CN109544545A - A kind of salt mine intelligent detecting method and system based on convolutional neural networks - Google Patents
A kind of salt mine intelligent detecting method and system based on convolutional neural networks Download PDFInfo
- Publication number
- CN109544545A CN109544545A CN201811449637.3A CN201811449637A CN109544545A CN 109544545 A CN109544545 A CN 109544545A CN 201811449637 A CN201811449637 A CN 201811449637A CN 109544545 A CN109544545 A CN 109544545A
- Authority
- CN
- China
- Prior art keywords
- vibration wave
- salt mine
- exposure mask
- unet model
- training
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30181—Earth observation
- G06T2207/30184—Infrastructure
Abstract
The present invention discloses a kind of salt mine intelligent detecting method based on convolutional neural networks, is related to technical field of image segmentation, including training part and detection part.Training part needs to collect vibration wave figure, Expert opinion and the salt mine for marking vibration wave figure, then setting resolution ratio carries out the segmentation of vibration wave figure, divided vibration wave figure is as training set, UNet model is built on deep learning platform, convolutional layer is introduced in the same level of UNet model, training set introducing UNet model is learnt, new UNet model is obtained.In detection part, vibration wave figure to be detected is introduced into new UNet model, given threshold divides vibration wave figure to be detected and obtains exposure mask figure, exposure mask figure is transformed into binaryzation exposure mask figure, high definition exposure mask figure is obtained after splicing all binaryzation exposure mask figures, salt mine threshold value is set, compares the ratio that salt mine area in high definition exposure mask figure glues the gross area, finds whether find salt mine in vibration wave figure.This method is used for the vibration wave figure of the big resolution ratio of intelligent measurement large scale.The present invention also provides a kind of salt mine intelligent checking systems.
Description
Technical field
The present invention relates to geologic image segmentation technology, specifically a kind of salt mine intelligence based on convolutional neural networks
It can detection method and system.
Background technique
Underground structure can be imaged using the difference of vibration wave spread speed in different media.The image of formation
Generally grayscale image, and the edge of different medium can be than more visible, but same media imaging is fuzzy, such as one left figure institute of attached drawing
Show.It is often used vibration wave imaging technique in engineering and analytically descends composition of ores, such as detection salt mine.Propagation of the vibration wave in salt mine
The more general rock of speed is fast, and on the boundary of salt mine and other lithologic subsurface structures, vibration wave image will form sharp boundary line.
Veteran geology detecting personnel can judge whether it is salt mine by vibration wave image.But in actual detection, a piece of big
It will form a large amount of vibration wave images in region, vibration wave figure in every image description different depth and finite region is completely dependent on spy
Survey personnel are differentiated time and effort consuming to all images one by one, and are easy error.
Previous salt mine detection remains a large amount of vibration wave figures, and has already been through professional and analyze, because having accumulated
A large amount of labeled data.Such as in attached drawing one, a figure left side is vibration wave figure, and the figure right side is generated according to the differentiation or practical exploitation situation of expert
Exposure mask figure, salt mine part indicates that non-salt mine part is indicated with black with white in figure.Utilize these labeled data, Ke Yixun
Practice neural network model to classify to newly generated vibration wave figure automatically.Specifically, which is in vibration wave figure
Whether each pixel is that salt mine is classified, and is typical image segmentation problem.
UNet model is a kind of convolutional neural networks model being specially split to biologic medical image.Such as 2 institute of attached drawing
Show, which is spliced by different convolution blocks, carries out down-sampling (on the left of model between different blocks or by maximum pondization
Part), or up-sampled (model right part) by transposition convolution, each piece of inside is connected by convolutional layer extracts image spy
Sign.It is worth noting that, there is also input relationships between the block in same level level, as shown in Fig. 3, block B ' input
The characteristic pattern and stacked with the characteristic pattern of level-one block B output that characteristic pattern is exported by preceding layer.
Summary of the invention
The present invention is directed to the demand and shortcoming of current technology development, provides a kind of salt mine based on convolutional neural networks
Intelligent detecting method and system.
