair
The maximum-contrast characteristic value answered；
S25 is final, and entire image will obtain rxr maximum-contrast characteristic value and be configured feature vector.
3) color decay characteristics
The color decay characteristics formula is as follows:
A (x)=Iυ(x)-Is(x)
Here, image is first transformed into HSV space from rgb space first, wherein Iυ(x) it represents maximum in some image block
Brightness value (i.e. V value), and Is(x) then represent full color saturation value in the image block: i.e. S value, the difference between them is with regard to table
The color decay characteristics of the image block are levied, mist is bigger in image, and brightness will increase, and color saturation just sharply declines,
Corresponding difference, i.e. color decay characteristics value will be bigger；
Piece image is divided into the image block of mxm rxr, can calculate a length according to formula above is mxm's
Feature vector.
As shown in figure 5, being the color decay characteristics flow chart of the specific embodiment of the invention.
Implementation step:
Image is transformed into the domain HSV from the domain RGB by S31；
The image block that HSV image segmentation is mxm at rxr size by S32；
The brightness V value of each pixel in S33 poll mxm image block, and compare to obtain maximum V；
The saturation degree S value of each pixel in S34 poll mxm image block, and compare to obtain maximum S.Maximum V value is subtracted
Maximum S value is exactly the corresponding color decay characteristics value of the image block；
S35 is final, and entire image will obtain rxr color decay characteristics value and be configured feature vector；
4) tone otherness feature
The tone otherness characteristic formula is as follows:
Wherein, IsiIt (x) is semi-inverted image, value is equal to max [Ic(x), 255-Ic(x)], c characterizes R, G, channel B, so
Original image and semi-inverted image are transformed into HSV from the domain RGB afterwards, and take the channel H, is obtainedAnd Ih(x), finally ask poor
And thoroughly deserve tone otherness characteristic value；For rxr image block, the absolute value that each pixel corresponds to difference is summed
And average tone difference characteristic value is obtained divided by rxr.Image mist is bigger, and tone difference characteristic value is relatively small；
Piece image is divided into mxm rxr image block, then calculates a tone difference characteristic for rxr image block
Value, such entire image will calculate the feature vector that a length is mxm.
As shown in fig. 6, being the tone difference feature stream journey figure of the specific embodiment of the invention.
Comprising steps of
S41 replicating original RGB image, acquires semi-inverted image；
Original image and semi-inverted image are transformed into HSV from the domain RGB by S42；
It is the image block of mxm that two HSV images are all divided into rxr size by S43；
S44 subtracts each other the H value of their corresponding each pixels of mxm image block, and seeks absolute value, and then absolute value is asked
The image block tone otherness characteristic value is obtained with and divided by mxm；
S45 is final, and entire image will obtain rxr tone otherness characteristic value and be configured feature vector.
In this embodiment, comprehensive four features recited above, it is special that each image can extract a two dimension
Parameter Map is levied, with RGB image 960x960, for tile size is 96x96, parameter characterization is as follows.
Then, it is instructed using the characteristic parameter figure of the deep learning convolutional neural networks algorithm to great amount of samples image
Practice and models.
In this embodiment, since deep learning is driven based on big data, a large amount of data are needed, in order to
Cost is reduced, uses transfer learning as sample training Learning Scheme in this specific embodiment, that is, is based on public image data set
The model of pre-training trains our oneself a small amount of sample data.
Using depth convolutional neural networks as classification method when the present invention models.Convolutional neural networks (CNN) are to be based on
The depth machine learning method of artificial neural network has been widely used for the computer vision fields such as image classification, identification, and
Obtain good effect.Compared to conventional machines learning method, such as SVM, CNN can be established more complicated using deep layer network structure
Model is come while simulating real world, moreover it is possible to keep good generalization ability.
Classical CNN has Alexnet, VGG etc..The Inception-V3 net released in this embodiment using Google
Network structure.
As shown in Figure 1, it is the Inception block structural diagram that the present invention uses, it can be seen from this figure that Inception
There are four parallel routes in block.Preceding three-line is taken out using window size is 1 × 1,3 × 3 and 5 × 5 convolutional layer respectively
Take the information under different spaces size.Wherein intermediate two routes can first do 1 × 1 convolution to input to reduce input channel number,
To reduce model complexity.Article 4 route then uses 3 × 3 maximum pond layers, is followed by 1 × 1 convolutional layer to change port number.Four
Route, which all employs suitable filling, to be come so that input and output height and width are consistent.Finally we are by the output of every route logical
Link in road dimension, and is input in next layer.
