CN110119753A - A kind of method of reconstituted texture identification lithology - Google Patents
A kind of method of reconstituted texture identification lithology Download PDFInfo
- Publication number
- CN110119753A CN110119753A CN201910016656.5A CN201910016656A CN110119753A CN 110119753 A CN110119753 A CN 110119753A CN 201910016656 A CN201910016656 A CN 201910016656A CN 110119753 A CN110119753 A CN 110119753A
- Authority
- CN
- China
- Prior art keywords
- image
- model
- rock
- texture
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/061—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using biological neurons, e.g. biological neurons connected to an integrated circuit
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Abstract
The invention discloses a kind of methods of reconstituted texture identification lithology, include the following steps: to establish type specimen image library;Establish standardized images library;The problem of extracting the feature of rock sample image using the discrete convolution and the operation of maximum pondization of convolutional neural networks MobileNet-V2 convolution model, and gradient explosion and gradient disappearance during image characteristics extraction effectively prevented using new algorithm;And the maximum rock category output of select probability is as a result, complete identification process.The present invention solves existing algorithm and identifies that rock lithology by rock structure is included distribution of particles and section texture developmental state at random in field, is included that rock color and clarity etc. are confined to the low problem low with recognition efficiency of precision caused by two dimension by image quality.The present invention with that can identify various lithology extensively, and not by background, the limitation of the external factor such as color improves recognition accuracy;Promote the intelligent recognition speed of service.
Description
Technical field
The present invention relates to computer or mobile intelligent terminal Intelligent treatment identification rock lithology fields, and more specifically it is
A kind of method of the reconstituted texture identification lithology of convenient and efficient.
Background technique
The current existing convenient recognizer of rock lithology is traditional machine learning algorithm.Data are located in advance first
Reason, determines classification method according to pretreated data characteristics, then carries out data analysis and visualization, judges where is data fit
Kind distribution, finally (decision tree, Bayes) tests, model integrated on existing algorithm.There are four on existing algorithm
Big defect: first is that the obvious Rock Species of identification feature difference are only capable of, such as the identification of coal petrography and other single rocks;Second is that by rock
Color and image quality influence degree are larger, and identification level cannot reach fuzzy level;Third is that grain texture feature after threshold
The small rock identification difficulty of difference;Fourth is that calculating speed is slow, test accuracy rate is low.
Therefore, a kind of new method is researched and developed on the basis of prior art, and above-mentioned four big defects can be allowed effectively to be solved
Certainly, rock recognition accuracy and recognition efficiency with higher become an important, urgent and significant job.
Summary of the invention
The purpose of the invention is to overcome the shortcoming of above-mentioned background technology, and a kind of reconstituted texture proposed identifies
The method of lithology, it solves existing algorithm and identifies that rock lithology by rock structure is included distribution of particles and section at random in field
Texture developmental state restricts, and is included rock color by image quality and clarity is confined to that precision caused by two dimension is low and identification is imitated
The low problem of rate.
The purpose of the present invention is what is reached by following measure: a kind of method of reconstituted texture identification lithology, including such as
Lower step:
Step A, it establishes type specimen image library: the rock sample of different kinds of rocks being carried out using photograph, photographic equipment more
Angle, multiple spurs from shooting, and complete number, name, image size, typing and petrographic description, unified picture format, classification storage
After depositing, inputting fixed format attribute, type specimen image library is established;
Step B, standardized images library is established: right using Image Argument (image parameter to match with this algorithm)
Image in type specimen image library is pre-processed, and is converted to four-dimensional (sample including feature normalization, by the format of image
Number, high, wide, number of dimensions) the format of input model, the optimization based on GPU, form standardized images, establish standardized images
Library, and the image in standardized images library is divided into training set and test set at random;
Step C, characteristics of image is extracted: every for standardized images library for convenience of the quick identification of computer and mobile terminal
Image in one rocks training set is established convolutional neural networks MobileNet-V2 and (is belonged in lightweight convolutional neural networks
A kind of existing model) convolution model, pass through the Conv (discrete convolution operation) and MaxPool of the model (maximum pondization operation)
And first algorithm core formula (Leaky_SELU) extract the feature of image after each Zhang Yanshi standardization, and it is feature is unified
It is stored in fixed format, in fixed route file, by geometry identification, fuzzy diagnosis, model identification and at random identifies successively
Feature extraction obtains preliminary texture reconstruct characteristic function;
Step D, training discrimination model: by the discrete convolution of the MobileNet-V2 convolution model of training set and maximum pond
The result of operation output expands into one-dimensional data, adds shot and long term memory models using stored feature as the function learnt
Input in LSTM (Long Short-Term Memory) as hidden layer carries out model by the way of backpropagation
Training, LSTM model by geometry, fuzzy, model, random training characteristics result superposition, Process fusion Leaky_SELU and
LSLE algorithm finally obtains new reconstituted texture characteristic function, chooses preferred plan with this, generates discrimination model;
Step E, optimize discrimination model: the method that training process uses learning-rate decay (learning rate decaying),
To improve the training speed of entire model;In the training process of backpropagation, to improve model to the discrimination of detail characteristic,
Penalty values in training are predicted using Large-Margin Softmax Loss Expand (abbreviation LSLE) the second core algorithm,
And penalty values constantly are reduced using a kind of Adam (algorithms most in use of deep learning) algorithm, so that Optimized model, is improved
The output accuracy of model.
Step F, characteristic matching exports result: for unknown rock image, substituting into discrimination model and extracts its thin portion spy
Sign, and Large-Margin Softmax (existing image similarity mode algorithm) algorithm is utilized, output newly acquires to be matched
The possible rock category of unknown rock probability, and the maximum rock category of select probability as unique output as a result, completing
Identification process.Its characteristic matching and result output process are as shown in Figure 1, pre-process the rock image obtained by image-forming module
Feature will be entered identifying system, then be merged in Leaky_SELU and LSLE algorithm and standard picture library by reconstituted texture
Image compares.If similarity reaches requirement, then the similarity highest being matched in outputting standard image library by comparison
Rock image and corresponding rock properties, and new picture is added in the standard picture library of the lithology and is stored.If similarity
It is unable to reach requirement, then needs to be manually entered the attribute of new lithology, and mark is included in as new sample by study module
In quasi- image library, stored.
