CN110135371A - A kind of Citrus Huanglongbing pathogen recognition methods and device based on Mixup algorithm - Google Patents
A kind of Citrus Huanglongbing pathogen recognition methods and device based on Mixup algorithm Download PDFInfo
- Publication number
- CN110135371A CN110135371A CN201910419761.3A CN201910419761A CN110135371A CN 110135371 A CN110135371 A CN 110135371A CN 201910419761 A CN201910419761 A CN 201910419761A CN 110135371 A CN110135371 A CN 110135371A
- Authority
- CN
- China
- Prior art keywords
- image
- pixel
- model
- mixup
- huanglongbing pathogen
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/188—Vegetation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Multimedia (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of Citrus Huanglongbing pathogen recognition methods based on Mixup algorithm and devices, this method comprises: image to be sorted is pre-processed, carry out data enhancing to pretreated image；The enhanced image of data is subjected to Mixup algorithm process, establishes the linear relationship between sample；BN batch normalized is carried out to the pixel of treated image, the pixel is made to meet normal distribution；Depth of the pixel input based on Model Fusion of the normal distribution is separated into convolution disaggregated model, exports the classification results of described image.This method solve traditional data Enhancement Methods excessively to rely on original image and the insufficient limitation of antagonism sample generalization, separates convolution disaggregated model using the depth based on Model Fusion, improves recognition efficiency, reduce over-fitting, strong robustness.
Description
Technical field
The present invention relates to deep learning field, specifically a kind of Citrus Huanglongbing pathogen recognition methods based on Mixup algorithm and
Device.
Background technique
It is most serious disease that Citrus Huanglongbing pathogen (Citrus Hlb, HLB), which is at present to the Citrus Industry in the world,.18
When century, which is found in the middle part of India, at that time referred to as top dry (dieback).China earlier
When, Citrus Huanglongbing pathogen is to find in southern planting industry, but the farther age can be traced in China in its generation history.Citrus
Yellow twig be all be most common fruit diseases all over the world, it is also widely distributed in China.The province of cultivation citrus has
More than 20, Guangdong Province is one of the main province for wherein planting citrus.But it is influenced in recent years by yellow twig, citrus cultivated area is most
It began to decline in recent years, growing area is shifted to Guangxi province, main to turn to Guangxi Wuzhou, He Prefecture, Guilin.Citrus orchard is plantation citrus
And the main platform of orange yield, mandarin tree will directly affect the yield of citrus with Citrus Huanglongbing pathogen, so using deep
The intellectualized detection mode of degree learning method has citriculture very important meaning.According to the reference text developed to orchard
It offers and orchard investigation on the spot is analyzed, show orchard management level, the intelligent level, Mechanization Level in China
It is lower.
Traditional citrus image-recognizing method, not only time and effort consuming, and compare the engineering for relying on and manually doing feature extraction,
And traditional lenet5, AlexNet model parameter amount is big, recognition accuracy is low.Traditional image in conventional images classification method
There are limitations for Enhancement Method, and network is all different to the parsing of different pictures, and network performance depends on the data of input,
Data are easy to be influenced by various aspects such as image definition, brightness, contrasts.
Therefore, how Citrus Huanglongbing pathogen is accurately identified, is colleague's practitioner's urgent problem to be solved.
Summary of the invention
In view of the above problems, present invention aim to address current traditional data Enhancement Method excessively rely on original image and
The problem of antagonism sample generalization insufficient limitation, realize accurate, the quick identification to Citrus Huanglongbing pathogen.
The embodiment of the present invention provides a kind of Citrus Huanglongbing pathogen recognition methods based on Mixup algorithm, comprising:
Image to be identified is pre-processed, data enhancing is carried out to pretreated image；
The enhanced image of data is subjected to Mixup algorithm process, establishes the linear relationship between sample；
BN batch normalized is carried out to the pixel of treated image, the pixel is made to meet normal distribution；
Depth of the pixel input based on Model Fusion of the normal distribution is separated into convolution disaggregated model, exports institute
State the recognition result of image.
In one embodiment, image to be identified is pre-processed, comprising:
Center is carried out according to the image length and width size to be identified and cuts fixed dimension, is uniformly cut to pre-set dimension
Image.