Inception V3 is Inception third generation version, which on the basis of two generations, introduces in front
Larger two-dimensional convolution is splitted into two smaller one-dimensional convolution by Factorization into small convolutions thought,
To save quantity of parameters, accelerate operation, mitigate over-fitting, increases by one layer of non-linear, extended model ability to express.Meanwhile it is non-
Symmetrical convolutional coding structure is split, and is become apparent from than symmetrically splitting identical small convolution kernel effect, can be handled more, richer space characteristics,
And increase characteristic polymorphic.To sum up, Inception V3 is the ideal convolutional neural networks structure for realizing image classification of the present invention.
In the present embodiment, acquire the picture of 3000 transmission lines of electricity, wherein clearly, middle mist and thick fog each 1000,
It is respectively present three respective file catalogues.Their template image is as shown in Figure 2.Specifically, the definition of clear image is several
It can't see mist, content is all seen clearer；The definition of middle mist is that the inside part has mist, at least 50% or more content.
After samples pictures are collected, two dimensional character Parameter Map is extracted respectively to each image, equally store by class, then make
It is trained with Inception-V3 convolutional neural networks frame, starts load and be directed to number on public data collection imagenet in advance
The trained weight model of million pictures, then trains ginsengs several layers of below for 3000 samples pictures of collection in worksite
Number, and classification number is generated 3, wherein 0 represents clear class, mist class in 1 representative, 2 represent thick fog class.By 5000 iteration
Training obtains sorting algorithm model.
Finally, being carried out judging whether there is mist to unknown images with model achieved above, and distinguish the concentration scale of mist: clear
Clear, middle mist and thick fog.
By test, the accuracy classified to new picture reaches 98% or more.Through the experimental result such as following table institute
Show:
| Images to be recognized | Picture number | Identify success count | Accuracy rate |
| Clearly | 100 | 99 | 99% |
| Middle mist | 90 | 88 | 97.78% |
| Thick fog | 150 | 148 | 98.67% |
| Total image recognition number | 340 | Average Accuracy | 98.48% |
The above embodiments are merely illustrative of the technical scheme of the present invention and are not intended to be limiting thereof, although referring to above-described embodiment pair
The present invention is described in detail, it should be understood by a person of ordinary skill in the art that still can be to of the invention specific
Embodiment is modified or replaced equivalently, and without departing from any modification of spirit and scope of the invention or equivalent replacement,
It is intended to be within the scope of the claims of the invention.
Claims (10)
1. a kind of method that mist body concentration is distinguished in power transmission line intelligent image monitoring, uses combination features method,
It is characterized in that, comprising the steps of:
1) several features in picture about mist are extracted, and several characteristic synthetics that each image is extracted are joined at a feature
Number table, forms a two dimensional character figure；
2) it is trained and builds using the characteristic parameter figure of the deep learning convolutional neural networks algorithm to great amount of samples image
Mould；
3) unknown images are carried out to judge whether there is mist with the model obtained, and distinguish the concentration scale of mist: clear, middle mist and
Thick fog.
2. the method for distinguishing mist body concentration in a kind of power transmission line intelligent image monitoring according to claim 1, special
Sign is that several features include that dark feature, maximum-contrast feature, color decay characteristics and tone otherness are special
Sign.
3. the method for distinguishing mist body concentration in a kind of power transmission line intelligent image monitoring according to claim 1, special
Sign is do not have correlation between several features.
4. the method for distinguishing mist body concentration in a kind of power transmission line intelligent image monitoring according to claim 2, special
Sign is that the dark characteristic formula is as follows:
Wherein: IC(y) be value of the image on tri- channels RGB on y-coordinate point, before plus min be take in RGB it is the smallest that
A value.
Ωr(x) be by image centered on x coordinate, the rectangular image block of the wide a height of rxr of x；Outmost min is to seek entire figure
As minimum value inside block；
If a sub-picture is divided into mxm rxr image block, will acquire a length by formula above is mxm's
Dark feature vector.
5. the method for distinguishing mist body concentration in a kind of power transmission line intelligent image monitoring according to claim 2, special
Sign is, the maximum-contrast feature:
Its formula is expressed as follows:
Wherein, Ωs(y) it is small template centered on y, such as 3x3 or 5x5 etc.Above in formula, radical sign part is first to ask
The quadratic sum of other pixel values and y value difference in template, then obtain mean value divided by number of pixels, finally evolution again；Respectively to image
Block Ωr(x) each position calculates such a value, then asks one of maximum value, i.e. maximum-contrast feature with max
Value；Wherein mist is bigger in image, and the value is also bigger；
Similarly, if a sub-picture to be divided into the image block of mxm rxr, one can be calculated by formula above
A length is the maximum-contrast feature vector of mxm.
