CN109740673A - A kind of neural network smog image classification method merging dark - Google Patents
A kind of neural network smog image classification method merging dark Download PDFInfo
- Publication number
- CN109740673A CN109740673A CN201910010088.8A CN201910010088A CN109740673A CN 109740673 A CN109740673 A CN 109740673A CN 201910010088 A CN201910010088 A CN 201910010088A CN 109740673 A CN109740673 A CN 109740673A
- Authority
- CN
- China
- Prior art keywords
- dark
- smog
- network
- image
- channel
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 31
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 19
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims abstract description 16
- 238000012549 training Methods 0.000 claims abstract description 16
- 230000004927 fusion Effects 0.000 claims abstract description 13
- 238000012545 processing Methods 0.000 claims abstract description 12
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 8
- 239000000284 extract Substances 0.000 claims abstract description 8
- 238000013461 design Methods 0.000 claims abstract description 3
- 238000010606 normalization Methods 0.000 claims description 9
- 238000000605 extraction Methods 0.000 claims description 7
- 238000013480 data collection Methods 0.000 claims description 6
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims description 5
- 238000013507 mapping Methods 0.000 claims description 4
- 230000002708 enhancing effect Effects 0.000 claims description 2
- 238000004880 explosion Methods 0.000 claims description 2
- 230000000644 propagated effect Effects 0.000 claims description 2
- 238000012360 testing method Methods 0.000 abstract 1
- 238000001514 detection method Methods 0.000 description 9
- 239000000779 smoke Substances 0.000 description 9
- 230000008569 process Effects 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 230000000007 visual effect Effects 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 235000019504 cigarettes Nutrition 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 239000003595 mist Substances 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000002485 combustion reaction Methods 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 239000006185 dispersion Substances 0.000 description 1
- 238000006073 displacement reaction Methods 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000007429 general method Methods 0.000 description 1
- 239000011521 glass Substances 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 239000002245 particle Substances 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 239000000047 product Substances 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 239000013589 supplement Substances 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The present invention provides a kind of neural network smog image classification methods for merging dark, this method comprises: preparing smog image and two class sample of non-smog image, it is normalized to identical size, dark processing is carried out to all sample images, original image and corresponding dark channel image are divided into training set, verifying collection and test set, the data input as subsequent network training；Then, data set is trained with the binary channels convolutional neural networks of design, wherein first channel network, which increases residual block, improves classification performance, input RGB original image is used to extract the extensive feature in original image；Network is simplified using improved AlexNet in Article 2 channel, and input dark channel image extracts the minutia of smog in dark；Two passes are respectively trained, last Fusion Features, generate training pattern and classify to image；The result shows that this method effectively improves the accuracy rate of smog image classification.
Description
Technical field
The present invention relates to a kind of neural network smog image classification methods for merging dark, belong to image procossing, machine
Vision, deep learning, field of environment protection.
Background technique
Fire incident inevitably threatens to the security of the lives and property of people.Traditional fire detection is generally adopted
With the precise sensors sampled based on particle, temperature sampling, relative humidity sample and smog is analyzed.Although these sensor costs
Low, principle is simple, but since these sensors need to be physically contacted with combustion product ambient gas, can not detect sensing
Fire behavior other than around device, so that the generation of fire can not be avoided in time.The study found that smog often occurs early than open fire,
Traditional smoke detector is similar to fire detector principle, and delay is longer, and efficiency is lower under biggish space environment.With
The high speed development of digital camera and machine vision is caused in terms of fire early detection extensive based on the detection system of image
Concern.Compared with conventional detectors, the Smoke Detection based on machine vision utilizes the visual signature of smog, is in larger range
Incipient fire provides effectively accurately prevention.
The research method of early stage concentrates on the obvious characteristic of observation smog, by color, the movement, edge mould of studying smog
It pastes with features such as frequency domains and detects smog.In common visual signature, texture be it is most reliable, the shape and color of smog are logical
Often because of the type of incendiary material and illumination condition, or due to environment and weather condition there are a great differences.And study found that according to
The dark channel prior of He Kaiming is theoretical, and smog has a kind of special property for dark channel prior theory, i.e., general with other
Object is compared, and cigarette value of dark pixel in dark is higher, and the discovery of this feature is greatly promoted the research of Smoke Detection.