In the above-mentioned technical solutions, the step A further comprises:
Step A1: photograph, photographic equipment include camera, video camera, tablet computer, mobile phone, camera, are imaged by it
The rock sample image of module acquisition should be colored, save as unified format, such as * .jpg format, image is more visible, should not
Shooting angle, shooting distance are asked, does not require image size.
Step A2: the image of prescribed form should be stored in different files according to different rock categories in step A1,
Type specimen image library should include the image of storage, storage path, picture number, the name of image sample, petrographic description including solid
The attribute for the formula that fixes is expressed.
In the above-mentioned technical solutions, the step B further comprises:
Step B1: every rock sample image carries out characteristics of image normalizing in the type specimen image library generated to step A2
Change operation.Characteristics of image normalization is to find one group of parameter using the not bending moment of image, can eliminate other transforming function transformation functions
Influence to image transformation.Unique canonical form is namely converted images into, the influence of geometry deformation can be resisted, is increased
Recognition accuracy of the discrimination model to similar image, formula are as follows: output=(pixel value-picture pixels value average value)/
(variances of picture pixels).
Step B2: the format conversion operation of image is carried out to the image that step B1 is generated.Image format conversion refer to by
The format of image is converted to the format of four-dimensional input model, and respectively (color of picture is logical for sample number, height, width, number of dimensions
Road).
Step B3: the operation of the optimization based on GPU is carried out to the step B2 image generated.Optimization based on GPU, which refers to, to be schemed
Float as being converted to float32 (existing float format) facilitates and carries out GPU acceleration operation.
Step B4: the standardized images that all step B3 are generated are ordered together with storage path, picture number, image sample
Name, petrographic description include that standardized images library is established in the attribute expression of fixed format.
Step B5: the image of every kind of rock type in standardized images library that step B4 is generated is randomly divided by 1:1
For training the training set of discrimination model and for verifying the test set for differentiating effect.
In the above-mentioned technical solutions, the step C further comprises:
Step C1: MobileNet-V2 convolution model is established.The composition of MobileNet-V2 convolution model is as shown in the table:
Dimension therein refers to the color channel of picture, and step-length refers to the distance that convolution kernel moves every time, the dimension after expansion
Degree refers to the new color channel number of the constant longitudinal intensification of length and width.
The step uses new algorithm core formula, can provide canonical more stable than before in each step of calculating
Formula, to show be mean value with 0,1 for variance sample distribution, gradient is effectively prevented for such extensive sample training problem
The problem of explosion and gradient disappear.
First algorithm core formula (referred to as Leaky_SELU) is as follows:
Wherein λ is regular terms, and x is input item, and b is bias term,
C is empirical when rock characteristic learns, and α is learning rate, and e is constant.
Step C2: the training set chosen in step B5 is substituted into the MobileNet-V2 convolution model that step C1 is established, is led to
The discrete convolution and maximum pondization operation for crossing the model extract the feature of each picture, and complete geometry identification, a mould
The successively feature extraction that paste identification, model are identified and identified at random obtains preliminary texture reconstruct characteristic function.The texture of image
Reconstruct is come shield portions characteristics of image by the way of a part of random shielded image, and same image generates multiple block
The copy of different piece, to guarantee that the minutia of image local preferably can be saved and extract, and different copies is still
Can recognize that the due common trait of original rocks institute, an especially picture texture block at random be split as multiple thin portions it
Afterwards, the thin portion of texture is special before still retaining after new algorithm of the present invention processing by randomly assigne again spliced new texture
Sign.Texture reconstruct in the present invention passes through a geometry identification, fuzzy diagnosis, circulation model identification and identified at random
Feature extraction goes to realize, the first texture reconstruct number of same image does not set the circulation upper limit, be such as set as at least more than 10 times
100 times.
In the above-mentioned technical solutions, the step D further comprises:
Step D1: by result that step C2 is trained every time successively with 1. geometry (by freely being converted to image geometry space,
Known according to object in the mobile mathematical feature, including translation, offset, rotation, control with changed scale etc. in transformation with invariance in space
Feature Dui Bi not extracted), 2. obscure that (image is converted gray scale, extracts the related operation of texture information and is incorporated into a texture
In unit, similarity between texture being carried out characteristic quantification and the individual measured), 3. model is (to same photo rock texture
Cutting reassembles into new picture and carries out feature extraction using conventional histogramming algorithm after fragmentating), it is 4. random (at random to can
A same type of wherein picture for energy uses SIFT algorithm (i.e. Scale invariant features transform " Scale-invariant
The image characteristic extracting method of feature transform ") carry out feature extraction) four kinds the knot obtained is identified to characteristics of image
Fruit carries out feature superposition, and a process as reconstituted texture obtains preliminary characteristic function, is launched into one-dimensional data and is added to
Input in shot and long term memory models LSTM (Long Short-Term Memory) as hidden layer, in a manner of backpropagation
LSTM model is trained.Leaky_SELU and LSLE algorithm is merged simultaneously, obtains new image training characteristics, i.e., new rock
Stone image texture reconstructs characteristic function.
Wherein back-propagation algorithm is mainly by two links (excitation is propagated, weight updates) iterative cycles iteration, until net
Until network reaches scheduled target zone to the response of input.The learning process of shot and long term memory models LSTM is by forward-propagating
Journey and back-propagation process composition.During forward-propagating, input information, through hidden layer, is successively handled and is passed by input layer
To output layer.If cannot get desired output valve in output layer, take the quadratic sum of output and desired error as target
Function, is transferred to backpropagation, successively finds out objective function to the partial derivative of each neuron weight, constitute objective function to weight to
The ladder amount of amount, as the foundation of modification weight, the study of network is completed during weight is modified.Error reaches desired value
When, e-learning terminates.
Step D2: training and the loss curve predicted in observation of steps D1, if the penalty values of training are much smaller than prediction
Penalty values, then needing to increase the quantity of picture or adding regularization parameter λ on loss function;If the loss of training
Value then needing to deepen the depth of model, and modifies empirical c when rock characteristic study much larger than the penalty values of prediction.
Then think that model training is completed until the result of test set output reaches certain accuracy rate.
In the above-mentioned technical solutions, the step E further comprises:
Step E1: in step C and step D, the method that training process uses learning-rate decay, to improve
The training speed of entire model.The formula of learning-rate decay method is decaying learning rate=learning rate × attenuation rate
× (current period/damped cycle), such as: frequency of training can be set as 1000 times, damped cycle 100 is arranged trained
Batch-size (existing a kind of batch of dimensional parameters) is 64, i.e., 64 pictures of selection are trained every time, and specific training is wanted
Ask to set according to actual needs.