A kind of salt mine intelligent detecting method based on convolutional neural networks of the invention solves what above-mentioned technical problem used
Technical solution is as follows:
A kind of salt mine intelligent detecting method based on convolutional neural networks, the detection method include:
1) training part:
Vibration wave figure 1a) is collected, expert determines and marks to the salt mine in vibration wave figure；
1b) divide vibration wave figure according to the resolution ratio of setting, using vibration wave figures all after segmentation as training set；
UNet model 1c) is built on deep learning platform, introduces convolutional layer, setting in the same level of UNet model
UNet model training stop parameter, by training set introducing UNet model learn, obtain can to salt mine in vibration wave figure into
The new UNet model of row segmentation；
2) detection part:
New UNet model 2a) is deployed to computer equipment；
Vibration wave figure to be detected is introduced new UNet model and is divided into multiple vibration wave figure units, owned by 2b) given threshold
The exposure mask figure of vibration wave figure unit, and according to the threshold value of setting, exposure mask figure is transformed into binaryzation exposure mask figure；
High definition exposure mask figure is obtained after 2c) splicing all binaryzation exposure mask figures, salt mine threshold value is set, compares in high definition exposure mask figure
Salt mine area glues the ratio of the gross area, to find whether find salt mine in vibration wave figure.
In step 1a), after expert determines vibration wave figure, the part that salt mine is judged as in vibration wave figure is labeled as
Pixel 1, remaining is labeled as pixel 0.
In step 2b), in exposure mask figure more than given threshold part output pixel value be 1, otherwise output pixel value be
0, it can be realized and exposure mask figure is transformed into binaryzation exposure mask figure.
In training part, label is number after same Zhang Zhen's wave figure is divided according to setting resolution ratio, is numbered different
All vibration wave figure groups are combined into training set；When training set introducing UNet model is learnt, it is firstly introduced into identical number
Vibration wave figure is trained respectively, and all vibration wave figures of training set are subsequently introduced and are trained, and multiple training obtains can be to vibration wave figure
The new UNet model that middle salt mine is split.
In step 2b), during building UNet model, UNet model same level two blocks B and B ' it
Between introduce convolutional layer, the convolutional layer have an output characteristic pattern channel, feature of the convolutional layer first to the last layer in block B
Figure carries out convolution algorithm, secondly stacks to the characteristic pattern of preceding layer in block B ', finally rolls up the characteristic pattern of heap poststack
Product operation, the new UNet model after introducing convolutional layer can obtain original high resolution, the corresponding high definition of large scale vibration wave figure is covered
Film figure.
Optionally, expert sets resolution ratio as PxQ, finds out the vibration wave figure for collecting that vibration wave figure intermediate-resolution is more than PxQ, and will
Resolution ratio is more than that the vibration wave figure of PxQ is divided into PxQ size, and same Zhang Zhen's wave figure is divided into multiple vibration wave figure units, Duo Gezhen
The sum of area of wave figure unit is not less than the gross area of Zhang Zhen's wave figure.
Based on above-mentioned detection method, the present invention also provides a kind of salt mine intelligent checking system based on convolutional neural networks,
The detection system includes:
Collection module, for collecting vibration wave figure；
Mark module, for marking the salt mine in Expert opinion vibration wave figure, expert will be judged as salt mine in vibration wave figure
Part is labeled as pixel 1, remaining is labeled as pixel 0；
Setting segmentation module, for setting resolution ratio and according to the vibration wave figure after resolution ratio segmentation expert's label；
Collection modules, for all vibration wave figures after set-partition；
Training building module, for building UNet model on deep learning platform, and in the same level of UNet model
Convolutional layer is introduced, parameter is stopped by the training of setting UNet model, vibration wave set of graphs introducing UNet model is trained, structure
Building can be to the new UNet model that salt mine in vibration wave figure is split；
Vibration wave figure to be detected, is split, the exposure mask of vibration wave figure after being divided by new UNet model based on given threshold
Figure, and is transformed into binaryzation exposure mask figure for exposure mask figure according to given threshold, splices and obtains high definition after all binaryzation exposure mask figures and cover
Film figure sets salt mine threshold value, compares salt mine area in high definition exposure mask figure and glues the ratio of the gross area, thus find in vibration wave figure whether
It was found that salt mine.
In exposure mask figure more than given threshold part output pixel value be 1, otherwise output pixel value be 0, can be realized by
Exposure mask figure is transformed into binaryzation exposure mask figure.
Detection system further include:
Coding module, for encoding to vibration wave figure, all vibration wave figure groups are combined into training set, and same Zhang Zhen's wave figure is pressed
It is number according to label after setting resolution ratio segmentation；Training building module is firstly introduced into the difference of the vibration wave figure with identical number
It is trained, all vibration wave figures of training set is subsequently introduced and are trained, repeatedly training obtains to carry out salt mine in vibration wave figure
The new UNet model of segmentation.
Optionally, involved convolutional layer has an output characteristic pattern channel, and the convolutional layer is first to the last layer in block B
Characteristic pattern carry out convolution algorithm, secondly the characteristic pattern of preceding layer in block B ' is stacked, finally by the characteristic pattern of heap poststack
Convolution algorithm is carried out, it is corresponding that the new UNet model after introducing convolutional layer can obtain original high resolution, large scale vibration wave figure
High definition exposure mask figure.