For existing feature extracting method there is no satisfied effect is obtained in Smoke Detection, main cause is smog in face
Color and change in shape very big.In addition, smog has obscured visual scene, cause the feature extracted unstable.Therefore, from image
Accurate detection smog is still a challenging task.In order to improve detection accuracy, it would be desirable to find a kind of synthesis
Feature extracting method, each feature extracted has respective value, so enhancing robustness in conjunction with different features
It is critically important.In recent years, deep learning be because feature can be automatically extracted, have stronger learning ability in image classification and
Identification aspect achieves huge success.It can automatically extract the feature needed when classification image, avoid processing etc. one by hand
The cumbersome process of series, and compare for more general method with very high accuracy.
Therefore, the invention proposes a kind of neural network smog image classification methods for merging dark.Pass through design
Neural network binary channels, the RGB channel image and dark channel image of input data set, realize generalized character and dark respectively
The extraction of feature, syncretizing effect.
Summary of the invention
The invention proposes a kind of neural network smog image classification methods for merging dark, by RGB artwork data collection
Two passes are input to image dark channel data set to be respectively trained, and increase residual block in first channel to optimize
Category of model performance, Article 2 channel are used to extract the dark feature of smog image, and it is complete to obtain finally to carry out Fusion Features
The characteristics of image in face improves the classification performance of network model.
Technical solution of the present invention, including the following steps:
Step 1: preparing smog and non-smog image, day aerial cloud, smooth wall, vehicle body and water are added in data
Face image is to enrich training sample；
Step 2: being 227*227 to the image size normalization in step 1, and carry out dark processing, as subsequent net
The data set of network training；
Step 3: the structure of convolutional neural networks is designed as binary channels network, two passes network is trained simultaneously, wherein
Residual block is added on the basis of AlexNet in first network, by input artwork data collection, extract original image in Generalization Capability compared with
Good feature, Article 2 network inputs dark channel image extract the minutia of smog in dark, and two network end-points carry out
Feature connection, fusion generate training network model file and save；
Step 4: being classified with training pattern to practical smog and non-smog image, obtain classification results.
Compared with prior art, the beneficial effects of the present invention are:
Method of the present invention by the way that data set is carried out dark processing extracts figure using binary channels neural network simultaneously
As the method for feature, model is allowed not only to extract original image information, moreover it is possible to extract the information in dark, obtain model comprehensively
Characteristic information, classify it is more accurate.Also, dual channel network structure makes model that can ensure that model with stand-alone training
Robustness, improve the classification performance of model.
Detailed description of the invention
Fig. 1 smog and non-smoke data collection image；
The original image and dark channel image of the non-smog of Fig. 2 and smog；
The original image of Fig. 3 smog and the visualization feature figure of dark channel image；
Fig. 4 extraction dark feature simplifies CNN structure chart；
Fig. 5 residual block principle assumption diagram；
The residual error network structure of Fig. 6 extraction original image feature；
Fig. 7 binary channels network structure.
Specific embodiment
The present invention is described in further detail With reference to embodiment.A kind of fusion dark of the invention
Neural network smog image classification method include:
1. preparing smog and non-smog image, day aerial cloud, smooth wall, vehicle body and water surface figure are added in data
As to enrich training sample；
2. the image size normalization in pair step 1 is 227*227, and carries out dark processing, instructed as subsequent network
Experienced data set；
3. the structure of convolutional neural networks is designed as binary channels network, two passes network training simultaneously, wherein first
Residual block is added on the basis of AlexNet in network, and by inputting artwork data collection, it is preferable to extract Generalization Capability in original image
Feature, Article 2 network inputs dark channel image extract the minutia of smog in dark, and two network end-points carry out feature
Connection, fusion generate training network model file and save；
4. being classified with training pattern to practical smog and non-smog image, classification results are obtained.
The specific implementation process of technical solution of the present invention is illustrated below.