Step E2: in the backpropagation training process of step D1, using Large-Margin Softmax Loss
The second core algorithm of Expand (abbreviation LSLE) come predict training in penalty values.Because rock characteristic can not embody, because
This specially enhances the difficulty of similar study when learning similar sample, this difficulty is bigger than inhomogeneous difficulty.It is such
It treats with a certain discrimination so that the ga s safety degree of feature enhances.So the second algorithm core formula existsBasis
Upper (reference: Liu, Weiyang, et al. " Large-Margin Softmax Loss for Convolutional Neural
Networks. " ICML.2016.) it joined more feasible added value:
Second algorithm core formula is as follows:
Wherein LiFor the penalty values for predicting i-th, Li=0 indicates that judgement is completely correct;When sample m quantity is bigger, the boundary of classification is bigger, and learning difficulty is certainly
It is higher.Meanwhile the D (θ) in formula must be a monotonous descending function and D (θ)=cos (π/m) D (m)=cos (π/m), to protect
Demonstrate,proving ψ (θ) is a continuous function.WyiIt is the output weight of ith feature, xiThe input of i-th of classification, λ represent when just
Then change coefficient, for constraining the range of formula, more reliable penalty values and precision can be provided, it will not be because of the slightly change of picture
It moves and violent prediction fault occurs, wherein lose value metric is the difference between predicted value and true value.
Step E3: (a kind of using Adam in order to constantly reduce penalty values in the backpropagation training process of step D1
The algorithms most in use of deep learning) it optimizes.The algorithm is optimized according to the result and actual result that acquire, to reduce
Penalty values.
In the above-mentioned technical solutions, the step F further comprises:
Step F1: for the unknown rock image obtained by image-forming module, being pre-processed using step B, is formed
Standardized images.
Step F2: by the standardized images of the obtained unknown rock image of F1, it is input to the differentiation mould that step D is trained
Type seeks the class probability of unknown rock using the addition operation of LSLE algorithm, and the maximum classification of select probability is as differentiation knot
Fruit output.
The accuracy of rock intelligent recognition, precision are always the main problem that industry is concerned about, a kind of reconstituted texture identification
The method of lithology is standardized image using Image Argument (image enhancement technique), is based on GPU (Graphics
Processing Analysis, graphics processor) acceleration is optimized to graphics process, exclude the interference of descriptive geometry deformation.
This method is put forward for the first time the texture reconstruct identification lithofacies technology of image, enhances the detail characteristic of image by algorithm, increases and differentiate
Model improves training effectiveness to the differentiation accuracy rate of similar similar picture, and in the training process, innovation uses Leaky_
SELU and LSLE algorithm is expanded existing algorithm to promote discrimination model to the recognition capability of thin portion texture, and is merged
Algorithm is further by learning-rate decay (learning rate decaying) algorithm and Adam (a kind of algorithms most in use of deep learning)
Raising training effectiveness, above-mentioned four big defects of very good solution, rock recognition accuracy and recognition efficiency with higher.
The beneficial effects of the present invention are:
1, various lithology can be identified extensively, even the small different rocks of feature difference also can be identified effectively;
2, pretreatment, novelty reconstruct rock texture and particle space structure are standardized to rock image, are not carried on the back
Scape, the external factor such as color limitation, so that image quality is smaller to the influence degree of recognition result, improves training effectiveness;
3, the texture reconstruct after incorporating Leaky_SELU and LSLE algorithm can solve that rock particles is tiny very well and texture compared with
Approximation identifies bring difficulty to rock, improves recognition accuracy;
4, original image is based on GPU and accelerates optimization, promote the intelligent recognition speed of service.
Detailed description of the invention
Fig. 1 Image Feature Matching of the present invention and result export entire flow schematic diagram.
The method schematic diagram of Fig. 2 reconstituted texture identification lithology of the present invention.
Texture restructuring procedure basic skills schematic diagram Fig. 3 of the invention.
Specific embodiment
In order to which the technological means, technical characteristic, the method for reaching purpose that realize the present invention are easier to understand, tie below
Close attached drawing operating procedure be just embodied and do further details of elaboration, but they and do not constitute a limitation of the invention, only make
Citing.Simultaneously by illustrating that advantages of the present invention will become clearer and be readily appreciated that.
A kind of method of reconstituted texture identification lithology includes the following steps: in operation
Step A, it establishes type specimen image library: the rock sample of different kinds of rocks being carried out using photograph, photographic equipment more
Angle, multiple spurs from shooting, and complete number, name, image size, typing petrographic description, unified picture format, classification storage
It deposits, input fixed format attribute, establish type specimen image library.
Step A1: photograph, picture pick-up device include camera, video camera, tablet computer, mobile phone, camera, are imaged by it
The rock sample image of module acquisition should be colored, saves as unified format, such as should be * .jpg, image clearly does not require to shoot
Angle, shooting distance do not require image size.
Step A2: the image of prescribed form should be stored in different files according to different rock categories in step A1,
Type specimen image library should include the image of storage, storage path, picture number, the name of image sample, petrographic description including solid
The attribute for the formula that fixes is expressed.
Step B, standardized images library is established: right using Image Argument (image parameter to match with this algorithm)
Image in type specimen image library is pre-processed, and is converted to four-dimensional (sample including feature normalization, by the format of image
Number, high, wide, number of dimensions) the format of input model, the optimization based on GPU, form standardized images, establish standardized images
Library, and the image in standardized images library is divided into training set and test set at random.
Step B1: every rock sample image carries out characteristics of image normalizing in the type specimen image library generated to step A2
Change operation.Characteristics of image normalization is to find one group of parameter using the not bending moment of image, can eliminate other transforming function transformation functions
Influence to image transformation.Unique canonical form is namely converted images into, the influence of geometry deformation can be resisted, is increased
Recognition accuracy of the discrimination model to similar image, formula are as follows: output=(pixel value-picture pixels value average value)/
(variances of picture pixels).
Step B2: the format conversion operation of image is carried out to the image that step B1 is generated.Image format conversion refer to by
The format of image is converted to the format of four-dimensional input model, and respectively (dimension refers to picture for sample number, height, width, number of dimensions
Color channel).