A kind of salt mine intelligent detecting method and system based on convolutional neural networks of the invention, has compared with prior art
Beneficial effect be:
1) detection method of the invention is for detecting salt mine, including training part and detection part；Wherein, training part needs
Vibration wave figure is collected, the salt mine in Expert opinion vibration wave figure is simultaneously marked, the segmentation that resolution ratio carries out vibration wave figure is then set,
Divided vibration wave figure builds UNet model as training set on deep learning platform, introduces in the same level of UNet model
The training of convolutional layer, setting UNet model stops parameter, training set introducing UNet model is learnt, obtaining can be to vibration wave
The new UNet model that salt mine is split in figure；In detection part, vibration wave figure to be detected is introduced into new UNet model, passes through setting
Threshold segmentation vibration wave figure to be detected obtains exposure mask figure, and exposure mask figure is transformed into binaryzation exposure mask figure, splices all binaryzations and covers
High definition exposure mask figure is obtained after film figure, then by setting salt mine threshold value, compare the ratio that salt mine area in high definition exposure mask figure glues the gross area
Example finds whether find salt mine in vibration wave figure；This method has convolution by segmentation vibration wave figure, and by vibration wave figure introducing after segmentation
The UNet model of layer is learnt, to construct new UNet model, while improving detection efficiency, also reduces new UNet
The calculation amount of model is particularly suitable for the vibration wave figure of the big resolution ratio of intelligent measurement large scale；
2) detection system of the invention is combined with the above method, also includes training part and detection part；In training department
Point, the collection of vibration wave figure is completed using collection module, utilizes the salt mine part of mark module label the determined vibration wave figure of expert, benefit
Vibration wave figure is divided into small size with setting segmentation module, using vibration wave figure after collection modules set-partition, is constructed using training
Module builds UNet model on deep learning platform, and introduces convolutional layer in the same level of UNet model, to shaking after segmentation
Building can be to the new UNet model that salt mine in vibration wave figure is split after wave figure carries out deep learning；It is in detection part, then sharp
Vibration wave figure is split with new UNet model, exposure mask figure is transformed into two-value by the exposure mask figure of vibration wave figure after then being divided
Change exposure mask figure, obtains high definition exposure mask figure after splicing all binaryzation exposure mask figures, set salt mine threshold value, compare salt in high definition exposure mask figure
The ratio of the viscous gross area of ore faces product, finds whether find salt mine in vibration wave figure；This system is while improving detection efficiency, also
The calculation amount for reducing new UNet model is particularly suitable for the vibration wave figure of the big resolution ratio of intelligent measurement large scale.
Detailed description of the invention
Attached drawing 1 is underground vibration wave imaging (left side) of the present invention and its salt mine exposure mask figure (right side) after marking；
Attached drawing 2 is the structural schematic diagram of Unet model of the present invention；
Attached drawing 3 is Unet model of the present invention with the connection schematic diagram between the block of left and right two of level；
Attached drawing 4 is Unet model of the present invention with the connection schematic diagram for introducing convolutional layer between the block of left and right two of level；
Attached drawing 5 is the structure connection block diagram of the embodiment of the present invention two.
Each label information indicates in attached drawing:
10, collection module, 20, mark module, 30, setting segmentation module, 40, collection modules,
50, training building module, 60, new UNet model, 70, coding module.
Specific embodiment
The technical issues of to make technical solution of the present invention, solving and technical effect are more clearly understood, below in conjunction with tool
Body embodiment is checked technical solution of the present invention, is completely described, it is clear that described embodiment is only this hair
Bright a part of the embodiment, instead of all the embodiments.Based on the embodiment of the present invention, those skilled in the art are not doing
All embodiments obtained under the premise of creative work out, all within protection scope of the present invention.
Embodiment one:
The present embodiment proposes a kind of salt mine intelligent detecting method based on convolutional neural networks, which includes:
1) training part:
Vibration wave figure 1a) is collected, expert determines and marks to the salt mine in vibration wave figure, with reference to attached drawing 1,；
1b) divide vibration wave figure according to the resolution ratio of setting, using vibration wave figures all after segmentation as training set；
UNet model 1c) is built on deep learning platform, with reference to attached drawing 2, introduces volume in the same level of UNet model
The training of lamination, setting UNet model stops parameter, training set introducing UNet model is learnt, obtaining can be to vibration wave figure
The new UNet model that middle salt mine is split；
2) detection part:
New UNet model 2a) is deployed to computer equipment；
Vibration wave figure to be detected is introduced new UNet model and is divided into multiple vibration wave figure units, owned by 2b) given threshold
The exposure mask figure of vibration wave figure unit, and according to the threshold value of setting, exposure mask figure is transformed into binaryzation exposure mask figure；
High definition exposure mask figure is obtained after 2c) splicing all binaryzation exposure mask figures, salt mine threshold value is set, compares in high definition exposure mask figure
Salt mine area glues the ratio of the gross area, to find whether find salt mine in vibration wave figure.