1 data set
The present invention constructs data set: http://staff referring to the common data sets on website.ustc.edu.Cn/~
yfn/vsd.html.The present invention has collected machine again and is difficult to the image distinguished with smog, increases the diversity of data set, such
Image can be divided into three classes: with smog physics forming process having the same, with smog transparency having the same and identical same
Matter.The first kind includes the object of shared similar physical forming process, such as mist, cloud and steam.Since shade and glass have class
Like the transparent attribute of smog, therefore the representative sample of the second class can be treated them as.Third class includes smooth wall, day
Sky, clothes and vehicle body, because their image patch has height homogeney.In addition, machine is being difficult the water surface and smog classification just
Really, because both having the transparency and high homogeney.So the present invention has collected the figure that wherein four classes have classification challenge
Picture, the supplement including the aerial cloud in day, smooth wall, vehicle body and the water surface, as non-smog image.Smog image and have point
The non-smog image of class challenge is as shown in Figure 1.
2 image preprocessings
2.1 image normalization
The data images being collected into are subjected to size normalization, image of the invention needs the input size of 227*227,
And by normalized image, the convergence of network can be accelerated to a certain extent.
The processing of 2.2 data set darks
This task had collected 9794 smog and non-smog image for meeting requirement of experiment, it is whole with Matlab
Dark processing is carried out, as the network inputs image for extracting dark feature.
The neural network of 3 fusion darks
The fusion of 3.1 dark features
In the smog image detection based on machine vision, what is extracted is the color, form, Texture eigenvalue of smog,
And these features are all the features on smog surface, the present invention is according to the higher speciality of dark channel image dark pixel of smog, by cigarette
The dark feature of mist is merged with other features, increases the comprehensive of feature, improves nicety of grading.Dark is a base
This is it is assumed that certain some pixel, which always has at least one Color Channel, to be had very i.e. in most of non-sky regional area
Low value:
Wherein Jdark(x) value of dark pixel, J are indicatedc(y) pixel of RGB triple channel is indicated.In this formula, dark picture
The value of element is the minimum pixel value in some region, and this pixel value is taken from the minimum value in RGB triple channel.Secretly
Channel prior theorem is pointed out:
Jdark(x)→0 (2)
That is the value of dark pixel always tends to 0, so the dark channel image of object is always showed than darker state, and
For smog, the dark pixel of smog wants high relative to other things values, the original image and dark channel image of non-smog and smog
As shown in Fig. 2, the dark channel image of smog shows more detailed smog information, is conducive to the present invention and smog is carried out comprehensively
Feature extraction.Original image and dark channel image are passed through convolutional neural networks by the present invention respectively, will pass through the original after the first convolutional layer
Figure and the characteristic pattern of dark channel image are visualized, and such as Fig. 3 is extracted in dark channel image as can be seen from Figure
Validity feature figure it is more, the information extracted is more, so, the present invention is mentioned using the technology of fusion smog dark feature
The classification performance of high model.
Dark channel image show extracted in original image less than smoke characteristics, such as smog color contrast, profile information.
Therefore data set is carried out dark processing by the present invention, that is, is taken the minimum value in three channels in RGB to constitute grayscale image, then carried out
Minimum filtering, obtains the value for representing dark, is input in individual network and individually extracts dark feature, finally by itself and one
As feature combine, improve the comprehensive of smoke characteristics.Since the dark extracted is characterized in for assisting master network, for spy
Sign increase it is comprehensive, so the channel network of dark input is had devised one and is simplified based on AlexNet network
CNN network, Fig. 4 are Article 2 channel network configuration.
The 3.2 residual error networks based on AlexNet
In first channel, we increase two residual blocks to improve the performance of AlexNet network, to form one
A new residual error network.It is generally acknowledged that each layer of neural network both corresponds to extract the characteristic information of different levels, including low
Layer, middle layer and upper layer.And network is deeper, the hierarchical information of extraction is more, and the combination of hierarchical information is more.However, with depth
Intensification, gradient disappear, gradient explosion will occur.Traditional corresponding solution is initialization data and regularization, but this
Although solving the problems, such as gradient, it can also deepen the depth of network, error rate increases.Residual error network can effectively solve gradient and ask
Topic, improves the performance of network, residual block theory structure such as Fig. 5.