Step B3: the operation of the optimization based on GPU is carried out to the step B2 image generated.Optimization based on GPU, which refers to, to be schemed
Float as being converted to float32 facilitates and carries out GPU acceleration operation.
Step B4: the standardized images that all step B3 are generated are ordered together with storage path, picture number, image sample
Name, petrographic description include that standardized images library is established in the attribute expression of fixed format.
Step B5: the image of every kind of rock type in standardized images library that step B4 is generated is randomly divided by 1:1
For training the training set of discrimination model and for verifying the test set for differentiating effect.
Step C, characteristics of image is extracted: every for standardized images library for convenience of the quick identification of computer and mobile terminal
Image in one rocks training set is established convolutional neural networks MobileNet-V2 and (is belonged in lightweight convolutional neural networks
A kind of existing model) convolution model, pass through the Conv (discrete convolution operation) and MaxPool of the model (maximum pondization operation)
And first algorithm core formula (Leaky_SELU) extract the feature of image after each Zhang Yanshi standardization, and it is feature is unified
It is stored in fixed format, in fixed route file, by geometry identification, fuzzy diagnosis, model identification and at random identifies successively
Feature extraction obtains preliminary texture reconstruct characteristic function.
Step C1: MobileNet-V2 convolution model is established.The composition of MobileNet-V2 convolution model is as shown in the table:
Dimension therein refers to the color channel of picture, and step-length refers to the distance that convolution kernel moves every time, the dimension after expansion
Degree refers to the new color channel number of the constant longitudinal intensification of length and width.
The step uses the first core algorithm formula, and it is more stable can to provide patent than before in each step of calculating
Canonical formula, to show be mean value with 0,1 for variance sample distribution,
The problem of gradient explosion and gradient disappearance are effectively prevented for such extensive sample training problem.
First core algorithm formula (referred to as Leaky_SELU) is as follows:
Wherein λ is regular terms, and x is input item, and b is bias term,
C is empirical when rock characteristic learns, and α is learning rate, and e is constant.
Step C2: the training set chosen in step B5 is substituted into the MobileNet-V2 convolution model that step C1 is established, is led to
The discrete convolution and maximum pondization operation for crossing the model extract the feature of each picture, and complete geometry identification, a mould
The successively feature extraction that paste identification, model are identified and identified at random obtains preliminary texture reconstruct characteristic function.The texture of image
Reconstruct is come shield portions characteristics of image by the way of a part of random shielded image, and same image generates multiple block
The copy of different piece, to guarantee that the minutia of image local preferably can be saved and extract, and different copies is still
Can recognize that the due common trait of original rocks institute, an especially picture texture block at random be split as multiple thin portions it
Afterwards, the thin portion of texture is special before still retaining after new algorithm of the present invention processing by randomly assigne again spliced new texture
Sign.Texture reconstruct in the present invention passes through a geometry identification, fuzzy diagnosis, circulation model identification and identified at random
Feature extraction goes to realize, the first texture reconstruct number of same image does not set the circulation upper limit, be such as set as at least more than 10 times
100 times.
Step D, training discrimination model: by the discrete convolution of the MobileNet-V2 convolution model of training set and maximum pond
The result of operation output expands into one-dimensional data, adds shot and long term memory models using stored feature as the function learnt
As the input of hidden layer in LSTM, by the way of backpropagation, model is trained, LSTM model passes through geometry, mould
The superposition of paste, model, random training characteristics result, Process fusion Leaky_SELU and LSLE algorithm, finally obtains new reconstruct
The characteristic function of texture chooses preferred plan with this, generates discrimination model, and process is as shown in Figure 2.In Fig. 2, a kind of reconstruct
The method of texture recognition lithology includes image-forming module, storage system, learning system, parameter system and textured recombination function
Five part of identifying system, final output.Image-forming module is for acquiring rock image and extracting characteristics of image.Stocking system is used
According to fixed format store various types image, written historical materials.Learning system is for controlling convolutional neural networks MobileNet-V2
The process of convolution model study rock characteristics of image.Parameter system is for managing convolutional neural networks MobileNet-V2 convolution mould
Parameter of the type after training updates.Identifying system contains geometry identification, fuzzy diagnosis, model identification and random identification, uses
In carrying out the first differentiation of rock lithologic properties to having been subjected to pretreated unknown rock characteristics of image, and followed using four identification
Ring completes image texture reconstruct, final output comparison result.
Step D1: by result that step C2 is trained every time successively with 1. geometry (by freely being converted to image geometry space,
Known according to object in the mobile mathematical feature, including translation, offset, rotation, control with changed scale etc. in transformation with invariance in space
Feature Dui Bi not extracted), 2. obscure that (image is converted gray scale, extracts the related operation of texture information and is incorporated into a texture
In unit, similarity between texture being carried out characteristic quantification and the individual measured), 3. model is (to same photo rock texture
Cutting reassembles into new picture and carries out feature extraction using conventional histogramming algorithm after fragmentating), it is 4. random (at random to can
A same type of wherein picture for energy uses SIFT algorithm (i.e. Scale invariant features transform " Scale-invariant
The image characteristic extracting method of feature transform ") carry out feature extraction) four kinds the knot obtained is identified to characteristics of image
Fruit carries out feature superposition, and a process as reconstituted texture obtains preliminary characteristic function, is launched into one-dimensional data and is added to
Input in shot and long term memory models LSTM (Long Short-Term Memory) as hidden layer, in a manner of backpropagation
LSTM model is trained.Leaky_SELU and LSLE algorithm is merged simultaneously, obtains new image training characteristics, i.e., new rock
Stone image texture reconstructs characteristic function.As shown in figure 3, reconstructing characteristic function for building texture, convolution is established by training set image
Neural network MobileNet-V2 convolution model, and carry out discrete convolution and the operation of maximum pondization, through geometry, obscure, model, with
Four kinds of feature recognition algorithms of machine complete first texture reconstruct, and the result of output is expanded into one-dimensional data, stored spy
It levies as the input in the function addition shot and long term memory models LSTM learnt as hidden layer, using the side of backpropagation
Formula is trained model.The accuracy that identification feature is improved in training process, uses and has merged Leaky_
SELU and LSLE algorithm, training result are new rock image texture reconstruct characteristic function.
