CN109376695A - A kind of smog detection method based on depth hybrid neural networks - Google Patents
A kind of smog detection method based on depth hybrid neural networks Download PDFInfo
- Publication number
- CN109376695A CN109376695A CN201811421607.1A CN201811421607A CN109376695A CN 109376695 A CN109376695 A CN 109376695A CN 201811421607 A CN201811421607 A CN 201811421607A CN 109376695 A CN109376695 A CN 109376695A
- Authority
- CN
- China
- Prior art keywords
- layer
- neural networks
- neural network
- depth
- value
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
-
- 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
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention relates to a kind of smog detection methods based on depth hybrid neural networks, and depth normalization neural network and a kind of completely new convolutional neural networks based on the company of jump are combined into depth hybrid neural networks by the method for Fusion Features.The smog detection method realized by depth hybrid neural networks, the smog that can be emerged to factory chimney, torch identify, judge smog for white cigarette, non-white cigarette.Due to the ratio difference for the toxic harmful exhaust gas that different types of smog contains, it need to be judged in time to control.Reducing the parameter in a large amount of neural networks compared with existing method by exhaust treatment system is applied to based on the smog detection method of depth hybrid neural networks, improving recognition accuracy while improving arithmetic speed.Can generation to exhaust gas and discharge process carry out accurate real-time control, can not only substantially reduce the discharge of toxic and harmful gas, while also reducing energy consumption.
Description
Technical field
The smog detection method that the present invention utilizes depth hybrid neural networks to be realized, can emit factory chimney, torch
Smog out identified, judge smog for white cigarette or non-white cigarette, the toxic harmful exhaust gas contained due to different types of smog
Ratio it is different, it need to be judged in time to control.Smog detection method based on depth hybrid neural networks is answered
For exhaust treatment system, generation and discharge process to exhaust gas carry out accurate real-time control, can not only substantially reduce
The discharge of malicious pernicious gas, while also reducing energy consumption.Image knowledge was both belonged to Smoke Detection based on depth hybrid neural networks
Other field, and belong to artificial intelligence field.
Background technique
In recent years, China has been devoted to protection environment, and energy-saving and emission-reduction reduce atmosphere pollution.Wherein, " environment is empty for new edition
Gas quality standard " and be for the discharge standard of the industries such as thermoelectricity, steel, cement, chemical industry and non-electrical coal-burning boiler respectively
Promote the most strong policy of each enterprise implement smoke treated engineering.But since the industries such as traditional thermoelectricity, chemical industry are in China's warp
Occupy biggish specific gravity in Ji, the discharge amount of exhaust gas is still very big, and lacks the technology discharged to exhaust gas and implement real-time control
Means are especially difficult to distinguish smog type, lead to air pollution problems inherent still and annoying many enterprises.
When coal and oil combustion, burning waste gas is mainly nitrogen, nitrogen oxides, CO2、O2And vapor, gas are colourless
, even if there is a small amount of dust together with smog and Water vapor condensation, gas also can be white.Why do not formed white
Cigarette is primarily due to smog and contains solid carbon, caused by the imperfect combustion often caused by various fuel gas.In non-white cigarette
Contain a large amount of NO2、SO2、O3、CO、PM2.5、PM10Equal toxic and harmful gas, so carrying out effective identification pair to two kinds of smog
In the control that combustion process and exhaust gas are discharged to which the discharge for reducing toxic and harmful gas is significant.
The method of traditional smoke detection relies primarily on artificial observation or sensor.But since human resources are limited, cost compared with
Height, the method based on artificial observation cannot fast and effeciently monitor smog for a long time.On the other hand, due to the influence of environmental change,
Smoke sensor device based on smoke particle sampling or relative humidity sampling is also likely to will appear serious time lag, while also not
Detection zone can be completely covered.In general, existing smog detection method is difficult to meet demand.In this regard, using traditional figure
Method, the method for extraction image shape feature and colouring information as recognition methods, such as using image local histogram variances
And smog is detected based on the support vector machine method of wavelet analysis, but since the feature of smog image is difficult to extract, this
The effect is unsatisfactory for kind method.
In recent years, significant progress is achieved using the technology that convolutional neural networks carry out image recognition, especially with
The raising of modern computer operational capability can be efficiently extracted using depth convolutional neural networks by the study to great amount of samples
Clarification of objective is to realize accurate image recognition.Z.Yin, B.Wan, F.Yuan, X.Xia, J.Sh propose a kind of depth
Normalization convolutional neural networks (DNCNN) is simultaneously applied to Smoke Detection, achieves certain effect.But depth normalizes
Convolutional neural networks have that parameter is excessive, operation is slower, are unfavorable for timely and effectively detecting smog.We
Method reduces the parameter of depth normalization convolutional neural networks 4/5ths by depth hybrid neural networks, and identifies standard
True rate is higher than depth and normalizes convolutional neural networks, and artificial intelligence application is real in Smoke Detection, being conducive to carry out smog
When and correctly classify, thus effectively increase to burning and exhaust gas discharge control efficiency, reduce toxic and harmful gas
Discharge.