In step 1a), after expert determines vibration wave figure, the part that salt mine is judged as in vibration wave figure is labeled as
Pixel 1, i.e., white, remaining is labeled as pixel 0, i.e. black.
In step 2b), in exposure mask figure more than given threshold part output pixel value be 1, otherwise output pixel value be
0, it can be realized and exposure mask figure is transformed into binaryzation exposure mask figure.
In training part, label is number after same Zhang Zhen's wave figure is divided according to setting resolution ratio, is numbered different
All vibration wave figure groups are combined into training set；When training set introducing UNet model is learnt, it is firstly introduced into identical number
Vibration wave figure is trained respectively, and all vibration wave figures of training set are subsequently introduced and are trained, and multiple training obtains can be to vibration wave figure
The new UNet model that middle salt mine is split.
In step 2b), during building UNet model, UNet model same level two blocks B and B ' it
Between introduce convolutional layer, the convolutional layer have an output characteristic pattern channel, feature of the convolutional layer first to the last layer in block B
Figure carries out convolution algorithm, secondly stacks to the characteristic pattern of preceding layer in block B ', finally rolls up the characteristic pattern of heap poststack
Product operation, the new UNet model after introducing convolutional layer can obtain original high resolution, the corresponding high definition of large scale vibration wave figure is covered
Film figure.
Optionally, expert sets resolution ratio as PxQ, finds out the vibration wave figure for collecting that vibration wave figure intermediate-resolution is more than PxQ, and will
Resolution ratio is more than that the vibration wave figure of PxQ is divided into PxQ size, and same Zhang Zhen's wave figure is divided into multiple vibration wave figure units, Duo Gezhen
The sum of area of wave figure unit is not less than the gross area of Zhang Zhen's wave figure.
With reference to attached drawing 3,4, two blocks B and B ' of same level, if the characteristic pattern of the last layer is fm in B1Size be H ×
W × C1 respectively indicates height, width and the port number of characteristic pattern.The characteristic pattern of B ' preceding layer input is fm2, size is H × W
Two characteristic patterns are stacked in × C2, B ', size is H × W × (C1+C2), if the characteristic pattern that first layer convolution exports in B ' is
fm3, size is H × W × K, and convolution kernel size is R × S × (C1+C2) × K, and R is convolution kernel height, and S is convolution kernel width,
It (C1+C2) is convolution kernel port number, K is convolution kernel number.Then by characteristic pattern fm1Obtain characteristic pattern fm3Calculation amount are as follows:
Q1=R × S × (C1+C2) × H × W × K
In order to increase model to the learning ability of feature, a convolutional layer is added, between two blocks B and B ' in order to reduce
Operand, enabling the port number of output characteristic pattern is 1.I.e. in characteristic pattern fm1With fm2Before being stacked, to fm1Carry out convolution fortune
It calculates, and its convolution kernel size is R1 × S1 × C1 × 1, obtains characteristic pattern fm1', by fm1' and fm2It stacks, carries out subsequent convolution
Operation obtains characteristic pattern fm3.After improving, by characteristic pattern fm1Obtain characteristic pattern fm3Calculation amount are as follows:
Q2=R × S × (C2+1) × H × W × K+R1 × S1 × C1 × 1 × H × W
In addition, according to design feature C1, C2 and K of UNet, there are relation of equality.Convolution kernel width, height are generally all provided with
It is set to 3, i.e. R=S=R1=S1.
Therefore, Q1=2 × Q2, i.e. calculation amount reduce one times.
UNet model is suitble to the image segmentation of small scale, but vibration wave figure is because apart from remote and scale is generally large, directly
Image serious distortion can be made by narrowing down to input size, lose a large amount of detailed information, classification is caused to fail.On the other hand, directly
A part of image is cut, diminution search coverage is equivalent to, the detection result of zonule can not react whole salt mine looks.This implementation
Convolutional layer is introduced between two blocks of level in the UNet model of example, high-resolution, large scale vibration wave figure are carried out orderly
Shearing keeps image detail, the image of all shearings is input to model and is successively classified, according to input picture in original graph
All classification results are spliced in position as in, obtain the corresponding high definition exposure mask of original high resolution, large scale vibration wave figure
Figure.