Wherein back-propagation algorithm is mainly by two links (excitation is propagated, weight updates) iterative cycles iteration, until net
Until the response to input of network reaches scheduled target zone.The learning process of shot and long term memory models LSTM is by forward-propagating
Process and back-propagation process composition.During forward-propagating, input information is by input layer through hidden layer, and successively processing is simultaneously
It is transmitted to output layer.If cannot get desired output valve in output layer, take the quadratic sum of output and desired error as mesh
Scalar functions, are transferred to backpropagation, successively find out objective function to the partial derivative of each neuron weight, constitute objective function to weight
The ladder amount of vector, as the foundation of modification weight, the study of network is completed during weight is modified.Error reaches desired value
When, e-learning terminates.
Step D2: training and the loss curve predicted in observation of steps D1, if the penalty values of training are much smaller than prediction
Penalty values then needing to increase the quantity of picture or adding regularization parameter λ on loss function, and modify rock characteristic
Empirical c when study.If the penalty values of training need to deepen the depth of model much larger than the penalty values of prediction.
Then think that model training is completed until the result of test set output reaches certain accuracy rate.
Step E, optimize discrimination model: the method that training process uses learning-rate decay (learning rate decaying),
To improve the training speed of entire model.In the training process of backpropagation, to improve model to the discrimination of detail characteristic,
Using the penalty values in the prediction training of LSLE algorithm, and come constantly using a kind of Adam (algorithms most in use of deep learning) algorithm
Reduction penalty values greatly improve the output accuracy of model thus Optimized model.
Step E1: in step C and step D, the method that training process uses learning-rate decay, to improve
The training speed of entire model.The formula of learning-rate decay method is decaying learning rate=learning rate × attenuation rate
× (current period/damped cycle), such as: frequency of training can be set as 1000 times, damped cycle 100 is arranged trained
Batch-size (existing a kind of batch of dimensional parameters) is 64, i.e., 64 pictures of selection are trained every time, and specific training is wanted
Ask to set according to actual needs.
Step E2: it in the backpropagation training process of step D1, is predicted using the second core LSLE algorithm in training
Penalty values.Because rock characteristic can not embody, when learning similar sample, the difficulty of similar study is specially enhanced
Degree, this difficulty is more greatly than inhomogeneous difficulty.It is such to treat with a certain discrimination so that the ga s safety degree of feature enhances.So the
Two core algorithms exist
On the basis of (reference: Liu, Weiyang, et al. " Large-Margin Softmax Loss for Convolutional
Neural Networks. " ICML.2016.) it joined more feasible added value:
I.e. new core algorithm LSLE.
Second core algorithm formula is as follows:
Wherein LiFor the penalty values for predicting i-th, Li=0 indicates that judgement is completely correct;
When sample m quantity is bigger, the boundary of classification is bigger, and learning difficulty is certainly higher.Meanwhile the D (θ) in formula must be one
A monotonous descending function and D (θ)=cos (π/m) D (m)=cos (π/m), to guarantee that ψ (θ) is a continuous function.WyiIt is i-th
The output weight of feature, xiIt is the input of i-th of classification, the when regularization coefficient that λ is represented can for constraining the range of formula
To provide more reliable penalty values and precision, will not occur violent prediction fault because of slightly changing for picture, wherein damaging
Lose value metric is the difference between predicted value and true value.
Step E3: (a kind of using Adam in order to constantly reduce penalty values in the backpropagation training process of step D1
The algorithms most in use of deep learning) algorithm optimizes.The algorithm is optimized according to the result and actual result that acquire, thus
Reduce penalty values.
Step F, characteristic matching exports result: for unknown rock image, substituting into discrimination model and extracts its thin portion spy
Sign, and Large-Margin Softmax (existing image similarity mode algorithm) algorithm is utilized, output newly acquires to be matched
The possible rock category of unknown rock probability, and the maximum rock category of select probability as unique output as a result, completing
Identification process.Its characteristic matching and result output process are as shown in Figure 1, pre-process the rock image obtained by image-forming module
Feature will be entered identifying system, then be merged in Leaky_SELU and LSLE algorithm and standard picture library by reconstituted texture
Image compares.If similarity reaches requirement, then the similarity highest being matched in outputting standard image library by comparison
Rock image and corresponding rock properties, and new picture is added in the standard picture library of the lithology and is stored.If similarity
It is unable to reach requirement, then needs to be manually entered the attribute of new lithology, and mark is included in as new sample by study module
In quasi- image library, stored.
Step F1: for the unknown rock image obtained by image-forming module, being pre-processed using step B, is formed
Standardized images.
Step F2: by the standardized images of the obtained unknown rock image of F1, it is input to the differentiation mould that step D is trained
Type seeks the class probability of unknown rock using the addition operation of LSLE algorithm, and the maximum classification of select probability is as differentiation knot
Fruit output.
The technical effect invented in order to better understand, with a kind of rock portable intelligent recognition methods (number of patent application
CN201610370185.4 it) compares, the method for the present invention is using intelligence degree, arithmetic speed, identification precision, training speed
There are bigger innovation, raising and change in terms of rate, is shown in Table 1.
A kind of rock portable intelligent recognition methods (number of patent application CN201610370185.4) of table 1
Compared with the method for the present invention
It is other it is unspecified be the prior art.