Summary of the invention
Present invention obtains a kind of smog detection methods based on depth hybrid neural networks, combine depth normalization volume
It is combined with a kind of completely new convolutional neural networks based on the company of jump, it is mixed to constitute depth by the advantages of product neural network
Neural network is closed, accurate, classification in real time is carried out to smog by the smog image of input to realize, is solved to smog
The problem that image is measured in real time.Accurately real-time control is carried out for the discharge process to combustion process and exhaust gas to create
Condition；
Present invention employs the following technical solution and realize step:
A kind of smog detection method based on depth hybrid neural networks, by judging the smog image of input,
Judge smog for white cigarette or non-white cigarette；
Characterized by comprising the following steps:
(1) it proposes and builds a kind of new depth hybrid neural networks；
Depth hybrid neural networks are that the convolutional neural networks and depth normalization neural network by being connected based on jump pass through spy
It is obtained combined by the mode for levying fusion, specific structure is shown in Fig. 1.Wherein, the convolutional neural networks based on the company of jump are a kind of complete
New neural network, its structure is shown in Fig. 2, specific as follows:
First layer is convolutional layer, includes 32 9 dimension convolution kernels, and convolution kernel moving step length is 1, uses ReLU activation primitive.
The second layer is convolutional layer, includes 64 1 dimension convolution kernels, and convolution kernel moving step length is 1, uses ReLU activation primitive.Third layer,
Four layers, layer 5 be normalization convolutional layer, include 64 3 dimension convolution kernels, convolution kernel moving step length be 1, activated using ReLU
Function.Layer 6 is normalization convolutional layer, includes 64 1 dimension convolution kernels, and convolution kernel moving step length is 1, activates letter using ReLU
Number.Layer 7 is pond layer, and pond range is 3 dimensions, moving step length 2, using maximum pond method.8th layer, the 9th layer is to return
One changes convolutional layer, includes 128 3 dimension convolution kernels, and convolution kernel moving step length is 1, uses ReLU activation primitive.Tenth layer is pond
Change layer, pond range is 3 dimensions, moving step length 2, using maximum pond method.Eleventh floor, Floor 12 are normalization convolution
Layer includes 256 3 dimension convolution kernels, and convolution kernel moving step length is 1, uses ReLU activation primitive.13rd layer is normalization
Convolutional layer includes 21 dimension convolution kernels, and convolution kernel moving step length is 1, uses ReLU activation primitive.14th layer is pond
Layer, using the average pond of the overall situation.
Neural network is normalized for depth, since its number of parameters focuses primarily upon the full articulamentum of second part, institute
To eliminate full articulamentum, by the output of its characteristic extraction part before full articulamentum and the convolutional Neural net based on the company of jump
Network has carried out the operation of Fusion Features between its Floor 12, the 13rd layer, to together constitute depth composite nerve net
Network.
Four concrete operations about depth hybrid neural networks are given below.
1. convolution and normalization operation.Convolution algorithm is carried out to input picture by the convolution kernel trained, then to convolution
Feature vector that operation extracts carries out batch normalized, is improved the training speed of network with this and is avoided falling into
The problem of fitting.To each feature vector fjIt is normalized to obtain normalization characteristic vectorFormula it is as follows:
In formula (1),WithRespectively batch mean value and batch variance, n are every
The dimension of a feature vector, fj,iFor j-th of feature of i-th of sample in sample batch, wherein i is to be less than or equal to more than or equal to 1
Any positive integer of total sample number, j be more than or equal to 1 be less than or equal to feature vector sum any positive integer, ε for a value very
Small any normal number, is set as 0.0001 herein.But normalization can be such that the representative power of each feature vector declines therefore, draw
Enter two free parameters, and is asked by ratio in batch standardization and shift step transfer standard feature to solve this
Topic, shown in specific conversion regime such as formula (2):
Wherein B (fj) it is the standardized feature being converted to, α, β initial value are random in formula, are obtained most by subsequent training
The figure of merit.