The detection method of the present embodiment is for detecting salt mine, including training part and detection part；Wherein, training part needs
Vibration wave figure is collected, the salt mine in Expert opinion vibration wave figure is simultaneously marked, the segmentation that resolution ratio carries out vibration wave figure is then set,
Divided vibration wave figure builds UNet model as training set on deep learning platform, introduces in the same level of UNet model
The training of convolutional layer, setting UNet model stops parameter, training set introducing UNet model is learnt, obtaining can be to vibration wave
The new UNet model that salt mine is split in figure；In detection part, vibration wave figure to be detected is introduced into new UNet model, passes through setting
Threshold segmentation vibration wave figure to be detected obtains exposure mask figure, and exposure mask figure is transformed into binaryzation exposure mask figure, splices all binaryzations and covers
High definition exposure mask figure is obtained after film figure, then by setting salt mine threshold value, compare the ratio that salt mine area in high definition exposure mask figure glues the gross area
Example finds whether find salt mine in vibration wave figure.
Vibration wave figure after segmentation is introduced the UNet with convolutional layer by segmentation vibration wave figure by the detection method of the present embodiment
Model is learnt, to construct new UNet model, while improving detection efficiency, also reduces the meter of new UNet model
Calculation amount is particularly suitable for the vibration wave figure of the big resolution ratio of intelligent measurement large scale.
Embodiment two:
In conjunction with attached drawing 5, the present embodiment proposes a kind of salt mine intelligent checking system based on convolutional neural networks, the detection system
System includes:
Collection module 10, for collecting vibration wave figure；
Mark module 20, for marking the salt mine in Expert opinion vibration wave figure, expert will be judged as salt mine in vibration wave figure
Part be labeled as pixel 1, i.e., white, remaining is labeled as pixel 0, i.e. black；
Setting segmentation module 30, for setting resolution ratio and according to the vibration wave figure after resolution ratio segmentation expert's label；
Collection modules 40, for all vibration wave figures after set-partition；
Training building module 50, for building UNet model on deep learning platform, and in the same layer of UNet model
Grade introduces convolutional layer, stops parameter by the training of setting UNet model, and vibration wave set of graphs introducing UNet model is trained,
Building can be to the new UNet model 60 that salt mine in vibration wave figure is split；
Vibration wave figure to be detected, is split, vibration wave figure covers after being divided by new UNet model 60 based on given threshold
Film figure, and exposure mask figure is transformed by binaryzation exposure mask figure according to given threshold, high definition is obtained after splicing all binaryzation exposure mask figures
Exposure mask figure sets salt mine threshold value, compares salt mine area in high definition exposure mask figure and glues the ratio of the gross area, thus find be in vibration wave figure
No discovery salt mine.
In exposure mask figure more than given threshold part output pixel value be 1, otherwise output pixel value be 0, can be realized by
Exposure mask figure is transformed into binaryzation exposure mask figure.
Detection system further include:
Coding module 70, for encoding to vibration wave figure, all vibration wave figure groups are combined into training set, and same Zhang Zhen's wave figure
It is number according to label after setting resolution ratio segmentation；Training building module 50 is firstly introduced into the vibration wave figure with identical number
It is trained respectively, all vibration wave figures of training set is subsequently introduced and are trained, multiple training obtains can be to salt mine in vibration wave figure
The new UNet model being split.
Involved convolutional layer has an output characteristic pattern channel, feature of the convolutional layer first to the last layer in block B
Figure carries out convolution algorithm, secondly stacks to the characteristic pattern of preceding layer in block B ', finally rolls up the characteristic pattern of heap poststack
Product operation, the new UNet model after introducing convolutional layer can obtain original high resolution, the corresponding high definition of large scale vibration wave figure is covered
Film figure.