Claims (7)
1. a kind of method of reconstituted texture identification lithology, includes the following steps:
Step A, establish type specimen image library: using photograph, photographic equipment to the rock sample of different kinds of rocks carry out multi-angle,
Multiple spurs from shooting, and complete number, name, image size, typing and petrographic description, it is unified picture format, classified storage, defeated
After entering fixed format attribute, type specimen image library is established;
Step B, it establishes standardized images library: the image in type specimen image library being located in advance using Image Argument
Reason is converted to the format of four-dimensional input model including feature normalization, by the format of image, based on the optimization of GPU, forms mark
Standardization image establishes standardized images library, and the image in standardized images library is divided into training set and test set at random;
Step C, characteristics of image is extracted: for convenience of the quick identification of computer and mobile terminal, one kind every for standardized images library
Image in rock training set establishes convolutional neural networks MobileNet-V2 convolution model, by the Conv of the model and
MaxPool and the first algorithm core formula Leaky_SELU extracts the feature of image after each Zhang Yanshi standardization, and will be special
Sign is uniformly stored in fixed format, in fixed route file, passes through geometry identification, fuzzy diagnosis, model identification and identifies at random
Successively feature extraction, obtain preliminary texture reconstruct characteristic function;
Step D, training discrimination model: by the discrete convolution of the MobileNet-V2 convolution model of training set and maximum pondization operation
The result of output expands into one-dimensional data, adds shot and long term memory models LSTM using stored feature as the function learnt
The middle input as hidden layer is trained model by the way of backpropagation, and LSTM model is by geometry, fuzzy, mould
It is special to finally obtain new reconstituted texture for the superposition of type, random training characteristics result, Process fusion Leaky_SELU and LSLE algorithm
Function is levied, preferred plan is chosen with this, generates discrimination model;
Step E, optimize discrimination model: the method that training process uses learning-rate decay, to improve entire model
Training speed;In the training process of backpropagation, to improve model to the discrimination of detail characteristic, using Large-Margin
Penalty values in the prediction training of Softmax Loss the second core algorithm of Expand, and constantly dropped using Adam algorithm
Low penalty values, so that Optimized model, improves the output accuracy of model;
Step F, characteristic matching exports result: for unknown rock image, substitutes into discrimination model and extract its detail characteristic, and
And Large-Margin Softmax algorithm is utilized, output newly acquires the general of the possible rock category of unknown rock to be matched
Rate, and the maximum rock category of select probability is exported as unique as a result, completing identification process.
2. a kind of method of reconstituted texture identification lithology according to claim 1, which is characterized in that in the step A into
One step includes:
Step A1: photograph, photographic equipment include camera, video camera, tablet computer, mobile phone, camera, pass through its image-forming module
The rock sample image of acquisition should be colored, save as unified format;
Step A2: the image of prescribed form should be stored in different files according to different rock categories in step A1, original
Sample image library should include the image of storage, storage path, picture number, image sample is named, petrographic description includes fixed grating
The attribute of formula is expressed.
3. a kind of method of reconstituted texture identification lithology according to claim 1, which is characterized in that in the step B into
One step includes:
Step B1: every rock sample image carries out characteristics of image normalization behaviour in the type specimen image library generated to step A2
Make, characteristics of image normalization is to find one group of parameter using the not bending moment of image, can eliminate other transforming function transformation functions to figure
As the influence of transformation;Unique canonical form is namely converted images into, the influence of geometry deformation can be resisted, increases and differentiates
Recognition accuracy of the model to similar image, formula are as follows: output=(pixel value-picture pixels value average value)/(picture
The variance of pixel);
Step B2: the format conversion operation of image is carried out to the image that step B1 is generated.The format conversion of image refers to image
Format be converted to the format of four-dimensional input model, respectively sample number, height, width, number of dimensions;
Step B3: the operation of the optimization based on GPU is carried out to the step B2 image generated.Optimization based on GPU, which refers to, turns image
It is changed to the float of float32, facilitates and carries out GPU acceleration operation;
Step B4: the standardized images that all step B3 are generated, together with storage path, picture number, the name of image sample, rock
Property description include fixed format attribute expression establish standardized images library;
Step B5: the image of every kind of rock type in standardized images library that step B4 is generated is randomly divided by 1:1 and is used for
Train the training set of discrimination model and for verifying the test set for differentiating effect.
4. a kind of method of reconstituted texture identification lithology according to claim 1, which is characterized in that in the step C into
One step includes:
Step C1: MobileNet-V2 convolution model is established.The composition of MobileNet-V2 convolution model is as shown in the table:
Dimension therein refers to the color channel of picture, and step-length refers to the distance that convolution kernel moves every time, and the dimension after expansion is
Refer to the new color channel number of the constant longitudinal intensification of length and width;
The step uses new algorithm core formula, and canonical formula more stable than before can be provided in each step of calculating, is in
To reveal be mean value with 0,1 for variance sample distribution, gradient explosion is effectively prevented for such extensive sample training problem
The problem of disappearing with gradient;
First algorithm core formula Leaky_SELU is as follows:
Wherein λ is regular terms, and x is input item, and b is bias term,
C is empirical when rock characteristic learns, and α is learning rate, and e is constant;
Step C2: the training set chosen in step B5 is substituted into the MobileNet-V2 convolution model that step C1 is established, by this
The discrete convolution of model and maximum pondization operation extract the feature of each picture, and complete a geometry identification, fuzzy knowledge
Not, the successively feature extraction that model is identified and identified at random obtains preliminary texture reconstruct characteristic function;The texture of image reconstructs
It is that same image generation is multiple to block difference come shield portions characteristics of image by the way of a part of random shielded image
Partial copy, to guarantee that the minutia of image local preferably can be saved and extract, and different copies remains to know
Not Chu the due common trait of original rocks institute, an especially picture texture blocks at random to be split as after multiple thin portions, is pressed
Randomly assigne again spliced new texture by new algorithm of the present invention processing after still retain before texture detail characteristic.
5. a kind of method of reconstituted texture identification lithology according to claim 1, which is characterized in that in the step D into
One step includes:
Step D1: the result that step C2 is trained every time successively with 1. geometry, i.e., by freely being converted to image geometry space, root
According to object in space movement and the mathematical feature with invariance in transformation, including translation, offset, rotation, control with changed scale identification pair
Than extracting feature, 2. obscuring, i.e., image is converted gray scale, extracts the related operation of texture information and is incorporated into a texture cell
Interior, similarity, 3. model between texture being carried out characteristic quantification and the individual measured cut same photo rock texture
New picture is reassembled into after fragmentating and feature extraction, 4. random is carried out using conventional histogramming algorithm, i.e., at random to may
A same type of wherein picture using SIFT algorithm carry out feature extraction;Four kinds identify the result obtained to characteristics of image
Feature superposition is carried out, a process as reconstituted texture obtains preliminary characteristic function, is launched into one-dimensional data and is added to length
Input in short-term memory model LSTM as hidden layer, is trained LSTM model in a manner of backpropagation;Melt simultaneously
Leaky_SELU and LSLE algorithm is closed, new image training characteristics are obtained, i.e., new rock image texture reconstructs characteristic function;
Wherein back-propagation algorithm is mainly by two link iterative cycles iteration, until response of the network to input reaches scheduled
Until target zone, the learning process of shot and long term memory models LSTM is made of forward-propagating process and back-propagation process;?