In one embodiment, the mode of the data enhancing includes one or more of:
Mistake is cut: the degree that mistake is cut is between -5 to 5, and wrong corner cut degree is -16 ° to 16 °；
Gaussian noise: will add some Gaussian noises in original image, by N times of data extending；
Interpolation: interpolation sequence is [0,1], using closest interpolation and bilinear interpolation；
Image is affine: keeping the grazing of image, including rotation, scales, translation, mistake cuts operation；
Mean value disturbance: mean value is taken to disturb from closest pixel；
Intermediate value disturbance: it is disturbed by arest neighbors median.
In one embodiment, the enhanced image of data is subjected to Mixup algorithm process, established linear between sample
Relationship, comprising:
K=λ ki+(1-λ)kj
Y=λ yi+(1-λ)yj
In formula, feature vector ki, i ∈ 1 ..., and m }, yj, j ∈ 1 ..., and m }, m sample, i, j are sample volume
Number, yjFor one-hot coding, λ belongs to [0,1] section.
In one embodiment, the depth based on Model Fusion separates convolution disaggregated model and is trained by following step
It generates, comprising:
Citrus Huanglongbing pathogen image is expanded in acquisition, is pre-processed, and carries out data enhancing to pretreated image；
The enhanced image of data is subjected to Mixup algorithm process, establishes the linear relationship between sample；
BN batch normalized is carried out to the pixel of treated image, the pixel is made to meet normal distribution；
Construct a convolutional neural networks model；The convolutional neural networks model includes one containing the separable volume of depth
Volume module, classifier, output network three parts；The convolutional neural networks model is to separate the basis of convolution model in depth
The network of upper exploitation right reconnection in-between, uses ResNeXt module as output layer；
The pixel of the normal distribution is input to depth and separates convolution module, makees down-sampling by pond layer, subtracts
Few parameter amount, calculates the penalty values between convolutional neural networks output valve and true value；
Using back-propagation algorithm, the directional trim declined according to stochastic gradient descent algorithm along the penalty values is entire
The weight w of network；
When the penalty values are intended to convergence, deconditioning model generates the separable volume of the depth based on Model Fusion
Integrate class model.
In one embodiment, the classifier is sigmoid classifier, addition in the last layer, to input picture into
Row two is classified.
In one embodiment, the penalty values between convolutional neural networks output valve and true value are calculated, following public affairs are passed through
Formula is calculated:
In formula, E indicates the difference between predicted value and true value, and r ' is the output by activation primitive, 0 and 1 it
Between.
Second aspect, the present invention also provides a kind of Citrus Huanglongbing pathogen identification devices based on Mixup algorithm, comprising:
Data enhance module, for pre-processing image to be identified, carry out data increasing to pretreated image
By force；
Calculation processing module establishes the line between sample for the enhanced image of data to be carried out Mixup algorithm process
Sexual intercourse；
Module is normalized, BN batch normalized is carried out for the pixel to treated image, makes the pixel
Point meets normal distribution；
Output module, for depth of the pixel input based on Model Fusion of the normal distribution to be separated convolution point
Class model exports the recognition result of described image.
The beneficial effect of above-mentioned technical proposal provided in an embodiment of the present invention includes at least:
A kind of Citrus Huanglongbing pathogen recognition methods based on Mixup algorithm provided in an embodiment of the present invention, this method solve
Traditional data Enhancement Method excessively relies on original image and the insufficient limitation of antagonism sample generalization, melts using based on model
The depth of conjunction separates convolution disaggregated model and improves recognition efficiency, reduces over-fitting, strong robustness.
Other features and advantages of the present invention will be illustrated in the following description, also, partly becomes from specification
It obtains it is clear that understand through the implementation of the invention.The objectives and other advantages of the invention can be by written explanation
Specifically noted structure is achieved and obtained in book, claims and attached drawing.
Below by drawings and examples, technical scheme of the present invention will be described in further detail.
Detailed description of the invention
Attached drawing is used to provide further understanding of the present invention, and constitutes part of specification, with reality of the invention
It applies example to be used to explain the present invention together, not be construed as limiting the invention.In the accompanying drawings:
Fig. 1 is the flow chart of the Citrus Huanglongbing pathogen recognition methods provided in an embodiment of the present invention based on Mixup algorithm；
Fig. 2 is the process signal for the Citrus Huanglongbing pathogen recognition methods based on Mixup algorithm that the embodiment of the present invention 1 provides
Figure；
Fig. 3 is the affine effect picture of image provided in an embodiment of the present invention；
Fig. 4 is the effect picture of Mixup algorithm provided in an embodiment of the present invention；
Fig. 5 is characteristic pattern after convolution provided in an embodiment of the present invention；
Fig. 6 is ERM distribution probability figure provided in an embodiment of the present invention；
Fig. 7 is the prediction exemplary diagram of the Citrus Huanglongbing pathogen recognition methods provided in an embodiment of the present invention based on Mixup algorithm；
Fig. 8 is the block diagram of the Citrus Huanglongbing pathogen identification device provided in an embodiment of the present invention based on Mixup algorithm.