2. the company of jump operates.For the feature propagation of accelerans network, solve the problems, such as that gradient disappears and gradient is exploded, new
Based on the convolutional neural networks of the company of jump, the Feature Mapping F of the 1st layer of output is connected by the way of jump connection1With the 5th layer
The Feature Mapping F of output5, it is as follows that the company of jump operates used 1 dimension convolution kernel:
F6=max (0, ω6*[F1,F5]+b6) (3)
In above formula, [F1,F5] indicate F1And F5Two three-dimensional feature mapping matrixes are attached in third dimension.ω6With
b6Indicate the weight matrix and offset moment matrix of the 6th convolutional layer, the ginseng of weight matrix and offset moment matrix as the neural network
Number obtains optimal value by subsequent training, and initial value takes random value, F6The Feature Mapping obtained for layer 6.
3. the average pond of the overall situation.To reduce the parameter of depth hybrid neural networks, while over-fitting is avoided the problem that, in mind
Through the 13rd layer of network, 256 Feature Mappings are become into 2 features using the convolutional layer that the convolution kernel for being 1 from 2 dimensions is constituted
Mapping.This 2 Feature Mappings obtain the overall situation and are averaged pond value q by the 14th layer of the overall situation pond layer that be averagedt, the average pond of the overall situation
Change value is obtained by formula (4):
In formula (4), ps,tS-th of pixel value in t-th of Feature Mapping is represented, wherein s is to be less than more than or equal to 1 etc.
In any positive integer of pixel sum, due to there are two Feature Mappings, the value of t is { 1,2 }, and S is the total of whole pixels
Number.It provides image and is under the jurisdiction of the probability of u class and be calculated by following normalization exponential function:
In formula (5), QuTo provide the possibility that image is under the jurisdiction of u class, the type u incorporated into is { 1,2 }.Wherein v takes
Value is { 1,2 }, quIt is averaged pond value for the overall situation of u class, qvIt is averaged pond value for the overall situation of v class, e is natural constant.
4. Fusion Features.Herein by depth normalization neural network the resulting Feature Mapping of characteristic extraction part be based on
The Feature Mapping that the convolutional neural networks of the company of jump extract is merged.Concrete operations are by one 1 dimension convolution kernel to two
Feature Mapping carries out convolution algorithm, and the undetermined parameter of the 1 dimension convolution kernel as the neural network, initial value takes random value.
(2) data enhancing processing, both methods are carried out after first carrying out image normalization processing to training data set used
Concrete operations are as follows:
Pixel minimax method for normalizing is used to be more advantageous to feature to realize image normalization to protrude characteristics of image
It extracts, sees below formula (6).
In formula (6), ynFor the pixel value after normalized, yrFor original pixel value, ymaxAnd yminFor pixel value in image
Maxima and minima.
Data enhancing carries out white cigarette image horizontal aiming at the problem that white cigarette, non-white cigarette sample imbalance in training data
Overturning, flip vertical, 180 ° of operations overturn, concrete operations are shown in Fig. 3, are added to instruction for the image after overturning as new image
Practice in data set, increases the training sample number of less type.
(3) with by method shown in (2), treated that image is that input is trained neural network and its parameter；
Following three step is divided into for the training process of the neural network of proposition.Firstly, only training the convolution mind based on the company of jump
Through network.Second step merges the characteristic extraction part of depth normalization neural network and the convolutional neural networks based on the company of jump
For depth hybrid neural networks, freeze a part of parameter every time and only optimize remaining parameter, repeat this method until all parameters all
By training.Finally, bringing the parameter of depth hybrid neural networks whole into training together again, depth hybrid neural networks are obtained
Parameter.
Creativeness of the invention is mainly reflected in:
(1) present invention is difficult to the characteristics of extracting for the feature of smog image, proposes using the convolution mind based on the company of jump
Smog image is identified through network.By way of the company of jump, effective solution gradient in deep neural network disappears
The problem of with gradient explosion, feature propagation is accelerated, so that the characteristics of image extracted is more effective, neural metwork training effect is aobvious
It writes and improves；
(2) present invention devises global average pond layer, substantially reduces the number of parameters in depth hybrid neural networks,
The training speed of neural network is not only significantly improved, but also greatly improves the operation efficiency of image recognition, there is network
The features such as parameter is few, and operation is fast, and robustness is high；
(3) present invention employs the methods of Fusion Features normalizes convolutional neural networks from depth and based on the company of jump to merge
Convolutional neural networks in the feature that extracts, the advantages of combining depth normalization convolutional neural networks, to significantly improve
The accuracy rate of Smoke Detection.