It can be seen from the figure that X, which can be jumped directly two layers, is changed into input, as identical mapping.If H (X)=F
(X)+X: as F (X)=0, it is clear that H (X)=X is right if H (X)=X increasingly levels off to 0, H (X) and also increasingly levels off to X
In residual unit, can also be indicated to minor function:
yl=h (Xl)+F(Xl, Wl)(Wl={ WL, k|1≤k≤K}) (3)
Xl+1=f (yl) (4)
Wherein ylIndicate the output of this residual block, XlThe input of first of residual block is indicated, if its h (Xl)=Xl, Xl+1=
yl, it is available:
L-th residual unit can indicate the sum of several shallow-layer residual units, include all complex mappings.If display
Function is ε, then propagated forward are as follows:
Obviously, the multiplication between network structure cascade is not present, and the source that gradient disappears also is not present, which can be fine
Ground keeps gradient correlation, solves gradient problem, optimizes network.The present invention increases to residual block in AlexNet, for preventing
Gradient problem improves the classification performance of model.
In our network, the size of input picture is 227*227, and the size after first layer convolutional layer only has
55*55, such setting will lead to gradient problem, so we are optimized using residual error network, improve the strong of whole network
Strong property.Experiment shows that the presence of residual error network can optimize network performance, solves the problems, such as network significantly dimensionality reduction.Fig. 6 is us
The network structure in first channel.
3.3 converged network structures
In order to improve the robustness of whole network, the present invention devises two independent network channels for extracting feature.It is original
Image and dark channel image are trained in two networks respectively when starting；Then first channel the last layer is extracted
To the feature extracted of feature and second channel the last layer in 1: 1 ratio splicing fusion；Finally these features are inputted
Next full articulamentum of layer, enables full articulamentum to be fully integrated the respective feature of two channel networks.Whole binary channels of the invention
The structure of network is as shown in Figure 7.
The size of the input picture of two passes is cropped to 277*227*3 and 227*227*1 from original image respectively.When logical
When crossing convolutional layer, input picture carries out convolution by convolution kernel, in addition offset exports.The present invention uses the non-linear letter of unsaturation
Number ReLU is as activation primitive.In general, Wo Menyou:
Wherein kijIt is the weight of convolution kernel, bjIt is corresponding biasing.In the model, biggish for input image size
Transformation problem, the size of characteristic pattern can vary widely every time, and the quantity of feature can be multiplied, to avoid the occurrence of feature
The problem of excessive and dispersion, we standardize to feature using regularization method, and for residual error network, we use small
The method of batch stochastic gradient descent solves this problem.If following formula calculates the mean value and variance of each activation, to feature
It is normalized:
M represents the size of small batch, xI, fRepresent f-th of feature in batch in the i-th class sample.By mean value and variance,
Following processing can be done to each feature:
Wherein ξ is a small normal number, for improving numerical stability.However, the normalization of input feature vector reduces
Indicate the ability of input.In order to solve this problem, batch processing normalization introduces two parameter γ that can learnfWith feature βf。
Batch normalization is realized by the scaling and displacement of normalization characteristic:
The space size for being gradually reduced eigenmatrix using overlapping maximum Chi Huafa after regularization and ReLU, is being controlled
The parameter and calculation amount in network are reduced while over-fitting.
At the end of two network channels, the feature of two networks extraction is merged present invention uses concat layers, so
The feature of fusion is input in full articulamentum afterwards, recently enters in softmax and classifies.Full articulamentum and softmax
Layer forms linear classification module, feature vector is converted to smoke (' 1 ') and non-smoke (' 0 ') output probability, to reach
To the effect classified to the image of input.The principle formula of full articulamentum is as follows:
Wherein xi, W, biIt respectively inputs, weight matrix and biasing.Representative function mapping,Relative probability value.
Softmax classifier is applied in the present invention with higher probability and lower loss.
A kind of neural network smog image classification method of fusion dark of the invention, will merge the dark channel diagram of smog
As feature, the method for being inputted while being trained using two passes difference, wherein first channel is residual based on AlexNet
Poor network inputs RGB artwork data collection, and for extracting the preferable feature of Generalization Capability in original image, Article 2 channel is an essence
The CNN network of letter, input dark channel image is trained, for extracting the dark minutia of smog.Two passes difference
It is trained, finally carries out Fusion Features, raising feature is comprehensive, ensure that the robustness of network, improves classification performance.Classification
The result shows that this method can realize higher classification performance in smog image classification.