Specific embodiment
Exemplary embodiments of the present disclosure are described in more detail below with reference to accompanying drawings.Although showing the disclosure in attached drawing
Exemplary embodiment, it being understood, however, that may be realized in various forms the disclosure without should be by embodiments set forth here
It is limited.On the contrary, these embodiments are provided to facilitate a more thoroughly understanding of the present invention, and can be by the scope of the present disclosure
It is fully disclosed to those skilled in the art.
A kind of Citrus Huanglongbing pathogen recognition methods based on Mixup algorithm provided in an embodiment of the present invention, referring to Fig.1, the party
Method includes:
S11, image to be identified is pre-processed, data enhancing is carried out to pretreated image；For example it can pass through
Wireless camera is taken pictures to be crawled with regularization method, is acquired and is expanded to image；It include shape, color and side in the image
Edge feature.
S12, the enhanced image of data is subjected to Mixup algorithm process, establishes the linear relationship between sample；Such as it is defeated
Enter 64 batches 3 and ties up Citrus Huanglongbing pathogen image.Image progress mistake is cut, interpolation, Gauss disturbance, image is affine, mean value disturbs, middle position
The traditional datas Enhancement Methods such as number disturbance, and the linear relationship between sample is established in conjunction with Mixup algorithm.
S13, BN batch normalized is carried out to the pixel of treated image, the pixel is made to meet normal state point
Cloth；Each layer of input layer is carried out to the Gaussian Profile operation of BN batch normalization N (0,1), convolutional layer operation after input.
S14, depth of the pixel input based on Model Fusion of the normal distribution is separated into convolution disaggregated model, it is defeated
The recognition result of described image out.
In step s 11, center is carried out according to image length and width size to be identified and cuts fixed dimension, for example can unified
It is cut to the image of 299*299 size, the later period is facilitated to be uniformly processed.
In the present embodiment, it is somebody's turn to do the Citrus Huanglongbing pathogen recognition methods based on Mixup algorithm, solves traditional data Enhancement Method
Original image and the insufficient limitation of antagonism sample generalization are excessively relied on, using the separable volume of the depth based on Model Fusion
Class model is integrated, recognition efficiency is improved, reduces over-fitting, strong robustness.
The following detailed description of step of the invention.
The above-mentioned Citrus Huanglongbing pathogen recognition methods based on Mixup algorithm, wherein the separable volume of the depth based on Model Fusion
Class model is integrated, larger sample set is relied on, the pre- places such as center cutting, image sharpening and histogram equalization are carried out to sample
Pre-training is carried out after reason, is learnt afterwards using convolutional network；In view of convolutional network is lower in CPU and list GPU operational efficiency
Under, the convolution after double CUDA optimizations can be used in the embodiment of the present invention, realizes that improved efficiency is original about 1.9 times compared to single GPU
Left and right, the result after study identify in application for Citrus Huanglongbing pathogen, reduce time cost, high-efficient.
Transfer learning: using based on ImageNet image training depth can convolution model priori knowledge to our moulds
Type weight parameter w does initialization operation.Training process needs to debug certain partial parameters, can be closer using this method
Weight parameter w suboptimization point.
For example, one width N-dimensional Citrus Huanglongbing pathogen image of input.Image is carried out plus makes an uproar, overturn, scale, image is affine etc., and
In conjunction with Mixup algorithm, data enhancing is carried out, the linear relationship between sample is established, treats partial image with better generalization.
The mode of data enhancing may include following one or more modes:
Mistake is cut: Citrus Huanglongbing pathogen mistake being cut, degree is between -5 to 5, and wrong corner cut degree is -16 ° to 16 °；Wherein degree be-
It is ratio value, without Parameter units between 5 to 5；For example the x coordinate (or y-coordinate) of all the points is allowed to remain unchanged, and correspond to
Y-coordinate (or x coordinate) then translate in proportion, and translate size and the point to x-axis (or y-axis) vertical range at
Direct ratio；
Gaussian noise: will add some Gaussian noises in original image, by 2 times of data extending；
Interpolation: interpolation sequence is [0,1], using closest interpolation and bilinear interpolation；
Image is affine: keeping " grazing " of image, including rotation, scales, translation, mistake cuts operation；
Mean value disturbance: mean value is taken to disturb from closest pixel；
Intermediate value disturbance: it is disturbed by arest neighbors median.