With reference to attached drawing 3,4, two blocks B and B ' of same level, if the characteristic pattern of the last layer is fm in B1Size be H ×
W × C1 respectively indicates height, width and the port number of characteristic pattern.The characteristic pattern of B ' preceding layer input is fm2, size is H × W
Two characteristic patterns are stacked in × C2, B ', size is H × W × (C1+C2), if the characteristic pattern that first layer convolution exports in B ' is
fm3, size is H × W × K, and convolution kernel size is R × S × (C1+C2) × K, and R is convolution kernel height, and S is convolution kernel width,
It (C1+C2) is convolution kernel port number, K is convolution kernel number.Then by characteristic pattern fm1Obtain characteristic pattern fm3Calculation amount are as follows:
Q1=R × S × (C1+C2) × H × W × K
In order to increase model to the learning ability of feature, a convolutional layer is added, between two blocks B and B ' in order to reduce
Operand, enabling the port number of output characteristic pattern is 1.I.e. in characteristic pattern fm1With fm2Before being stacked, to fm1Carry out convolution fortune
It calculates, and its convolution kernel size is R1 × S1 × C1 × 1, obtains characteristic pattern fm1', by fm1' and fm2It stacks, carries out subsequent convolution
Operation obtains characteristic pattern fm3.After improving, by characteristic pattern fm1Obtain characteristic pattern fm3Calculation amount are as follows:
Q2=R × S × (C2+1) × H × W × K+R1 × S1 × C1 × 1 × H × W
In addition, according to design feature C1, C2 and K of UNet, there are relation of equality.Convolution kernel width, height are generally all provided with
It is set to 3, i.e. R=S=R1=S1.
Therefore, Q1=2 × Q2, i.e. calculation amount reduce one times.
The detection system of the present embodiment is combined with the detection method of embodiment one, also includes training part and test section
Point；In training part, the collection of vibration wave figure is completed using collection module 10, marks the determined vibration wave of expert using mark module 20
Vibration wave figure is divided into small size using setting segmentation module 30, after 40 set-partition of collection modules by the salt mine part of figure
Vibration wave figure builds UNet model using training building module 50 on deep learning platform, and in the same level of UNet model
Convolutional layer is introduced, constructs the new UNet that can be split to salt mine in vibration wave figure after carrying out deep learning to vibration wave figure after segmentation
Model 60；In detection part, then vibration wave figure is split using new UNet model 60, vibration wave figure covers after then being divided
Exposure mask figure is transformed into binaryzation exposure mask figure by film figure, is obtained high definition exposure mask figure after splicing all binaryzation exposure mask figures, is set salt mine
Threshold value compares the ratio that salt mine area in high definition exposure mask figure glues the gross area, finds whether find salt mine in vibration wave figure.The present embodiment
Detection system while improving detection efficiency, also reduce the calculation amount of new UNet model, be particularly suitable for intelligent inspection
Survey the vibration wave figure of the big resolution ratio of large scale.
Use above specific case elaborates the principle of the present invention and embodiment, these embodiments are
It is used to help understand core of the invention technology contents, the protection scope being not intended to restrict the invention, technical side of the invention
Case is not limited in above-mentioned specific embodiment.Based on above-mentioned specific embodiment of the invention, those skilled in the art
Without departing from the principle of the present invention, any improvement and modification to made by the present invention should all be fallen into of the invention special
Sharp protection scope.
Claims (10)
1. a kind of salt mine intelligent detecting method based on convolutional neural networks, which is characterized in that the detection method includes:
1) training part:
Vibration wave figure 1a) is collected, expert determines and marks to the salt mine in vibration wave figure；
1b) divide vibration wave figure according to the resolution ratio of setting, using vibration wave figures all after segmentation as training set；
UNet model 1c) is built on deep learning platform, is introduced convolutional layer in the same level of UNet model, is set UNet mould
The training of type stops parameter, and training set introducing UNet model is learnt, obtains to be split salt mine in vibration wave figure
New UNet model；
2) detection part:
New UNet model 2a) is deployed to computer equipment；
Vibration wave figure to be detected is introduced new UNet model and is divided into multiple vibration wave figure units, obtains all vibration waves by 2b) given threshold
The exposure mask figure of figure unit, and according to the threshold value of setting, exposure mask figure is transformed into binaryzation exposure mask figure；
High definition exposure mask figure is obtained after 2c) splicing all binaryzation exposure mask figures, salt mine threshold value is set, compares salt mine in high definition exposure mask figure
Area glues the ratio of the gross area, to find whether find salt mine in vibration wave figure.
2. a kind of salt mine intelligent detecting method based on convolutional neural networks according to claim 1, which is characterized in that
Step 1a) in, after expert determines vibration wave figure, the part that salt mine is judged as in vibration wave figure is labeled as pixel 1, remaining
Labeled as pixel 0.
3. a kind of salt mine intelligent detecting method based on convolutional neural networks according to claim 2, which is characterized in that
Step 2b) in, the part output pixel value in exposure mask figure being more than given threshold is 1, and otherwise output pixel value is 0, be can be realized
Exposure mask figure is transformed into binaryzation exposure mask figure.