6. the method for distinguishing mist body concentration in a kind of power transmission line intelligent image monitoring according to claim 2, special
Sign is, described its formula of color decay characteristics is as follows:
A (x)=Iυ(x)-Is(x)
Image is first transformed into HSV space from rgb space first, wherein Iυ(x) maximum brightness value (i.e. V in some image block is represented
Value), and Is(x) full color saturation value in the image block: i.e. S value is then represented, the difference between them just characterizes the image block
Color decay characteristics, mist is bigger in image, and brightness will increase, and color saturation just sharply declines, and corresponds to difference, i.e.,
Color decay characteristics value will be bigger；
Piece image is divided into the image block of mxm rxr, the feature that a length is mxm can be calculated according to formula above
Vector.
7. the method for distinguishing mist body concentration in a kind of power transmission line intelligent image monitoring according to claim 2, special
Sign is, the tone otherness feature:
Its formula is as follows:
Wherein, IsiIt (x) is semi-inverted image, value is equal to max [Ic(x), 255-Ic(x)], c characterizes R, G, channel B, then will
Original image and semi-inverted image are transformed into HSV from the domain RGB, and take the channel H, obtainAnd Ih(x), finally ask poor and exhausted
Tone otherness characteristic value is obtained to value；For rxr image block, the absolute value that each pixel corresponds to difference is summed and removed
Average tone difference characteristic value is obtained with rxr.Image mist is bigger, and tone difference characteristic value is relatively small；
Piece image is divided into mxm rxr image block, then calculates a tone difference characteristic value for rxr image block, this
Sample entire image will calculate the feature vector that a length is mxm.
8. the method for distinguishing mist body concentration in a kind of power transmission line intelligent image monitoring according to claim 1, special
Sign is, using depth convolutional neural networks as classification method when modeling, the convolutional neural networks are Inception-V3
Network structure has four parallel routes in Inception block, and preceding three-line is 1 × 1,3 × 3 respectively using window size
Convolutional layer with 5 × 5 extracts the information under different spaces size.Wherein intermediate two routes can first do 1 × 1 convolution to input
Input channel number is reduced, to reduce model complexity.Article 4 route then uses 3 × 3 maximum pond layers, is followed by 1 × 1 convolution
Layer changes port number.Four routes, which all employ suitable filling, to be come so that input and output height and width are consistent.Finally by every
The output of route links on channel is tieed up, and is input in next layer.
9. the method for distinguishing mist body concentration in a kind of power transmission line intelligent image monitoring according to claim 4, special
Sign is that the dark characteristic value implements step:
S11 divides the image into the image block that rxr size is mxm；
Each pixel on S12 poll mxm image block obtains tri- channel minimum values of its R, G, B by comparing, generates mxm
A minimum value；
Then S13 is compared this mxm minimum value again, obtain a minimum value.Here it is mxm image blocks to help secretly to deserved
Road characteristic value；
S14 is final, and entire image will obtain rxr dark characteristic value and be configured feature vector.
10. the method for distinguishing mist body concentration in a kind of power transmission line intelligent image monitoring according to claim 5, special
Sign is that maximum-contrast feature realizes step:
Image is transformed into gray scale domain from the domain RGB by S21；
It is the image block of mxm that gray image is divided into rxr size by S22；
S23 calculates periphery 3x3-1 or 5x5-1 adjacent pixel and center pixel gray scale centered on pixel each on mxm image block
The quadratic sum of the difference of value, then radical sign is opened again divided by number of pixels, obtain contrast metric value；
Then S24 is compared this mxm contrast metric value, acquire a maximum value.Here it is mxm image block is corresponding
Maximum-contrast characteristic value；