Above-mentioned pretreatment such as in such a way that center is cut, the image of arbitrary dimension, is fixed as the big of 299*299
It is small.
Assuming that m, n are that image is long and wide, image center m/2, n/2.Cutting center is 299/2,299/2, so x
[m/2-299/2:m/2+299/2, n/2-299/2:n/2+299/2,3].
Mixup calculation formula is as follows:
K=λ ki+(1-λ)kj
Y=λ yi+(1-λ)yj
In formula, feature vector ki, i ∈ 1 ..., and m }, yj, j ∈ 1 ..., and m }, m sample, i, j are sample volume
Number, yjFor one-hot coding, λ belongs to [0,1] section.
Each layer of input is subjected to the Gaussian Profile that batch normalizes N (0,1), later layer is then inputted and calculates.Assuming that defeated
Enter for k1, k2, k3......kn, its calculation formula is:
Wherein, numpy is the computational science library of python, and mean is library function, takes the average value of input k, and input point is
One mean value is 0, the Gaussian Profile that variance is 1.Just because of learning rate LR takes biggish value using the regularization mode of BN,
Step-length is 0.1.
It uses depth to separate convolution model as basic model, removes the last layer, introduce RexNeXt module, be added
Dropout layers and classification layer.Dropout uses parameter value for 0.3.
Activation primitive uses Leaky ReLU function:
Function are as follows: S=ax+1, S=max (ax, x), i.e. derivative are a.
Because of the feature of sigmoid index class function, formula are as follows:
Because exponential function both ends are gentle, derivative causes weight to update slowly, gradient disappearance problem occurs close to 0.
Using Focal Loss loss function, in order to solve the problems, such as positive and negative sample proportion serious unbalance in image recognition.
The calculation formula of loss function are as follows:
In formula, E indicates the difference between predicted value and true value, and r ' is the output by activation primitive, so in 0 and 1
Between.
Optimizer uses momentum gradient descent method, uses 0.9 to momentum parameter.Gradient descent method formula are as follows:
Under normal circumstances, gradient descent method W=v0, the renewal amount V of each weight parameter W are as follows:
Wherein, λ is learning rate.
V is thought of as this gradient slippage and the vector sum of part last time renewal amount, i.e.,More with last time x
The new V that measures is multiplied by the sum of a Coefficient m omentum between [0,1].
Loss function:Wherein pwFor predicted value.
It asks
The algorithmic procedure of gradient descent method is as follows:
First to w assignment, value follows Gaussian Profile.The value for changing w, so that E (w) is subtracted by the direction that gradient declines
It is few.
Formula process:
pi=wi-α(pw(x)-y)x(i)
LR initial value uses 0.1 size.
Because of the feature of sigmoid index class function, formula are as follows:
It needs to carry out Citrus Huanglongbing pathogen two classification, differentiates the classification of picture to be sorted.
Calling recalls to function ReduceLRPlateau function, when verifying collection penalty values no longer decline, reduces learning rate
Step sizes: LR=LR*0.2.
Early stop regularization method for example, can be used, when verifying collection penalty values no longer decline in 10 circulations, suspension is instructed
Practice.
Weight and bias are calculated according to back-propagation algorithm, and updates weight w and bias b.
, every batch of 64.
Double GPU data parallel modes, by model copy on two GPU, input is divided into the sub- batch that two sizes are 32,
It is separately operable on two GPU, by returning to the result that size is 64 after connection on CPU.
Illustrate the Citrus Huanglongbing pathogen recognition methods based on Mixup algorithm below by a complete embodiment:
Embodiment 1:
Referring to the step of shown in Fig. 2, explaining entire assorting process and model structure:
Step 1, under Ubuntu system environments, use tensorflow for the keras frame of rear end, to the image of collection
Data set is divided into training set, verifying collection and test set respectively with 6: 2: 2 ratio.
Step 2, input picture, carries out data prediction, the operation that center is cut, and fixed size is the figure of 299*299
Picture.
Data enhancing is carried out to 299*299 size image, uses scaling, mirror image switch, addition Gaussian noise, figure
As affine, rotating image etc. operates.