4. a kind of salt mine intelligent detecting method based on convolutional neural networks according to claim 1, which is characterized in that
Training part, label is number after same Zhang Zhen's wave figure is divided according to setting resolution ratio, numbers different all vibration wave figures
Group is combined into training set；When training set introducing UNet model is learnt, it is firstly introduced into the vibration wave figure difference with identical number
It is trained, all vibration wave figures of training set is subsequently introduced and are trained, repeatedly training obtains to carry out salt mine in vibration wave figure
The new UNet model of segmentation.
5. a kind of salt mine intelligent detecting method based on convolutional neural networks according to claim 1, which is characterized in that
Step 2b) in, during building UNet model, convolution is introduced between two blocks B and B ' of the same level of UNet model
Layer, the convolutional layer have an output characteristic pattern channel, which carries out convolution to the characteristic pattern of the last layer in block B first
Secondly operation stacks the characteristic pattern of preceding layer in block B ', the characteristic pattern of heap poststack is finally carried out convolution algorithm, introduces
New UNet model after convolutional layer can obtain the corresponding high definition exposure mask figure of original high resolution, large scale vibration wave figure.
6. a kind of salt mine intelligent detecting method based on convolutional neural networks according to claim 1, which is characterized in that specially
Family sets resolution ratio as PxQ, finds out the vibration wave figure for collecting that vibration wave figure intermediate-resolution is more than PxQ, and is more than the vibration of PxQ by resolution ratio
Wave figure is divided into PxQ size, and same Zhang Zhen's wave figure is divided into multiple vibration wave figure units, the sum of the area of multiple vibration wave figure units
Not less than the gross area of Zhang Zhen's wave figure.
7. a kind of salt mine intelligent checking system based on convolutional neural networks, which is characterized in that the detection system includes:
Collection module, for collecting vibration wave figure；
Mark module, for marking the salt mine in Expert opinion vibration wave figure, expert will be judged as the part of salt mine in vibration wave figure
Labeled as pixel 1, remaining is labeled as pixel 0；
Setting segmentation module, for setting resolution ratio and according to the vibration wave figure after resolution ratio segmentation expert's label；
Collection modules, for all vibration wave figures after set-partition；
Training building module is introduced for building UNet model on deep learning platform, and in the same level of UNet model
Convolutional layer stops parameter by the training of setting UNet model, vibration wave set of graphs introducing UNet model is trained, building can
With the new UNet model being split to salt mine in vibration wave figure；
Vibration wave figure to be detected, is split, the exposure mask figure of vibration wave figure after being divided by new UNet model based on given threshold, and
Exposure mask figure is transformed into binaryzation exposure mask figure according to given threshold, obtains high definition exposure mask figure after splicing all binaryzation exposure mask figures,
Salt mine threshold value is set, compares the ratio that salt mine area in high definition exposure mask figure glues the gross area, to find whether find in vibration wave figure
Salt mine.
8. a kind of salt mine intelligent checking system based on convolutional neural networks according to claim 7, which is characterized in that cover
The part output pixel value in film figure being more than given threshold is 1, and otherwise output pixel value is 0, can be realized and changes exposure mask figure
At binaryzation exposure mask figure.
9. a kind of salt mine intelligent checking system based on convolutional neural networks according to claim 7, which is characterized in that also
Include:
Coding module, for encoding to vibration wave figure, all vibration wave figure groups are combined into training set, and same Zhang Zhen's wave figure is according to setting
Label is number after determining resolution ratio segmentation；Training building module is firstly introduced into the vibration wave figure with identical number and carries out respectively
Training, is subsequently introduced all vibration wave figures of training set and is trained, and repeatedly training obtains to be split salt mine in vibration wave figure
New UNet model.