The foregoing is only a preferred embodiment of the present invention, is not intended to limit the scope of the present invention, should
Understand, the present invention is not limited to implementation as described herein, the purpose of these implementations description is to help this field
In technical staff practice the present invention.Any those of skill in the art are easy to do not departing from spirit and scope of the invention
In the case of be further improved and perfect, therefore the present invention is only by the content of the claims in the present invention and the limit of range
System, intention, which covers, all to be included the alternative in the spirit and scope of the invention being defined by the appended claims and waits
Same scheme.
Claims (7)
1. a kind of neural network smog image classification method for merging dark, including the following steps:
Step 1: preparing smog and non-smog image, day aerial cloud, smooth wall, vehicle body and water surface figure are added in data
As to expand training sample；
Step 2: being 227*227 to the image size normalization in step 1, and carry out dark processing, instructed as subsequent network
Experienced data set；
Step 3: the structure of convolutional neural networks is designed as binary channels network, two passes network training simultaneously, wherein first
Residual block network is added in network on the basis of AlexNet, inputs RGB artwork data collection, and it is preferably former to extract Generalization Capability
Figure feature, Article 2 network inputs dark channel image extract smog dark feature, and two network end-points carry out feature connection, melt
It closes, generate training network model file and saves；
Step 4: being classified with training pattern to practical smog and non-smog image, obtain classification results.
2. a kind of neural network smog image classification method for merging dark according to claim 1, which is characterized in that
In step 1, the nicety of grading of subject image enhancing model similar with smog characteristics of image is added.
3. a kind of neural network smog image classification method for merging dark according to claim 1, which is characterized in that
In step 2, all data are subjected to dark processing, the input as Article 2 network channel；Dark principle are as follows: big absolutely
In most non-sky regional areas, certain some pixel always has at least one Color Channel with very low value:
Wherein Jdark(x) value of dark pixel, J are indicatedc(x) pixel of RGB triple channel is indicated；In this formula, dark pixel
Value is the minimum pixel value in some region, and this pixel value is taken from the minimum value in RGB triple channel；Dark
Priori theorem is pointed out:
Jdark(x)→0 (2)
That is the value of dark pixel always tends to 0, so the dark of some scenery is always showed than darker state；And for
For smog, the dark pixel values of smog want high relative to other things, and the dark channel image of smog shows more detailed smog
Characteristic information is conducive to the present invention and carries out comprehensive feature extraction to smog.
4. a kind of neural network smog image classification method for merging dark according to claim 1, which is characterized in that
In step 3, there are binary channels network different data to input, and in first network channel, the design of network is with AlexNet
The promotion that residual block carries out classification performance to network is added in basis, and the input of first network is RGB original image, for mentioning
Take generalized character；And the gradient that there are problems that well solve of residual block disappears and gradient explosion, effectively improves network
Classification performance；Principle formula is as follows:
yl=h (Xl)+F(Xl, Wl)(Wl={ WL, k|1≤k≤K}) (3)
Xl+1=f (yl) (4)
Wherein ylIndicate the output of this residual block, XlThe input of first of residual block is indicated, if its h (Xl)=Xl, Xl+1=yl, can
To obtain:
L-th residual unit can indicate the sum of several shallow-layer residual units, include all complex mappings；If explicit function
For ε, then propagated forward are as follows:
Obviously, the multiplication between network structure cascade is not present, and the source that gradient disappears also is not present, which can protect well
Gradient correlation is held, gradient problem is solved, optimizes network；The present invention increases to residual block in AlexNet, for preventing gradient
Problem improves the classification performance of model.
5. a kind of neural network smog image classification method for merging dark according to claim 1, which is characterized in that
In step 3, Article 2 network channel is obtained by AlexNet network improvement, is directly trained using dark channel image as target data set,
For extracting the minutia of smog in dark channel image.
6. a kind of neural network smog image classification method for merging dark according to claim 1, which is characterized in that
In step 3, the network of two passes is respectively trained, and in concat layers of progress Fusion Features, then is input to full articulamentum to obtain
The comprehensive character of smog image.