Mistake is cut: Citrus Huanglongbing pathogen mistake being cut, degree is between -5 to 5, and wrong corner cut degree is -16 ° to 16 °；
Gaussian noise: will add some Gaussian noises in original image, by 2 times of data extending；
Interpolation: interpolation sequence is [0,1], using closest interpolation and bilinear interpolation；
Image is affine: keeping " grazing " of image, including rotation, scales, translation, mistake cuts operation, referring to Fig. 3, the figure
Show the effect that image is affine in traditional data enhancing algorithm.
Mean value disturbance: mean value is taken to disturb from closest pixel；
Intermediate value disturbance: it is disturbed by arest neighbors median.
Referring to Fig. 4, Citrus Huanglongbing pathogen data set establishes linear relationship to sample by mixup algorithm.
Step 3, BN batch normalization operation is carried out to input pixel, pixel is made to be distributed as mean value 0, variance 1
Gaussian Profile, can improve arrival optimize point efficiency.
Step 4, neural network training model structure is constructed.
The priori knowledge of training on ImageNet data set is moved to using transfer learning by Citrus Huanglongbing pathogen identification
Model carry out parameter value initialization operation, help to improve training effectiveness.
By the way of fine tuning, after adjustment member parameter, the weight file for importing oneself preservation is trained, Ke Yi great
Amplitude reduces the training time.
The Foundation neural network model of convolution model is separated in depth, 1*1 convolution kernel and shortcut machine is added
System increases model nonlinear capability of fitting.
Convolution kernel uses 3*3 size dimension, and using LReLU activation primitive, pond layer uses the maximum pond of 2*2.Reference
Fig. 5 learns that after convolutional layer, image thickens image by the library matplotlib drawing tool.
Output layer adds one layer dropout layers, and the regularization inactivated at random to the neuron of output layer 30% operates,
Help to reduce model complexity, effectively inhibits the degree of model over-fitting.
Step 5, classifier is established.Sigmoid classifier is added in the last layer, two classification are carried out to input picture.
Step 6, the mistake between predicted value and true value is calculated according to the following formula using Focal Loss loss function
Poor size E.As shown in fig. 6, being ERM distribution probability figure, it may be assumed that empirical risk minimization, the smallest distribution probability figure of penalty values.When
When hyper parameter is 0.2/0.4, empirical risk minimization, that is, penalty values are minimum.
Step 7, referring to shown in Fig. 7, model optimizer is used as using momentum gradient descent method, to predicted value and true value it
Between error amount optimize, constantly diminution error amount, approach w local optimum.It takes turns when penalty values continuous 5 and no longer declines, subtract
Small learning rate.
It is calculated using back-propagation algorithm, every wheel updates weighted value w size:
By the way of double GPU data parallels, iteration is trained the picture of 64 batch sizes each time.
For example, being trained to 560 Citrus Huanglongbing pathogens, data set reaches after the library opencv carries out data enhancing
7168 figures, every wheel iteration 112 times.When every 5 wheel stops training when verifying collection penalty values no longer decline.Save final weight
File and model.
The depth based on Model Fusion separates convolution disaggregated model during training and study, applied hard
Part environment and software environment, for example can refer to as follows:
Using two pieces of model 1080Ti, video memory is the GPU of 11G, saves as 32G in the CPU of use, is equipped with intel
6800k cpu i7,256 solid state hard disk of hard-disk capacity add 4T mechanical hard disk.Operating system is Ubuntu system.Using arriving
The libraries such as matplotlib, numpy, sklearn under the library opencv, python programming language and python system.
Based on the same inventive concept, the embodiment of the invention also provides a kind of, and the Citrus Huanglongbing pathogen based on Mixup algorithm is known
Other device, by the principle and a kind of aforementioned Citrus Huanglongbing pathogen recognition methods based on Mixup algorithm of the solved problem of the device
It is similar, therefore the implementation of the device may refer to the implementation of preceding method, overlaps will not be repeated.
The embodiment of the invention provides Citrus Huanglongbing pathogen identification device of the kind based on Mixup algorithm, referring to shown in Fig. 8, packet
It includes:
Data enhance module 81, for pre-processing image to be identified, carry out data to pretreated image
Enhancing；
Calculation processing module 82 is established between sample for the enhanced image of data to be carried out Mixup algorithm process
Linear relationship；
Module 83 is normalized, BN batch normalized is carried out for the pixel to treated image, makes the picture
Vegetarian refreshments meets normal distribution；
Output module 84, for depth of the pixel input based on Model Fusion of the normal distribution to be separated convolution
Disaggregated model exports the recognition result of described image.