Detailed description of the invention
Fig. 1 is depth hybrid neural networks structure chart of the present invention
Fig. 2 is that the present invention is based on the convolutional neural networks structure charts of the company of jump
Fig. 3 is data enhancement operations schematic diagram of the present invention
Fig. 4 is the present invention in 50 times to 300 times training, normalizes neural network with depth after each training of progress
(DNCNN), the recognition correct rate comparison of the convolutional neural networks (SCNN) based on the company of jump
Specific embodiment
Present invention obtains a kind of smog detection methods based on depth hybrid neural networks, combine depth normalization volume
It is combined with a kind of completely new convolutional neural networks based on the company of jump, it is mixed to constitute depth by the advantages of product neural network
Neural network is closed, accurate, classification in real time is carried out to smog by the smog image of input to realize, is solved to smog
The problem that image is measured in real time.Accurately real-time control is carried out for the discharge process to combustion process and exhaust gas to create
Condition；
Present invention employs the following technical solution and realize step:
1. proposing and building a kind of new depth hybrid neural networks；
Depth hybrid neural networks are that the convolutional neural networks and depth normalization neural network by being connected based on jump pass through spy
It is obtained combined by the mode for levying fusion, specific structure is shown in Fig. 1.Wherein, the convolutional neural networks based on the company of jump are a kind of complete
New neural network, its structure is shown in Fig. 2, specific as follows:
First layer is convolutional layer, includes 32 9 dimension convolution kernels, and convolution kernel moving step length is 1, uses ReLU activation primitive.
The second layer is convolutional layer, includes 64 1 dimension convolution kernels, and convolution kernel moving step length is 1, uses ReLU activation primitive.Third layer,
Four layers, layer 5 be normalization convolutional layer, include 64 3 dimension convolution kernels, convolution kernel moving step length be 1, activated using ReLU
Function.Layer 6 is normalization convolutional layer, includes 64 1 dimension convolution kernels, and convolution kernel moving step length is 1, activates letter using ReLU
Number.Layer 7 is pond layer, and pond range is 3 dimensions, moving step length 2, using maximum pond method.8th layer, the 9th layer is to return
One changes convolutional layer, includes 128 3 dimension convolution kernels, and convolution kernel moving step length is 1, uses ReLU activation primitive.Tenth layer is pond
Change layer, pond range is 3 dimensions, moving step length 2, using maximum pond method.Eleventh floor, Floor 12 are normalization convolution
Layer includes 256 3 dimension convolution kernels, and convolution kernel moving step length is 1, uses ReLU activation primitive.13rd layer is normalization
Convolutional layer includes 21 dimension convolution kernels, and convolution kernel moving step length is 1, uses ReLU activation primitive.14th layer is pond
Layer, using the average pond of the overall situation.
Neural network is normalized for depth, full articulamentum therein is eliminated, by its feature before full articulamentum
It extracts the output of part and the convolutional neural networks based on the company of jump has carried out the behaviour of Fusion Features between its 12nd layer, the 13rd layer
Make, to together constitute depth hybrid neural networks.
Four concrete operations about depth hybrid neural networks are given below.
(1) convolution and normalization operation.Convolution algorithm is carried out to input picture by the convolution kernel trained, then to convolution
The feature vector that operation extracts carries out batch normalized, and the number of samples of each training batch is set as 96 herein.
To each feature vector fjIt is normalized to obtain normalization characteristic vectorFormula it is as follows:
In formula (1),WithRespectively batch mean value and batch variance, n are every
The dimension of a feature vector, fj,iFor j-th of feature of i-th of sample in sample batch, wherein i is to be less than or equal to more than or equal to 1
Any positive integer of total sample number, j be more than or equal to 1 be less than or equal to feature vector sum any positive integer, ε for a value very
Small any normal number, is set as 0.0001 herein.Ratio is carried out to the feature extracted again and shifting function is converted, specifically
Shown in conversion regime such as formula (2):
Wherein B (fj) it is the standardized feature being converted to, α, β initial value are random in formula, obtain optimal value by training.
In third layer, the 4th layer, layer 5, layer 6, the 8th layer, the 9th layer, eleventh floor, Floor 12, the 13rd
Batch method for normalizing is used in layer convolutional layer.
(2) company of jump operates.The new convolutional neural networks based on the company of jump connect the 1st by the way of jump connection
The Feature Mapping F of layer output1The Feature Mapping F exported with the 5th layer5, it is as follows that the company of jump operates used 1 dimension convolution kernel:
F6=max (0, ω6*[F1,F5]+b6) (3)
In above formula, [F1,F5] indicate F1And F5Two three-dimensional feature mapping matrixes are attached in third dimension.ω6With
b6Indicate the weight matrix and offset moment matrix of the 6th convolutional layer, the ginseng of weight matrix and offset moment matrix as the neural network
Number obtains optimal value by subsequent training, and initial value takes random value, F6The Feature Mapping obtained for layer 6.