7. a kind of neural network smog image classification method for merging dark according to claim 1, which is characterized in that
In step 4, classifies to practical smog image and non-smog image, obtain classification results.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910010088.8A CN109740673A (en) | 2019-01-02 | 2019-01-02 | A kind of neural network smog image classification method merging dark |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910010088.8A CN109740673A (en) | 2019-01-02 | 2019-01-02 | A kind of neural network smog image classification method merging dark |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109740673A true CN109740673A (en) | 2019-05-10 |
Family
ID=66363465
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910010088.8A Pending CN109740673A (en) | 2019-01-02 | 2019-01-02 | A kind of neural network smog image classification method merging dark |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109740673A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110738624A (en) * | 2019-10-18 | 2020-01-31 | 电子科技大学 | area self-adaptive image defogging system and method |
| CN110825899A (en) * | 2019-09-18 | 2020-02-21 | 武汉纺织大学 | Clothing image retrieval method integrating color features and residual network depth features |
| CN111310622A (en) * | 2020-02-05 | 2020-06-19 | 西北工业大学 | Fish swarm target identification method for intelligent operation of underwater robot |
| CN111369472A (en) * | 2020-03-12 | 2020-07-03 | 北京字节跳动网络技术有限公司 | Image defogging method and device, electronic equipment and medium |
| CN111783900A (en) * | 2020-07-10 | 2020-10-16 | 宁波方太厨具有限公司 | Training method of oil smoke concentration detection model and control method of range hood gear |
| CN111898693A (en) * | 2020-08-06 | 2020-11-06 | 上海眼控科技股份有限公司 | Visibility classification model training method, visibility estimation method and device |
| WO2020258615A1 (en) * | 2019-06-19 | 2020-12-30 | 清华大学 | Target classification method, based on sound wave propagation equation, for two-way coupling deep learning |
| CN112861737A (en) * | 2021-02-11 | 2021-05-28 | 西北工业大学 | Forest fire smoke detection method based on image dark channel and YoLov3 |
| CN113361655A (en) * | 2021-07-12 | 2021-09-07 | 武汉智目智能技术合伙企业(有限合伙) | Differential fiber classification method based on residual error network and characteristic difference fitting |
| CN115082834A (en) * | 2022-07-20 | 2022-09-20 | 成都考拉悠然科技有限公司 | Engineering vehicle black smoke emission monitoring method and system based on deep learning |
| CN117173854A (en) * | 2023-09-13 | 2023-12-05 | 西安博深安全科技股份有限公司 | Coal mine open fire early warning method and system based on deep learning |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106846263A (en) * | 2016-12-28 | 2017-06-13 | 中国科学院长春光学精密机械与物理研究所 | The image defogging method being immunized based on fusion passage and to sky |
| CN107749067A (en) * | 2017-09-13 | 2018-03-02 | 华侨大学 | Fire hazard smoke detecting method based on kinetic characteristic and convolutional neural networks |
-
2019
- 2019-01-02 CN CN201910010088.8A patent/CN109740673A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106846263A (en) * | 2016-12-28 | 2017-06-13 | 中国科学院长春光学精密机械与物理研究所 | The image defogging method being immunized based on fusion passage and to sky |
| CN107749067A (en) * | 2017-09-13 | 2018-03-02 | 华侨大学 | Fire hazard smoke detecting method based on kinetic characteristic and convolutional neural networks |
Non-Patent Citations (2)
| Title |
|---|
| JIAN MA 等: "Image-based Air Pollution Estimation Using Hybrid Convolutional Neural Network", 《2018 24TH INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION》 * |
| 不可能打工: "【图像分类】ResNet论文详解（Deep Residual Learning for Image Recognition）", 《HTTPS://BLOG.CSDN.NET/EWEN_LEE/ARTICLE/DETAILS/106851954》 * |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020258615A1 (en) * | 2019-06-19 | 2020-12-30 | 清华大学 | Target classification method, based on sound wave propagation equation, for two-way coupling deep learning |
| CN110825899A (en) * | 2019-09-18 | 2020-02-21 | 武汉纺织大学 | Clothing image retrieval method integrating color features and residual network depth features |
| CN110738624A (en) * | 2019-10-18 | 2020-01-31 | 电子科技大学 | area self-adaptive image defogging system and method |
| CN110738624B (en) * | 2019-10-18 | 2022-02-01 | 电子科技大学 | Area-adaptive image defogging system and method |