Obviously, various changes and modifications can be made to the invention without departing from essence of the invention by those skilled in the art
Mind and range.In this way, if these modifications and changes of the present invention belongs to the range of the claims in the present invention and its equivalent technologies
Within, then the present invention is also intended to include these modifications and variations.
Claims (8)
1. a kind of Citrus Huanglongbing pathogen recognition methods based on Mixup algorithm characterized by comprising
Image to be identified is pre-processed, data enhancing is carried out to pretreated image；
The enhanced image of data is subjected to Mixup algorithm process, establishes the linear relationship between sample；
BN batch normalized is carried out to the pixel of treated image, the pixel is made to meet normal distribution；
Depth of the pixel input based on Model Fusion of the normal distribution is separated into convolution disaggregated model, exports the figure
The recognition result of picture.
2. a kind of Citrus Huanglongbing pathogen recognition methods based on Mixup algorithm as described in claim 1, which is characterized in that will be to
The image of identification is pre-processed, comprising:
Center is carried out according to the image length and width size to be identified and cuts fixed dimension, is uniformly cut to the figure of pre-set dimension
Picture.
3. a kind of Citrus Huanglongbing pathogen recognition methods based on Mixup algorithm as described in claim 1, which is characterized in that described
The mode of data enhancing includes one or more of:
Mistake is cut: the degree that mistake is cut is between -5 to 5, and wrong corner cut degree is -16 ° to 16 °；
Gaussian noise: will add some Gaussian noises in original image, by N times of data extending；
Interpolation: interpolation sequence is [0,1], using closest interpolation and bilinear interpolation；
Image is affine: keeping the grazing of image, including rotation, scales, translation, mistake cuts operation；
Mean value disturbance: mean value is taken to disturb from closest pixel；
Intermediate value disturbance: it is disturbed by arest neighbors median.
4. a kind of Citrus Huanglongbing pathogen recognition methods based on Mixup algorithm as described in claim 1, which is characterized in that will count
Mixup algorithm process is carried out according to enhanced image, establishes the linear relationship between sample, comprising:
K=λ ki+(1-λ)kj
Y=λ yi+(1-λ)yj
In formula, feature vector ki, i ∈ 1 ..., and m }, yj, j ∈ 1 ..., m }, m sample, i, j are sample number,
yjFor one-hot coding, λ belongs to [0,1] section.
5. a kind of Citrus Huanglongbing pathogen recognition methods based on Mixup algorithm as described in claim 1, which is characterized in that described
Depth based on Model Fusion separates convolution disaggregated model and is generated by following step training, comprising:
Citrus Huanglongbing pathogen image is expanded in acquisition, is pre-processed, and carries out data enhancing to pretreated image；
The enhanced image of data is subjected to Mixup algorithm process, establishes the linear relationship between sample；
BN batch normalized is carried out to the pixel of treated image, the pixel is made to meet normal distribution；
Construct a convolutional neural networks model；The convolutional neural networks model includes one and separates convolution mould containing depth
Block, classifier, output network three parts；The convolutional neural networks model is sharp on the basis of depth separates convolution model
With the network of weight connection in-between, use ResNeXt module as output layer；
The pixel of the normal distribution is input to depth and separates convolution module, makees down-sampling by pond layer, reduces ginseng
Quantity calculates the penalty values between convolutional neural networks output valve and true value；
Using back-propagation algorithm, the directional trim whole network declined according to stochastic gradient descent algorithm along the penalty values
Weight w；
When the penalty values are intended to convergence, deconditioning model generates the depth based on Model Fusion and separates convolution point
Class model.
6. a kind of Citrus Huanglongbing pathogen recognition methods based on Mixup algorithm as claimed in claim 5, which is characterized in that described
Classifier is sigmoid classifier, and addition carries out two classification in the last layer, to input picture.
7. a kind of Citrus Huanglongbing pathogen recognition methods based on Mixup algorithm as claimed in claim 5, which is characterized in that calculate
Penalty values between convolutional neural networks output valve and true value are calculated by following formula:
In formula, E indicates the difference between predicted value and true value, and r ' is the output by activation primitive, between zero and one.