(3) global average pond.At the 13rd layer of neural network, the convolution that constitutes of convolution kernel for being 1 by 2 dimensions is utilized
256 Feature Mappings are become 2 Feature Mappings by layer.This 2 Feature Mappings are obtained by the 14th layer of the overall situation pond layer that is averaged
To the average pond value q of the overall situationt, the average pond value of the overall situation obtains by formula (4):
In formula (4), ps,tS-th of pixel value in t-th of Feature Mapping is represented, wherein s is to be less than more than or equal to 1 etc.
In any positive integer of pixel sum, due to there are two Feature Mappings, the value of t is { 1,2 }, and S is the total of whole pixels
Number.It provides image and is under the jurisdiction of the probability of u class and be calculated by following normalization exponential function:
In formula (5), QuTo provide the possibility that image is under the jurisdiction of u class, the type u incorporated into is { 1,2 }.Wherein v takes
Value is { 1,2 }, quIt is averaged pond value for the overall situation of u class, qvIt is averaged pond value for the overall situation of v class, e is natural constant.
(4) Fusion Features.Herein by the resulting Feature Mapping of characteristic extraction part and base of depth normalization neural network
It is merged in the Feature Mapping that the convolutional neural networks of the company of jump extract.Concrete operations are by one 1 dimension convolution kernel to two
A Feature Mapping carries out convolution algorithm, which obtains optimal value by subsequent training as the parameter of the neural network,
Initial value takes random value.
2. pair training data set used carries out data enhancing processing, both methods tool after first carrying out image normalization processing
Gymnastics is made as follows:
Image normalization is realized using pixel minimax method for normalizing, sees below formula (6).
In formula (6), ynFor the pixel value after normalized, yrFor original pixel value, ymaxAnd yminFor pixel value in image
Maxima and minima.
Data enhancing, training sample concentration share 2200 non-white cigarette images and 8500 white cigarette images, to white cigarette image into
Row flip horizontal, flip vertical, 180 ° of operations overturn, are added to training dataset for the image after overturning as new image
In, increase white cigarette image number, Fig. 3 is shown in concrete operations.
3. to be that input is trained neural network and its parameter by step 2 treated image.
It is trained that the specific method is as follows to the parameter of neural network:
It introduces Glorot and is uniformly distributed initial method to complete the initialization to weight each in neural network.Then,
Refer to the training method of momentum and learning rate decaying and stochastic gradient descent, momentum coefficient 0.9, learning rate attenuation coefficient
It is 0.0001, the random coefficient of stochastic gradient descent method is set as 0.9, and initial learning rate is 0.01, the batch sample of small lot
Number is 96, and trained loss function is cross entropy loss function.
Following three step is divided into for the training process of the neural network of proposition.Firstly, only training the convolution mind based on the company of jump
Through network, frequency of training 300.Second step, by the characteristic extraction part of depth normalization neural network and the volume connected based on jump
Product neural network merge into depth hybrid neural networks, freeze every time 80% parameter only optimize remaining 20% parameter, instruction
Practicing number is 300, repeats this method until all parameters are all by training.Finally, again by the ginseng of depth hybrid neural networks whole
Number brings training into together, after training 300 times, obtains the optimized parameter of depth hybrid neural networks.
The depth hybrid neural networks trained by above method to test data concentrate 18522 groups of white cigarette images and
19060 groups of non-white cigarette images amount to 37582 groups of data and carry out class test, and accuracy is up to 99.5%, at 50 times to 300 times
In training, the convolutional neural networks for normalizing neural network (DNCNN) with depth after each training, connecting based on jump are carried out
(SCNN), Fig. 4 is shown in the recognition correct rate comparison of depth hybrid neural networks (DMNN).