S25 is final, and entire image will obtain the feature vector of rxr maximum-contrast characteristic value composition.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910222698.4A CN109961070A (en) | 2019-03-22 | 2019-03-22 | The method of mist body concentration is distinguished in a kind of power transmission line intelligent image monitoring |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910222698.4A CN109961070A (en) | 2019-03-22 | 2019-03-22 | The method of mist body concentration is distinguished in a kind of power transmission line intelligent image monitoring |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109961070A true CN109961070A (en) | 2019-07-02 |
Family
ID=67024656
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910222698.4A Pending CN109961070A (en) | 2019-03-22 | 2019-03-22 | The method of mist body concentration is distinguished in a kind of power transmission line intelligent image monitoring |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109961070A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110705619A (en) * | 2019-09-25 | 2020-01-17 | 南方电网科学研究院有限责任公司 | Fog concentration grade judging method and device |
| CN110930326A (en) * | 2019-11-15 | 2020-03-27 | 浙江大华技术股份有限公司 | Image and video defogging method and related device |
| CN112419745A (en) * | 2020-10-20 | 2021-02-26 | 中电鸿信信息科技有限公司 | Highway group fog early warning system based on degree of depth fusion network |
| CN112686105A (en) * | 2020-12-18 | 2021-04-20 | 云南省交通规划设计研究院有限公司 | Fog concentration grade identification method based on video image multi-feature fusion |
| CN112818886A (en) * | 2021-02-09 | 2021-05-18 | 广州富港万嘉智能科技有限公司 | Flying dust detection method, readable storage medium, flying dust detection machine and intelligent food machine |
| WO2021228088A1 (en) * | 2020-05-11 | 2021-11-18 | 南京邮电大学 | Method for recognizing haze concentration in haze image |
| WO2022012149A1 (en) * | 2020-07-17 | 2022-01-20 | 上海商汤智能科技有限公司 | Agglomerate fog recognition method and apparatus, electronic device, storage medium, and computer program product |
| CN114973110A (en) * | 2022-07-27 | 2022-08-30 | 四川九通智路科技有限公司 | On-line monitoring method and system for highway weather |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103442209A (en) * | 2013-08-20 | 2013-12-11 | 北京工业大学 | Video monitoring method of electric transmission line |
| CN103903008A (en) * | 2014-03-26 | 2014-07-02 | 国家电网公司 | Power transmission line fog level recognition method and system based on images |
| CN203825644U (en) * | 2014-03-26 | 2014-09-10 | 国家电网公司 | Image identification power transmission line-based fog level system |
| CN106779054A (en) * | 2016-12-31 | 2017-05-31 | 中国科学技术大学 | A kind of PM2.5 methods of estimation based on Misty Image |
| CN107909084A (en) * | 2017-11-15 | 2018-04-13 | 电子科技大学 | A kind of haze concentration prediction method based on convolution linear regression network |
| CN109237582A (en) * | 2018-11-15 | 2019-01-18 | 珠海格力电器股份有限公司 | Range hood control method and system based on image recognition and range hood |
| CN109255758A (en) * | 2018-07-13 | 2019-01-22 | 杭州电子科技大学 | Image enchancing method based on full 1*1 convolutional neural networks |
| CN109389569A (en) * | 2018-10-26 | 2019-02-26 | 大象智能科技(南京)有限公司 | Based on the real-time defogging method of monitor video for improving DehazeNet |
-
2019
- 2019-03-22 CN CN201910222698.4A patent/CN109961070A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103442209A (en) * | 2013-08-20 | 2013-12-11 | 北京工业大学 | Video monitoring method of electric transmission line |
| CN103903008A (en) * | 2014-03-26 | 2014-07-02 | 国家电网公司 | Power transmission line fog level recognition method and system based on images |
| CN203825644U (en) * | 2014-03-26 | 2014-09-10 | 国家电网公司 | Image identification power transmission line-based fog level system |
| CN106779054A (en) * | 2016-12-31 | 2017-05-31 | 中国科学技术大学 | A kind of PM2.5 methods of estimation based on Misty Image |
| CN107909084A (en) * | 2017-11-15 | 2018-04-13 | 电子科技大学 | A kind of haze concentration prediction method based on convolution linear regression network |
| CN109255758A (en) * | 2018-07-13 | 2019-01-22 | 杭州电子科技大学 | Image enchancing method based on full 1*1 convolutional neural networks |
| CN109389569A (en) * | 2018-10-26 | 2019-02-26 | 大象智能科技(南京)有限公司 | Based on the real-time defogging method of monitor video for improving DehazeNet |
| CN109237582A (en) * | 2018-11-15 | 2019-01-18 | 珠海格力电器股份有限公司 | Range hood control method and system based on image recognition and range hood |
Non-Patent Citations (4)
| Title |
|---|
| 中国气象局预测减灾司: "《气象部门拓展业务服务领域文集》", 31 March 2005, 气象出版社 * |
| 张敏华: "基于卷积神经网络的去雾算法优化研究与实现", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 许艳丽: "雾霾天气条件下能见度的检测与恢复算法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 黄伟政: "基于卷积神经网络的雾霾时空演化预测方法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110705619A (en) * | 2019-09-25 | 2020-01-17 | 南方电网科学研究院有限责任公司 | Fog concentration grade judging method and device |