8. a kind of Citrus Huanglongbing pathogen identification device based on Mixup algorithm characterized by comprising
Data enhance module, for pre-processing image to be identified, carry out data enhancing to pretreated image；
Calculation processing module establishes the linear pass between sample for the enhanced image of data to be carried out Mixup algorithm process
System；
Module is normalized, BN batch normalized is carried out for the pixel to treated image, makes the pixel glyph
Close normal distribution；
Output module, for depth of the pixel input based on Model Fusion of the normal distribution to be separated convolution classification mould
Type exports the recognition result of described image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910419761.3A CN110135371A (en) | 2019-05-20 | 2019-05-20 | A kind of Citrus Huanglongbing pathogen recognition methods and device based on Mixup algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910419761.3A CN110135371A (en) | 2019-05-20 | 2019-05-20 | A kind of Citrus Huanglongbing pathogen recognition methods and device based on Mixup algorithm |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110135371A true CN110135371A (en) | 2019-08-16 |
Family
ID=67571511
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910419761.3A Pending CN110135371A (en) | 2019-05-20 | 2019-05-20 | A kind of Citrus Huanglongbing pathogen recognition methods and device based on Mixup algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110135371A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110633751A (en) * | 2019-09-17 | 2019-12-31 | 上海眼控科技股份有限公司 | Training method of car logo classification model, car logo identification method, device and equipment |
| CN110940636A (en) * | 2019-12-05 | 2020-03-31 | 华南农业大学 | Intelligent identification and forest information monitoring system for citrus forest diseases and insect pests |
| CN112131421A (en) * | 2020-09-23 | 2020-12-25 | 平安科技（深圳）有限公司 | Medical image classification method, device, equipment and storage medium |
| CN112836820A (en) * | 2021-01-31 | 2021-05-25 | 云知声智能科技股份有限公司 | Deep convolutional network training method, device and system for image classification task |
| CN114764887A (en) * | 2022-04-12 | 2022-07-19 | 中国农业大学 | Crop disease identification method, device, equipment and storage medium |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104008551A (en) * | 2014-06-10 | 2014-08-27 | 华南农业大学 | Liberobacter asiaticum detection method based on visible light images |
| CN109065030A (en) * | 2018-08-01 | 2018-12-21 | 上海大学 | Ambient sound recognition methods and system based on convolutional neural networks |
| CN109272045A (en) * | 2018-09-25 | 2019-01-25 | 华南农业大学 | A kind of fruit image classification method and device based on neural network and transfer learning |
| CN109376763A (en) * | 2018-09-13 | 2019-02-22 | 山东师范大学 | Sample classification method, system and medium based on multisample reasoning neural network |
| CN109493933A (en) * | 2018-08-08 | 2019-03-19 | 浙江大学 | A kind of prediction meanss of the adverse cardiac events based on attention mechanism |
| CN109636791A (en) * | 2018-12-13 | 2019-04-16 | 华南农业大学 | A kind of Citrus Huanglongbing pathogen detection method based on deep learning, apparatus and system |
-
2019
- 2019-05-20 CN CN201910419761.3A patent/CN110135371A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104008551A (en) * | 2014-06-10 | 2014-08-27 | 华南农业大学 | Liberobacter asiaticum detection method based on visible light images |
| CN109065030A (en) * | 2018-08-01 | 2018-12-21 | 上海大学 | Ambient sound recognition methods and system based on convolutional neural networks |
| CN109493933A (en) * | 2018-08-08 | 2019-03-19 | 浙江大学 | A kind of prediction meanss of the adverse cardiac events based on attention mechanism |
| CN109376763A (en) * | 2018-09-13 | 2019-02-22 | 山东师范大学 | Sample classification method, system and medium based on multisample reasoning neural network |
| CN109272045A (en) * | 2018-09-25 | 2019-01-25 | 华南农业大学 | A kind of fruit image classification method and device based on neural network and transfer learning |
| CN109636791A (en) * | 2018-12-13 | 2019-04-16 | 华南农业大学 | A kind of Citrus Huanglongbing pathogen detection method based on deep learning, apparatus and system |
Non-Patent Citations (7)
| Title |
|---|
| DAOJUN LIANG等: "Understanding Mixup Training Method", 《IEEE ACCESS》 * |
| HONGYI ZHANG等: "mixup: BEYOND EMPIRICAL RISK MINIMIZATION", 《ARXIV:1710.09412V2 [CS.LG] 27 APR 2018》 * |