Claims (3)
1. a kind of smog detection method based on depth hybrid neural networks is sentenced by judging the smog image of input
Mist of giving up smoking is white cigarette or non-white cigarette, which comprises the following steps:
Step 1: proposing and building a kind of new depth hybrid neural networks；
Depth hybrid neural networks are the combinations of two neural networks, depth normalize neural network and it is a kind of it is completely new based on
The neural network of the company of jump；
Step 2: the smog image to input enhances algorithm process using image normalization and data；
Step 3: the parameter of depth hybrid neural networks is trained and optimized to substep, then the smog image of input is judged,
Judge smog for white cigarette or non-white cigarette；
In the first step:
Depth hybrid neural networks are that the neural network and depth normalization neural network by being connected based on jump pass through Fusion Features
It is obtained combined by mode；Wherein, the neural network based on the company of jump is a kind of completely new neural network, its structure is as follows:
First layer is convolutional layer, includes 32 9 dimension convolution kernels, and convolution kernel moving step length is 1, uses ReLU activation primitive；Second
Layer is convolutional layer, includes 64 1 dimension convolution kernels, and convolution kernel moving step length is 1, uses ReLU activation primitive；Third layer, the 4th
Layer, layer 5 are normalization convolutional layer, include 64 3 dimension convolution kernels, and convolution kernel moving step length is 1, activate letter using ReLU
Number；Layer 6 is normalization convolutional layer, includes 64 1 dimension convolution kernels, and convolution kernel moving step length is 1, uses ReLU activation primitive；
Layer 7 is pond layer, and pond range is 3 dimensions, moving step length 2, using maximum pond method；8th layer, the 9th layer is normalization
Convolutional layer includes 128 3 dimension convolution kernels, and convolution kernel moving step length is 1, uses ReLU activation primitive；Tenth layer is pond
Layer, pond range are 3 dimensions, moving step length 2, using maximum pond method；Eleventh floor, Floor 12 are normalization convolutional layer,
It include 256 3 dimension convolution kernels, convolution kernel moving step length is 1, uses ReLU activation primitive；13rd layer is normalization convolution
Layer includes 21 dimension convolution kernels, and convolution kernel moving step length is 1, uses ReLU activation primitive；14th layer is pond layer, is adopted
With the average pond of the overall situation；
Neural network is normalized for depth, full articulamentum therein is eliminated, by its feature extraction before full articulamentum
Partial output and the neural network based on the company of jump have carried out the operation of Fusion Features between its Floor 12, the 13rd layer,
To together constitute depth hybrid neural networks；
Four concrete operations about depth hybrid neural networks are given below；
(1) convolution and normalization operation；Convolution algorithm is carried out to input picture by the convolution kernel trained, then to convolution algorithm
The feature vector extracted carries out batch normalized, and the number of samples of each training batch is set as 96 herein；To every
A feature vector fjIt is normalized to obtain normalization characteristic vectorFormula it is as follows:
In formula (1),WithRespectively batch mean value and batch variance, n are each spy
Levy the dimension of vector, fj,iFor j-th of feature of i-th of sample in sample batch, wherein i is to be less than or equal to sample more than or equal to 1
Any positive integer of sum, j are any positive integer for being less than or equal to feature vector sum more than or equal to 1, and ε is a value very little
Any normal number；Ratio is carried out to the feature extracted again and shifting function is converted, specific conversion regime such as formula (2) institute
Show:
Wherein B (fj) it is the standardized feature being converted to, α, β initial value are random in formula, obtain optimal value by training；
In third layer, the 4th layer, layer 5, layer 6, the 8th layer, the 9th layer, eleventh floor, Floor 12, the tenth three-layer coil
Batch method for normalizing is used in lamination；
(2) company of jump operates；
The new neural network based on the company of jump connects the Feature Mapping F of the 1st layer of output by the way of jump connection1With
The Feature Mapping F of 5 layers of output5, it is as follows that convolution kernel is tieed up in cascade operation used 1:
F6=max (0, ω6*[F1,F5]+b6) (3)
In above formula, [F1,F5] indicate F1And F5Two three-dimensional feature mapping matrixes are attached in third dimension；ω6And b6Table
Show the 6th convolutional layer weight matrix and offset moment matrix, weight matrix and offset moment matrix as the neural network parameter by
Subsequent training obtains optimal value, and initial value takes random value, F6The Feature Mapping obtained for layer 6；
(3) global average pond；
At the 13rd layer of neural network, 256 Feature Mappings are become using the convolutional layer that the convolution kernel for being 1 by 2 dimensions is constituted
For 2 Feature Mappings；This 2 Feature Mappings obtain the overall situation and are averaged pond value q by the 14th layer of the overall situation pond layer that be averagedt,
The average pond value of the overall situation is obtained by formula (4):
In formula (4), ps,tS-th of pixel value in t-th of Feature Mapping is represented, wherein s is to be less than or equal to picture more than or equal to 1
Any positive integer of vegetarian refreshments sum, due to there are two Feature Mappings, the value of t is { 1,2 }, and S is the sum of whole pixels；It gives
The probability that image is under the jurisdiction of u class out is calculated by following normalization exponential function:
In formula (5), QuTo provide the possibility that image is under the jurisdiction of u class, the type u incorporated into is { 1,2 }；Wherein the value of v is
{ 1,2 }, quIt is averaged pond value for the overall situation of u class, qvIt is averaged pond value for the overall situation of v class, e is natural constant；
(4) Fusion Features；Herein by the resulting Feature Mapping of characteristic extraction part of depth normalization neural network and based on jump
The Feature Mapping that neural network even extracts is merged；Concrete operations are to be reflected by one 1 dimension convolution kernel to two features
Row convolution algorithm is injected, which obtains optimal value by subsequent training as the parameter of the neural network, and initial value takes
Random value.