| CN110930326A (en) * | 2019-11-15 | 2020-03-27 | 浙江大华技术股份有限公司 | Image and video defogging method and related device |
| US20220076168A1 (en) * | 2020-05-11 | 2022-03-10 | Nanjing University Of Posts And Telecommunications | Method for recognizing fog concentration of hazy image |
| US11775875B2 (en) * | 2020-05-11 | 2023-10-03 | Nanjing University Of Posts And Telecommunications | Method for recognizing fog concentration of hazy image |
| WO2021228088A1 (en) * | 2020-05-11 | 2021-11-18 | 南京邮电大学 | Method for recognizing haze concentration in haze image |
| JP2022545962A (en) * | 2020-07-17 | 2022-11-01 | シャンハイ センスタイム インテリジェント テクノロジー カンパニー リミテッド | Fog Recognition Method and Apparatus, Electronic Device, Storage Medium and Computer Program Product |
| WO2022012149A1 (en) * | 2020-07-17 | 2022-01-20 | 上海商汤智能科技有限公司 | Agglomerate fog recognition method and apparatus, electronic device, storage medium, and computer program product |
| CN112419745A (en) * | 2020-10-20 | 2021-02-26 | 中电鸿信信息科技有限公司 | Highway group fog early warning system based on degree of depth fusion network |
| CN112686105A (en) * | 2020-12-18 | 2021-04-20 | 云南省交通规划设计研究院有限公司 | Fog concentration grade identification method based on video image multi-feature fusion |
| CN112686105B (en) * | 2020-12-18 | 2021-11-02 | 云南省交通规划设计研究院有限公司 | Fog concentration grade identification method based on video image multi-feature fusion |
| CN112818886A (en) * | 2021-02-09 | 2021-05-18 | 广州富港万嘉智能科技有限公司 | Flying dust detection method, readable storage medium, flying dust detection machine and intelligent food machine |
| CN114973110A (en) * | 2022-07-27 | 2022-08-30 | 四川九通智路科技有限公司 | On-line monitoring method and system for highway weather |
| CN114973110B (en) * | 2022-07-27 | 2022-11-01 | 四川九通智路科技有限公司 | On-line monitoring method and system for highway weather |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109961070A (en) | The method of mist body concentration is distinguished in a kind of power transmission line intelligent image monitoring | |
| CN109800824B (en) | Pipeline defect identification method based on computer vision and machine learning | |
| CN111401372B (en) | Method for extracting and identifying image-text information of scanned document | |
| CN106910186B (en) | Bridge crack detection and positioning method based on CNN deep learning | |
| CN110852316B (en) | Image tampering detection and positioning method adopting convolution network with dense structure | |
| CN111784633B (en) | Insulator defect automatic detection algorithm for electric power inspection video | |
| CN106934386B (en) | A kind of natural scene character detecting method and system based on from heuristic strategies | |
| CN104077577A (en) | Trademark detection method based on convolutional neural network | |
| CN107945153A (en) | A kind of road surface crack detection method based on deep learning | |
| CN108181316B (en) | Bamboo strip defect detection method based on machine vision | |
| CN109753890A (en) | A kind of pavement garbage object intelligent recognition and cognitive method and its realization device | |
| CN104463196A (en) | Video-based weather phenomenon recognition method | |
| CN104063713B (en) | A kind of semi-autonomous on-line study method based on random fern grader | |
| CN110298297A (en) | Flame identification method and device | |
| CN109598681B (en) | No-reference quality evaluation method for image after repairing of symmetrical Thangka | |
| CN105550710A (en) | Nonlinear fitting based intelligent detection method for running exception state of contact network | |
| CN115294377A (en) | System and method for identifying road cracks | |
| CN116597438A (en) | Improved fruit identification method and system based on Yolov5 | |
| CN103279960B (en) | A kind of image partition method of human body cache based on X-ray backscatter images | |
| CN112800968B (en) | HOG blocking-based feature histogram fusion method for identifying identity of pigs in drinking area | |
| CN109741351A (en) | A kind of classification responsive type edge detection method based on deep learning | |
| CN111310899B (en) | Power defect identification method based on symbiotic relation and small sample learning | |
| CN106446832B (en) | Video-based pedestrian real-time detection method | |
| CN117058534A (en) | Small sample remote sensing image target detection method based on meta-knowledge adaptive migration network | |
| CN107545565A (en) | A kind of solar energy half tone detection method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190702 |
|
| RJ01 | Rejection of invention patent application after publication |