During forward-propagating, input information, through hidden layer, is successively handled by input layer and is transmitted to output layer;If obtained in output layer
Less than desired output valve, then the quadratic sum of output and desired error is taken to be transferred to backpropagation as objective function, successively ask
Objective function constitutes objective function and measures to the ladder of weight vector, as modification weight to the partial derivative of each neuron weight out
The study of foundation, network is completed during weight is modified;When error reaches desired value, e-learning terminates.
Step D2: training and the loss curve predicted in observation of steps D1, if the penalty values of training are much smaller than the loss of prediction
Value, then needing to increase the quantity of picture or adding regularization parameter λ on loss function;If the penalty values of training are remote
Greater than the penalty values of prediction, then needing to deepen the depth of model, and empirical c when rock characteristic study is modified;Until
The result of test set output reaches certain accuracy rate and then thinks that model training is completed.
6. a kind of method of reconstituted texture identification lithology according to claim 1, which is characterized in that in the step E into
One step includes:
Step E1: in step C and step D, the method that training process uses learning-rate decay is entire to improve
The training speed of model, the formula of learning-rate decay method be decaying learning rate=learning rate × attenuation rate × (when
The preceding period/damped cycle),
Step E2: in the backpropagation training process of step D1, using Large-Margin Softmax Loss Expand
Two core algorithms come predict training in penalty values, it is such to treat with a certain discrimination so that the ga s safety degree of feature enhances;So second calculates
Method core formula exists
On the basis of joined more feasible added value:
Second algorithm core formula is as follows:
Wherein Li is the penalty values for predicting i-th, and Li=0 indicates that judgement is completely correct;When sample m quantity is bigger, the boundary of classification is bigger, and learning difficulty is certainly
It is higher;Meanwhile the D (θ) in formula must be a monotonous descending function and D (θ)=cos (π/m) D (m)=cos (π/m), to protect
Demonstrate,proving ψ (θ) is a continuous function;WyiThe output weight of ith feature, xi is the input of i-th of classification, λ represent when just
Then change coefficient, for constraining the range of formula, more reliable penalty values and precision can be provided, it will not be because of the slightly change of picture
It moves and violent prediction fault occurs, wherein lose value metric is the difference between predicted value and true value;
Step E3: it in the backpropagation training process of step D1, in order to constantly reduce penalty values, is carried out using Adam excellent
Change;The algorithm is optimized according to the result and actual result that acquire, to reduce penalty values.
7. a kind of method of reconstituted texture identification lithology according to claim 1, which is characterized in that in the step F into
One step includes:
Step F1: for the unknown rock image obtained by image-forming module, being pre-processed using step B, forms standard
Change image;
Step F2: by the standardized images of the obtained unknown rock image of F1, it is input to the discrimination model that step D is trained, is adopted
The class probability of unknown rock is sought with the addition operation of LSLE algorithm, the maximum classification of select probability is defeated as differentiation result
Out.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910016656.5A CN110119753B (en) | 2019-01-08 | 2019-01-08 | Lithology recognition method by reconstructed texture |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910016656.5A CN110119753B (en) | 2019-01-08 | 2019-01-08 | Lithology recognition method by reconstructed texture |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110119753A true CN110119753A (en) | 2019-08-13 |
CN110119753B CN110119753B (en) | 2023-05-23 |
Family
ID=67520253
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910016656.5A Active CN110119753B (en) | 2019-01-08 | 2019-01-08 | Lithology recognition method by reconstructed texture |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110119753B (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111026902A (en) * | 2019-12-20 | 2020-04-17 | 贵州黔岸科技有限公司 | Intelligent identification system and method for building material category |
CN111368931A (en) * | 2020-03-09 | 2020-07-03 | 第四范式(北京)技术有限公司 | Method and device for training image classification model, computer device and storage medium |
CN111709423A (en) * | 2020-06-26 | 2020-09-25 | 东北石油大学 | Rock debris lithology recognition method under PDC drill bit condition matched with lithology feature library |
CN112989708A (en) * | 2021-04-08 | 2021-06-18 | 重庆科技学院 | Well logging lithology identification method and system based on LSTM neural network |
CN113128404A (en) * | 2021-04-20 | 2021-07-16 | 中国地质科学院 | Intelligent mineral identification method and system |
CN113177949A (en) * | 2021-04-16 | 2021-07-27 | 中南大学 | Large-size rock particle feature identification method and device |
CN113362372A (en) * | 2021-05-25 | 2021-09-07 | 同济大学 | Single target tracking method and computer readable medium |
RU2760105C1 (en) * | 2021-03-12 | 2021-11-22 | Публичное акционерное общество «Газпром нефть» | System, machine-readable media and method for core analysis by images |
CN115641455A (en) * | 2022-09-16 | 2023-01-24 | 杭州视图智航科技有限公司 | Image matching method based on multi-feature fusion |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101719286A (en) * | 2009-12-09 | 2010-06-02 | 北京大学 | Multiple viewpoints three-dimensional scene reconstructing method fusing single viewpoint scenario analysis and system thereof |
US20160124115A1 (en) * | 2014-10-23 | 2016-05-05 | Chevron U.S.A. Inc. | System and Method of Pore Type Classification for Petrophysical Rock Typing |
CN106055636A (en) * | 2016-05-30 | 2016-10-26 | 长江岩土工程总公司(武汉) | Portable intelligent identification method for rock |
CN107590824A (en) * | 2017-09-19 | 2018-01-16 | 北京科技大学 | Rock particles identification and displacement tracking method based on 3-D view treatment technology |
CN107633255A (en) * | 2017-08-11 | 2018-01-26 | 天津大学 | A kind of rock lithology automatic recognition classification method under deep learning pattern |
CN108253938A (en) * | 2017-12-29 | 2018-07-06 | 武汉大学 | TBM broken rock slags digital close range photogrammetry identifies and inversion method |
-
2019