2. a kind of smog detection method based on depth hybrid neural networks according to claim 1, characterized in that second
In step:
Data enhancing processing, both methods concrete operations are carried out after first carrying out image normalization processing to training data set used
It is as follows:
Image normalization is realized using pixel minimax method for normalizing, sees below formula (6)；
In formula (6), ynFor the pixel value after normalized, yrFor original pixel value, ymaxAnd yminFor the maximum of pixel value in image
Value and minimum value；
Data enhancing carries out flip horizontal, flip vertical, 180 ° of operations overturn to white cigarette image, the image after overturning is made
Training data is added to for new image to concentrate, and increases white cigarette image number.
3. a kind of smog detection method based on depth hybrid neural networks according to claim 1, characterized in that third
In step:
It is trained that the specific method is as follows to the parameter of neural network:
It introduces Glorot and is uniformly distributed initial method to complete the initialization to weight each in neural network；Then, it quotes
The training method of momentum and learning rate decaying and stochastic gradient descent, momentum coefficient 0.9, learning rate attenuation coefficient are
0.0001, the random coefficient of stochastic gradient descent method is set as 0.9, and initial learning rate is 0.01, the batch sample of small lot
Number is 96, and trained loss function is cross entropy loss function；
Following three step is divided into for the training process of the neural network of proposition；Firstly, the neural network based on the company of jump is only trained, instruction
Practicing number is 300；Second step merges the characteristic extraction part of depth normalization neural network and the neural network based on the company of jump
For depth hybrid neural networks, freeze every time 80% parameter only optimize remaining 20% parameter, frequency of training 300, weight
This multiple method is until all parameters all pass through training；Finally, bring the parameter of depth hybrid neural networks whole into training together again,
After training 300 times, the optimized parameter of depth hybrid neural networks is obtained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811421607.1A CN109376695B (en) | 2018-11-26 | 2018-11-26 | Smoke detection method based on deep hybrid neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811421607.1A CN109376695B (en) | 2018-11-26 | 2018-11-26 | Smoke detection method based on deep hybrid neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109376695A true CN109376695A (en) | 2019-02-22 |
| CN109376695B CN109376695B (en) | 2021-09-17 |
Family
ID=65377353
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811421607.1A Active CN109376695B (en) | 2018-11-26 | 2018-11-26 | Smoke detection method based on deep hybrid neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109376695B (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109815969A (en) * | 2019-03-05 | 2019-05-28 | 上海骏聿数码科技有限公司 | A kind of feature extracting method and device based on artificial intelligence image recognition |
| CN110059723A (en) * | 2019-03-19 | 2019-07-26 | 北京工业大学 | A kind of robust smog detection method based on integrated depth convolutional neural networks |
| CN110222559A (en) * | 2019-04-24 | 2019-09-10 | 深圳市微纳集成电路与***应用研究院 | Smog image detecting method and device based on convolutional neural networks |
| CN110765937A (en) * | 2019-10-22 | 2020-02-07 | 新疆天业（集团）有限公司 | Coal yard spontaneous combustion detection method based on transfer learning |
| CN112116526A (en) * | 2019-06-19 | 2020-12-22 | 中国石油化工股份有限公司 | Torch smoke image super-resolution method based on deep convolutional neural network |
| CN112418005A (en) * | 2020-11-06 | 2021-02-26 | 北京工业大学 | Smoke multi-classification identification method based on backward radiation attention pyramid network |
| CN113051963A (en) * | 2019-12-26 | 2021-06-29 | 中移(上海)信息通信科技有限公司 | Garbage detection method and device, electronic equipment and computer storage medium |
| CN113096103A (en) * | 2021-04-15 | 2021-07-09 | 北京工业大学 | Intelligent smoke image sensing method for emptying torch |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107481188A (en) * | 2017-06-23 | 2017-12-15 | 珠海经济特区远宏科技有限公司 | A kind of image super-resolution reconstructing method |
| US20180181838A1 (en) * | 2016-12-22 | 2018-06-28 | Samsung Electronics Co., Ltd. | Convolutional neural network system and operation method thereof |
| US20180268298A1 (en) * | 2017-03-15 | 2018-09-20 | Salesforce.Com, Inc. | Deep Neural Network-Based Decision Network |
| CN108564553A (en) * | 2018-05-07 | 2018-09-21 | 南方医科大学 | Low-dose CT image noise suppression method based on convolutional neural networks |
| CN108830809A (en) * | 2018-06-05 | 2018-11-16 | 陕西师范大学 | One kind is based on expansion convolved image denoising method |
-
2018