10. a kind of salt mine intelligent checking system based on convolutional neural networks according to claim 7, which is characterized in that
The convolutional layer has an output characteristic pattern channel, which carries out convolution to the characteristic pattern of the last layer in block B first
Secondly operation stacks the characteristic pattern of preceding layer in block B ', the characteristic pattern of heap poststack is finally carried out convolution algorithm, introduces
New UNet model after convolutional layer can obtain the corresponding high definition exposure mask figure of original high resolution, large scale vibration wave figure.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811449637.3A CN109544545A (en) | 2018-11-30 | 2018-11-30 | A kind of salt mine intelligent detecting method and system based on convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811449637.3A CN109544545A (en) | 2018-11-30 | 2018-11-30 | A kind of salt mine intelligent detecting method and system based on convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109544545A true CN109544545A (en) | 2019-03-29 |
Family
ID=65851529
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811449637.3A Pending CN109544545A (en) | 2018-11-30 | 2018-11-30 | A kind of salt mine intelligent detecting method and system based on convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109544545A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110060251A (en) * | 2019-04-26 | 2019-07-26 | 福州大学 | A kind of building surface crack detecting method based on U-Net |
| CN110287930A (en) * | 2019-07-01 | 2019-09-27 | 厦门美图之家科技有限公司 | Wrinkle disaggregated model training method and device |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107016665A (en) * | 2017-02-16 | 2017-08-04 | 浙江大学 | A kind of CT pulmonary nodule detection methods based on depth convolutional neural networks |
| AU2017101166A4 (en) * | 2017-08-25 | 2017-11-02 | Lai, Haodong MR | A Method For Real-Time Image Style Transfer Based On Conditional Generative Adversarial Networks |
| CN108052966A (en) * | 2017-12-08 | 2018-05-18 | 重庆邮电大学 | Remote sensing images scene based on convolutional neural networks automatically extracts and sorting technique |
| CN108229011A (en) * | 2017-12-29 | 2018-06-29 | 中国地质大学（武汉） | A kind of shale lithofacies development Dominated Factors judgment method, equipment and storage device |
-
2018
- 2018-11-30 CN CN201811449637.3A patent/CN109544545A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107016665A (en) * | 2017-02-16 | 2017-08-04 | 浙江大学 | A kind of CT pulmonary nodule detection methods based on depth convolutional neural networks |
| AU2017101166A4 (en) * | 2017-08-25 | 2017-11-02 | Lai, Haodong MR | A Method For Real-Time Image Style Transfer Based On Conditional Generative Adversarial Networks |
| CN108052966A (en) * | 2017-12-08 | 2018-05-18 | 重庆邮电大学 | Remote sensing images scene based on convolutional neural networks automatically extracts and sorting technique |
| CN108229011A (en) * | 2017-12-29 | 2018-06-29 | 中国地质大学（武汉） | A kind of shale lithofacies development Dominated Factors judgment method, equipment and storage device |
Non-Patent Citations (1)
| Title |
|---|
| 尹世豪: "基于深度学习的同步碎石封层碎石覆盖率检测算法研究", 《万方》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110060251A (en) * | 2019-04-26 | 2019-07-26 | 福州大学 | A kind of building surface crack detecting method based on U-Net |
| CN110287930A (en) * | 2019-07-01 | 2019-09-27 | 厦门美图之家科技有限公司 | Wrinkle disaggregated model training method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109118479B (en) | Capsule network-based insulator defect identification and positioning device and method | |
| CN103049763B (en) | Context-constraint-based target identification method | |
| CN107346420A (en) | Text detection localization method under a kind of natural scene based on deep learning | |
| CN104952073B (en) | Scene Incision method based on deep learning | |
| CN109766936A (en) | Image change detection method based on information transmitting and attention mechanism | |
| CN103780899B (en) | A kind of detect the most disturbed method of video camera, device and video monitoring system | |
| CN104240256A (en) | Image salient detecting method based on layering sparse modeling | |
| CN112149547A (en) | Remote sensing image water body identification based on image pyramid guidance and pixel pair matching | |
| CN104182985A (en) | Remote sensing image change detection method | |
| CN103544488B (en) | A kind of face identification method and device | |
| EP3680608A1 (en) | Antenna downward inclination angle measurement method based on multi-scale detection algorithm | |
| CN105426894A (en) | Image detection method and device for railroad plug nails | |
| CN109544545A (en) | A kind of salt mine intelligent detecting method and system based on convolutional neural networks | |
| CN106780727A (en) | A kind of headstock detection model method for reconstructing and device | |
| CN102509299B (en) | Image salient area detection method based on visual attention mechanism | |
| CN102867192A (en) | Scene semantic shift method based on supervised geodesic propagation | |
| CN116029979A (en) | Cloth flaw visual detection method based on improved Yolov4 | |
| Chu et al. | Deep learning method to detect the road cracks and potholes for smart cities | |
| CN106295556A (en) | A kind of Approach for road detection based on SUAV Aerial Images | |
| Yan et al. | The research of building earthquake damage object-oriented change detection based on ensemble classifier with remote sensing image | |
| CN104778461A (en) | Coal-rock recognition method based on similar measure learning | |
| CN104732239B (en) | Coal petrography sorting technique based on the asymmetric generalized gaussian model of wavelet field | |
| CN106682668A (en) | Power transmission line geological disaster monitoring method using unmanned aerial vehicle to mark images | |
| Wang et al. | Rock fracture tracing based on image processing and SVM | |
| CN110163081A (en) | Regional invasion real-time detection method, system and storage medium based on SSD |
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: 20190329 |
|
| RJ01 | Rejection of invention patent application after publication |