- 2019-01-08 CN CN201910016656.5A patent/CN110119753B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101719286A (en) * | 2009-12-09 | 2010-06-02 | 北京大学 | Multiple viewpoints three-dimensional scene reconstructing method fusing single viewpoint scenario analysis and system thereof |
US20160124115A1 (en) * | 2014-10-23 | 2016-05-05 | Chevron U.S.A. Inc. | System and Method of Pore Type Classification for Petrophysical Rock Typing |
CN106055636A (en) * | 2016-05-30 | 2016-10-26 | 长江岩土工程总公司(武汉) | Portable intelligent identification method for rock |
CN107633255A (en) * | 2017-08-11 | 2018-01-26 | 天津大学 | A kind of rock lithology automatic recognition classification method under deep learning pattern |
CN107590824A (en) * | 2017-09-19 | 2018-01-16 | 北京科技大学 | Rock particles identification and displacement tracking method based on 3-D view treatment technology |
CN108253938A (en) * | 2017-12-29 | 2018-07-06 | 武汉大学 | TBM broken rock slags digital close range photogrammetry identifies and inversion method |
Non-Patent Citations (3)
Title |
---|
YUJUAN SUN ET AL.: "3D reconstruction from single texture image based on patch matching and optimization", 《2015 ASIA-PACIFIC SIGNAL AND INFORMATION PROCESSING ASSOCIATION ANNUAL SUMMIT AND CONFERENCE》 * |
吉春旭: "基于卷积神经网络的岩石图像分类研究与应用", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
李发生等: "一种快速的三维模型纹理重构方法", 《第七届全国虚拟现实与可视化学术会议》 * |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111026902A (en) * | 2019-12-20 | 2020-04-17 | 贵州黔岸科技有限公司 | Intelligent identification system and method for building material category |
CN111026902B (en) * | 2019-12-20 | 2023-06-16 | 贵州黔岸科技有限公司 | Intelligent identification system and method for building material category |
CN111368931A (en) * | 2020-03-09 | 2020-07-03 | 第四范式(北京)技术有限公司 | Method and device for training image classification model, computer device and storage medium |
CN111368931B (en) * | 2020-03-09 | 2023-11-17 | 第四范式(北京)技术有限公司 | Method for determining learning rate of image classification model |
CN111709423A (en) * | 2020-06-26 | 2020-09-25 | 东北石油大学 | Rock debris lithology recognition method under PDC drill bit condition matched with lithology feature library |
CN111709423B (en) * | 2020-06-26 | 2022-03-15 | 东北石油大学 | Rock debris lithology recognition method under PDC drill bit condition matched with lithology feature library |
RU2760105C1 (en) * | 2021-03-12 | 2021-11-22 | Публичное акционерное общество «Газпром нефть» | System, machine-readable media and method for core analysis by images |
CN112989708A (en) * | 2021-04-08 | 2021-06-18 | 重庆科技学院 | Well logging lithology identification method and system based on LSTM neural network |
CN113177949B (en) * | 2021-04-16 | 2023-09-01 | 中南大学 | Large-size rock particle feature recognition method and device |
CN113177949A (en) * | 2021-04-16 | 2021-07-27 | 中南大学 | Large-size rock particle feature identification method and device |
CN113128404A (en) * | 2021-04-20 | 2021-07-16 | 中国地质科学院 | Intelligent mineral identification method and system |
CN113362372A (en) * | 2021-05-25 | 2021-09-07 | 同济大学 | Single target tracking method and computer readable medium |
CN115641455A (en) * | 2022-09-16 | 2023-01-24 | 杭州视图智航科技有限公司 | Image matching method based on multi-feature fusion |
CN115641455B (en) * | 2022-09-16 | 2024-01-09 | 杭州视图智航科技有限公司 | Image matching method based on multi-feature fusion |
Also Published As
Publication number | Publication date |
---|---|
CN110119753B (en) | 2023-05-23 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110119753A (en) | A kind of method of reconstituted texture identification lithology | |
CN110717553B (en) | Traffic contraband identification method based on self-attenuation weight and multiple local constraints | |
CN108985376A (en) | It is a kind of based on convolution-Recognition with Recurrent Neural Network rotary kiln sequence operating mode's switch method | |
CN110348399B (en) | Hyperspectral intelligent classification method based on prototype learning mechanism and multidimensional residual error network | |
Mathur et al. | Crosspooled FishNet: transfer learning based fish species classification model | |
CN109902018B (en) | Method for acquiring test case of intelligent driving system | |
CN111339935B (en) | Optical remote sensing picture classification method based on interpretable CNN image classification model | |
CN111291639B (en) | Cross-source ship feature fusion learning and identifying method based on layered variation self-coding | |
CN108399420B (en) | Visible light ship false alarm rejection method based on deep convolutional network | |
Alajaji et al. | Few shot scene classification in remote sensing using meta-agnostic machine | |
CN110991257B (en) | Polarized SAR oil spill detection method based on feature fusion and SVM | |
CN110390347A (en) | Conditions leading formula confrontation for deep neural network generates test method and system | |
CN113408340B (en) | Dual-polarization SAR small ship detection method based on enhanced feature pyramid | |
CN109582960B (en) | Zero example learning method based on structured association semantic embedding | |
CN114842343A (en) | ViT-based aerial image identification method | |
CN114511710A (en) | Image target detection method based on convolutional neural network | |
CN111126155B (en) | Pedestrian re-identification method for generating countermeasure network based on semantic constraint | |
CN115311508A (en) | Single-frame image infrared dim target detection method based on depth U-type network | |
CN114821200B (en) | Image detection model and method applied to industrial vision detection field | |
CN113313185B (en) | Hyperspectral image classification method based on self-adaptive spatial spectrum feature extraction | |
CN112052881B (en) | Hyperspectral image classification model device based on multi-scale near-end feature splicing | |
CN108647784A (en) | A kind of lifelong machine learning method based on depth belief network | |
CN117315381B (en) | Hyperspectral image classification method based on second-order biased random walk | |
Zhao et al. | Research on Traffic Sign Detection Algorithm Based on Improved YOLOv3 | |
Li et al. | Intelligent Detection of Marine Organisms with Deep Learning based on Attention Mechanism |
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 | ||
CB02 | Change of applicant information | ||
CB02 | Change of applicant information |
Address after: 430010 No. 1863 Jiefang Avenue, Jiang'an District, Hubei, Wuhan Applicant after: Changjiang Geotechnical Engineering Co.,Ltd. Applicant after: CHANGJIANG SURVEY PLANNING DESIGN AND RESEARCH Co.,Ltd. Address before: 430010 No. 1863 Jiefang Avenue, Jiang'an District, Hubei, Wuhan Applicant before: CHANGJIANG GEOTECHNICAL ENGINEERING Corp. Applicant before: CHANGJIANG SURVEY PLANNING DESIGN AND RESEARCH Co.,Ltd. |
|
GR01 | Patent grant | ||
GR01 | Patent grant |