| RUTU GANDHI等: "Plant disease detection using CNNs and GANs as an augmentative approach", 《2018 IEEE INTERNATIONAL CONFERENCE ON INNOVATIVE RESEARCH AND DEVELOPMENT (ICIRD)》 * |
| YINGWEI FU等: "Mixup Based Privacy Preserving Mixed Collaboration Learning", 《2019 IEEE INTERNATIONAL CONFERENCE ON SERVICE-ORIENTED SYSTEM ENGINEERING》 * |
| 刘庆飞等: "基于深度可分离卷积的实时农业图像逐像素分类研究", 《中国农业科学》 * |
| 许茗等: "复杂热红外监控场景下行人检测", 《中国图象图形学报》 * |
| 陈伟等: "基于RS-LSTM的滚动轴承故障识别", 《中国科技论文》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110633751A (en) * | 2019-09-17 | 2019-12-31 | 上海眼控科技股份有限公司 | Training method of car logo classification model, car logo identification method, device and equipment |
| CN110940636A (en) * | 2019-12-05 | 2020-03-31 | 华南农业大学 | Intelligent identification and forest information monitoring system for citrus forest diseases and insect pests |
| CN110940636B (en) * | 2019-12-05 | 2020-12-08 | 华南农业大学 | Intelligent identification and forest information monitoring system for citrus forest diseases and insect pests |
| CN112131421A (en) * | 2020-09-23 | 2020-12-25 | 平安科技（深圳）有限公司 | Medical image classification method, device, equipment and storage medium |
| CN112131421B (en) * | 2020-09-23 | 2023-09-15 | 平安科技（深圳）有限公司 | Medical image classification method, device, equipment and storage medium |
| CN112836820A (en) * | 2021-01-31 | 2021-05-25 | 云知声智能科技股份有限公司 | Deep convolutional network training method, device and system for image classification task |
| CN112836820B (en) * | 2021-01-31 | 2023-12-19 | 云知声智能科技股份有限公司 | Deep convolution network training method, device and system for image classification task |
| CN114764887A (en) * | 2022-04-12 | 2022-07-19 | 中国农业大学 | Crop disease identification method, device, equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110135371A (en) | A kind of Citrus Huanglongbing pathogen recognition methods and device based on Mixup algorithm | |
| CN111079639B (en) | Method, device, equipment and storage medium for constructing garbage image classification model | |
| Kumar et al. | Image based tomato leaf disease detection | |
| CN104834922B (en) | Gesture identification method based on hybrid neural networks | |
| CN108648191A (en) | Pest image-recognizing method based on Bayes's width residual error neural network | |
| CN111898547A (en) | Training method, device and equipment of face recognition model and storage medium | |
| CN108133188A (en) | A kind of Activity recognition method based on motion history image and convolutional neural networks | |
| CN109272045A (en) | A kind of fruit image classification method and device based on neural network and transfer learning | |
| CN105844285A (en) | Cucumber disease identification method and apparatus based on image information | |
| CN107392919B (en) | Adaptive genetic algorithm-based gray threshold acquisition method and image segmentation method | |
| CN111274916A (en) | Face recognition method and face recognition device | |
| CN107918772B (en) | Target tracking method based on compressed sensing theory and gcForest | |
| Salve et al. | Iris recognition using SVM and ANN | |
| EP3671635B1 (en) | Curvilinear object segmentation with noise priors | |
| CN115424093A (en) | Method and device for identifying cells in fundus image | |
| Zhuang et al. | A handwritten Chinese character recognition based on convolutional neural network and median filtering | |
| Balakrishnan et al. | Meticulous fuzzy convolution C means for optimized big data analytics: adaptation towards deep learning | |
| Kanaparthi et al. | A survey on training issues in chili leaf diseases identification using deep learning techniques | |
| CN110097067A (en) | It is a kind of based on layer into the Weakly supervised fine granularity image classification method of formula eigentransformation | |
| CN109034143A (en) | The micro- expression recognition method of face based on video amplifier and deep learning | |
| Zhang et al. | Application of gesture recognition based on simulated annealing bp neural network | |
| EP3671634B1 (en) | Curvilinear object segmentation with geometric priors | |
| CN107729992B (en) | Deep learning method based on back propagation | |
| Zhu et al. | Support vector machine optimized using the improved fish swarm optimization algorithm and its application to face recognition | |
| Mouna et al. | A vehicle detection approach using deep learning network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190816 |