| CN111310622A (en) * | 2020-02-05 | 2020-06-19 | 西北工业大学 | Fish swarm target identification method for intelligent operation of underwater robot |
| CN111369472A (en) * | 2020-03-12 | 2020-07-03 | 北京字节跳动网络技术有限公司 | Image defogging method and device, electronic equipment and medium |
| CN111369472B (en) * | 2020-03-12 | 2021-04-23 | 北京字节跳动网络技术有限公司 | Image defogging method and device, electronic equipment and medium |
| CN111783900A (en) * | 2020-07-10 | 2020-10-16 | 宁波方太厨具有限公司 | Training method of oil smoke concentration detection model and control method of range hood gear |
| CN111898693A (en) * | 2020-08-06 | 2020-11-06 | 上海眼控科技股份有限公司 | Visibility classification model training method, visibility estimation method and device |
| CN112861737A (en) * | 2021-02-11 | 2021-05-28 | 西北工业大学 | Forest fire smoke detection method based on image dark channel and YoLov3 |
| CN113361655A (en) * | 2021-07-12 | 2021-09-07 | 武汉智目智能技术合伙企业(有限合伙) | Differential fiber classification method based on residual error network and characteristic difference fitting |
| CN115082834A (en) * | 2022-07-20 | 2022-09-20 | 成都考拉悠然科技有限公司 | Engineering vehicle black smoke emission monitoring method and system based on deep learning |
| CN115082834B (en) * | 2022-07-20 | 2023-03-17 | 成都考拉悠然科技有限公司 | Engineering vehicle black smoke emission monitoring method and system based on deep learning |
| CN117173854A (en) * | 2023-09-13 | 2023-12-05 | 西安博深安全科技股份有限公司 | Coal mine open fire early warning method and system based on deep learning |
| CN117173854B (en) * | 2023-09-13 | 2024-04-05 | 西安博深安全科技股份有限公司 | Coal mine open fire early warning method and system based on deep learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109740673A (en) | A kind of neural network smog image classification method merging dark | |
| CN109522819B (en) | Fire image identification method based on deep learning | |
| Shao et al. | Cloud detection in remote sensing images based on multiscale features-convolutional neural network | |
| CN102332092B (en) | Flame detection method based on video analysis | |
| CN111027549A (en) | Method for identifying key points of women's upper clothes | |
| CN105608454B (en) | Character detecting method and system based on text structure component detection neural network | |
| CN107563355A (en) | Hyperspectral abnormity detection method based on generation confrontation network | |
| CN108647625A (en) | A kind of expression recognition method and device | |
| CN110033040B (en) | Flame identification method, system, medium and equipment | |
| CN104504362A (en) | Face detection method based on convolutional neural network | |
| CN108345850A (en) | The scene text detection method of the territorial classification of stroke feature transformation and deep learning based on super-pixel | |
| CN107564022A (en) | Saliency detection method based on Bayesian Fusion | |
| CN106778687A (en) | Method for viewing points detecting based on local evaluation and global optimization | |
| CN109522965A (en) | A kind of smog image classification method of the binary channels convolutional neural networks based on transfer learning | |
| CN113221655B (en) | Face spoofing detection method based on feature space constraint | |
| CN105512622B (en) | A kind of visible remote sensing image sea land dividing method based on figure segmentation and supervised learning | |
| Yarlagadda et al. | A novel method for human age group classification based on Correlation Fractal Dimension of facial edges | |
| CN108647690A (en) | The sparse holding projecting method of differentiation for unconstrained recognition of face | |
| CN103366373B (en) | Multi-time-phase remote-sensing image change detection method based on fuzzy compatible chart | |
| Guo et al. | Image esthetic assessment using both hand-crafting and semantic features | |
| CN111401213A (en) | Flame detection experimental device and method for chemical gas detection | |
| CN109214331A (en) | A kind of traffic haze visibility detecting method based on image spectrum | |
| Zhang et al. | Fire detection and identification method based on visual attention mechanism | |
| CN111507416B (en) | Smoking behavior real-time detection method based on deep learning | |
| Xu et al. | Generalized morphological component analysis for hyperspectral unmixing |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190510 |
|
| WD01 | Invention patent application deemed withdrawn after publication |