- 2018-11-26 CN CN201811421607.1A patent/CN109376695B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180181838A1 (en) * | 2016-12-22 | 2018-06-28 | Samsung Electronics Co., Ltd. | Convolutional neural network system and operation method thereof |
| US20180268298A1 (en) * | 2017-03-15 | 2018-09-20 | Salesforce.Com, Inc. | Deep Neural Network-Based Decision Network |
| CN107481188A (en) * | 2017-06-23 | 2017-12-15 | 珠海经济特区远宏科技有限公司 | A kind of image super-resolution reconstructing method |
| CN108564553A (en) * | 2018-05-07 | 2018-09-21 | 南方医科大学 | Low-dose CT image noise suppression method based on convolutional neural networks |
| CN108830809A (en) * | 2018-06-05 | 2018-11-16 | 陕西师范大学 | One kind is based on expansion convolved image denoising method |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109815969A (en) * | 2019-03-05 | 2019-05-28 | 上海骏聿数码科技有限公司 | A kind of feature extracting method and device based on artificial intelligence image recognition |
| CN110059723A (en) * | 2019-03-19 | 2019-07-26 | 北京工业大学 | A kind of robust smog detection method based on integrated depth convolutional neural networks |
| CN110059723B (en) * | 2019-03-19 | 2021-01-05 | 北京工业大学 | Robust smoke detection method based on integrated deep convolutional neural network |
| CN110222559A (en) * | 2019-04-24 | 2019-09-10 | 深圳市微纳集成电路与***应用研究院 | Smog image detecting method and device based on convolutional neural networks |
| CN112116526A (en) * | 2019-06-19 | 2020-12-22 | 中国石油化工股份有限公司 | Torch smoke image super-resolution method based on deep convolutional neural network |
| CN112116526B (en) * | 2019-06-19 | 2024-06-11 | 中国石油化工股份有限公司 | Super-resolution method of torch smoke image based on depth convolution neural network |
| CN110765937A (en) * | 2019-10-22 | 2020-02-07 | 新疆天业（集团）有限公司 | Coal yard spontaneous combustion detection method based on transfer learning |
| CN113051963A (en) * | 2019-12-26 | 2021-06-29 | 中移(上海)信息通信科技有限公司 | Garbage detection method and device, electronic equipment and computer storage medium |
| CN112418005A (en) * | 2020-11-06 | 2021-02-26 | 北京工业大学 | Smoke multi-classification identification method based on backward radiation attention pyramid network |
| CN112418005B (en) * | 2020-11-06 | 2024-05-28 | 北京工业大学 | Smoke multi-classification identification method based on reverse radiation attention pyramid network |
| CN113096103A (en) * | 2021-04-15 | 2021-07-09 | 北京工业大学 | Intelligent smoke image sensing method for emptying torch |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109376695B (en) | 2021-09-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109376695A (en) | A kind of smog detection method based on depth hybrid neural networks | |
| CN109522819B (en) | Fire image identification method based on deep learning | |
| CN103971342B (en) | A kind of image noise detection method based on convolutional neural networks | |
| CN107153789A (en) | The method for detecting Android Malware in real time using random forest grader | |
| CN109977790A (en) | A kind of video smoke detection and recognition methods based on transfer learning | |
| CN104504389B (en) | A kind of satellite cloudiness computational methods based on convolutional neural networks | |
| CN107274451A (en) | Isolator detecting method and device based on shared convolutional neural networks | |
| CN107609399A (en) | Malicious code mutation detection method based on NIN neutral nets | |
| CN106790019A (en) | The encryption method for recognizing flux and device of feature based self study | |
| CN110287805A (en) | Micro- expression recognition method and system based on three stream convolutional neural networks | |
| CN108769993A (en) | Based on the communication network abnormal user detection method for generating confrontation network | |
| CN110363183A (en) | Service robot visual method for secret protection based on production confrontation network | |
| CN106250845A (en) | Flame detecting method based on convolutional neural networks and device | |
| CN108665005A (en) | A method of it is improved based on CNN image recognition performances using DCGAN | |
| CN107808358A (en) | Image watermark automatic testing method | |
| CN108596041A (en) | A kind of human face in-vivo detection method based on video | |
| CN106920256A (en) | A kind of effective missing child searching system | |
| CN108229566A (en) | A kind of hierarchy sorting technique and device | |
| CN109712127A (en) | A kind of electric transmission line fault detection method for patrolling video flowing for machine | |
| CN113221655A (en) | Face spoofing detection method based on feature space constraint | |
| CN102609728B (en) | Method for detecting special pornographic image | |
| CN110415309A (en) | The method automatically generated based on generation confrontation network implementations fingerprint picture | |
| CN109284754A (en) | Ship geometric profile automatic identification method in bridge area based on computer vision | |
| CN112183438A (en) | Image identification method for illegal behaviors based on small sample learning neural network | |
| CN106250840A (en) | Face based on degree of depth study opens closed state detection method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |