CN107798356A - Crop leaf disease recognition method based on depth convolutional neural networks - Google Patents
Crop leaf disease recognition method based on depth convolutional neural networks Download PDFInfo
- Publication number
- CN107798356A CN107798356A CN201711195140.9A CN201711195140A CN107798356A CN 107798356 A CN107798356 A CN 107798356A CN 201711195140 A CN201711195140 A CN 201711195140A CN 107798356 A CN107798356 A CN 107798356A
- Authority
- CN
- China
- Prior art keywords
- leaf
- image
- crop
- layer
- convolutional neural
- 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
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
-
- 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)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The disclosure is directed to a kind of crop leaf disease recognition method based on depth convolutional neural networks, belong to image procossing and machine learning techniques field, this method includes:Raw image database is expanded to obtain and expands image data base;Depth convolutional neural networks are built, and the depth convolutional neural networks are trained to obtain crop disease identification model using the expansion image data base;The leaf image of crop to be identified is inputted into the crop disease identification model and obtains characteristic vector, and image classification is carried out to the characteristic vector;The Damage Types of the crop to be identified are obtained according to the class label of the crop to be identified and image classification result.This method can improve the accuracy rate of crop disease identification model, while also improve the applicability of crop disease identification model.
Description
Technical field
This disclosure relates to image procossing and machine learning techniques field, in particular to one kind based on depth convolution god
Crop leaf disease recognition method through network.
Background technology
Crop disease can both influence the growth of crop, can influence the yield and quality of crop again;Therefore, crop is detected in time
Disease occurs and type identification, prevention and preventing and treating for crop disease have very important significance.Wherein, to crop disease
The key effectively prevented and treated is how quickly to make a definite diagnosis Damage Types and state of development, and according to Damage Types and state of development system
Determine effectively preventing method and dosage.
, can be by extracting the face of disease leaf image in traditional disease recognition method based on crop leaf image
Color, shape and Texture eigenvalue are realized.Wherein, feature extraction and selection be the disease recognition method core;Due to simultaneously
The not all feature extracted is all effective to follow-up disease recognition, and some features can reduce the identification of crop leaf disease on the contrary
Rate, therefore feature how is automatically extracted from crop disease leaf image, and carry out accurately classification and know as crop disease
The hot issue do not studied.
At present, there is scholar to study crop disease leaf image identification theory, and propose and be much based on blade
The crop disease recognition methods of image.For example, Arivazhagan et al. proposes a kind of crop disease leaf image detection and known
Other method;King offers cutting edge of a knife or a sword et al. by extracting the color, shape, Texture eigenvalue of disease leaf image scab, with reference to plant growth
Environmental information, utilize techniques of discriminant analysis identification cucumber scab classification;Zhang et al. carries out the spot point of leaf image first
Cut, then extract color, shape and the textural characteristics of scab, then 5 kinds of maize leafs are carried out by K- arest neighbors sorting algorithm
Identification etc.;Above-mentioned algorithm is mainly the special characteristic by extracting disease leaf image, in conjunction with conventional sorting methods to disease
It is identified, although achieving preferable recognition effect, due to the complexity of crop disease leaf image so that extract
Special characteristic can not completely and preferably characterize the scab feature of crop disease leaf image, have certain limitation, such as
Some features are only suitable for representing the scab with clear texture, and other feature is only suitable for representing with clear profile
Scab;Further, different features can also have a negative impact to recognition effect, so, select one group to be adapted to Damage Types
Knowing another characteristic needs substantial amounts of experimental study and summary of experience, is a time-consuming stubborn problem again.
Given this, it is desirable to provide a kind of new crop leaf disease recognition method based on depth convolutional neural networks.
It should be noted that information is only used for strengthening the reason to the background of the disclosure disclosed in above-mentioned background section
Solution, therefore can include not forming the information to prior art known to persons of ordinary skill in the art.
The content of the invention
The purpose of the disclosure is to provide a kind of crop leaf disease recognition method based on depth convolutional neural networks, entered
And one or more problem caused by the limitation of correlation technique and defect is at least overcome to a certain extent.
According to an aspect of this disclosure, there is provided a kind of crop leaf disease recognition side based on depth convolutional neural networks
Method, including:
Raw image database is expanded to obtain and expands image data base;
Depth convolutional neural networks are built, and the depth convolutional neural networks are entered using the expansion image data base
Row training obtains crop disease identification model;
The leaf image of crop to be identified is inputted into the crop disease identification model and obtains characteristic vector, and to institute
State characteristic vector and carry out image classification;
The disease class of the crop to be identified is obtained according to the class label of the crop to be identified and image classification result
Type.
In a kind of exemplary embodiment of the disclosure, raw image database is expanded to obtain and expands view data
Storehouse includes:
Rotation process, translation, size change over operations are carried out to each history leaf image in raw image database
And noise processed obtains multiple new leaf images;
The expansion image data base is formed using each new leaf image.
In a kind of exemplary embodiment of the disclosure, each history leaf image in raw image database is revolved
Turn operation and obtain multiple new leaf images to include:
Multiple predetermined angles are set, and according to each predetermined angle to each history leaf in the raw image database
Picture is rotated to obtain multiple new leaf images.
In a kind of exemplary embodiment of the disclosure, each history leaf image in raw image database is put down
Shifting operation, which obtains multiple new leaf images, to be included:
Presetted pixel number is set, and according to the presetted pixel number to each history in the raw image database
Leaf image carries out translation and obtains multiple new leaf images.
In a kind of exemplary embodiment of the disclosure, chi is carried out to each history leaf image in raw image database
Very little map function, which obtains multiple new leaf images, to be included:
Set to preset and cut out number of pixels, and number of pixels is cut out to each in raw image database according to described preset
History leaf image is cut out to obtain multiple new leaf images.
In a kind of exemplary embodiment of the disclosure, each history leaf image in raw image database is made an uproar
Sonication, which obtains multiple new leaf images, to be included:
Salt-pepper noise and white Gaussian noise are set;And according to the salt-pepper noise and white Gaussian noise to described original
Each history leaf image in image data base carries out noise processed and obtains multiple new leaf images.
In a kind of exemplary embodiment of the disclosure, the crop leaf disease recognition method also includes:
Average value processing is carried out to each history leaf image and new leaf image.
In a kind of exemplary embodiment of the disclosure, structure depth convolutional neural networks include:
Step S10, build the level 0 of the depth convolutional neural networks;Wherein, level 0 is input layer;The input
Each leaf image of layer input is the long and wide RGB image with identical size;
Step S20, build the first layer of the depth convolutional neural networks;Wherein, the first layer includes convolutional layer, most
Great Chiization layer and batch normalization operation;
The convolutional layer is used to obtain multiple fisrt feature figures to RGB image progress convolution;The maximum pond layer
For carrying out dimension-reduction treatment to the fisrt feature figure;Described batch of normalization operation is used for the fisrt feature figure after dimension-reduction treatment
Carry out batch normalized and obtain multiple second feature figures;
Step S30, build the second layer of the depth convolutional neural networks;Wherein, the second layer is used for described the
Two characteristic patterns carry out convolution, maximum pond and batch normalized and obtain multiple third feature figures;
Step S40, build the third layer of the depth convolutional neural networks;Wherein, the third layer is used for described the
Three characteristic patterns carry out convolution, maximum pond and batch normalized and obtain multiple fourth feature figures;
Step S50, build the 4th layer of the depth convolutional neural networks;Wherein, described 4th layer is used for described the
Four characteristic patterns carry out convolution and batch normalized obtains multiple fifth feature figures;
Step S60, build the layer 5 of the depth convolutional neural networks;Wherein, the layer 5 is used for described the
Five characteristic patterns carry out global average pondization operation and obtain the characteristic value of the fifth feature figure;Wherein, the dimension of the characteristic value
Equal to the quantity of fifth feature figure;
Step S70, build the layer 6 of the depth convolutional neural networks;Wherein, the layer 6 is used for the spy
Value indicative is classified.
In a kind of exemplary embodiment of the disclosure, in the step S20, dimensionality reduction is carried out to the fisrt feature figure
Processing includes:
Dimension-reduction treatment is carried out to the fisrt feature figure using the method for local maximizing.
In a kind of exemplary embodiment of the disclosure, in the step S20, to the fisrt feature figure after dimension-reduction treatment
Carry out batch normalized and obtain multiple second feature figures including:
The excitation in the adjacent fisrt feature figure after dimension-reduction treatment is normalized using batch normalization algorithm, obtained
To multiple second feature figures.
A kind of crop leaf disease recognition method based on depth convolutional neural networks of the disclosure, by using expansion image
Database is trained to obtain crop disease identification model to depth convolutional neural networks;Again by the leaf image of crop to be identified
Input obtains characteristic vector into crop disease identification model, and carries out image classification to characteristic vector;Finally according to be identified
The class label of crop obtains the Damage Types of crop to be identified with image classification result;On the one hand, by original image number
Expanded to obtain according to storehouse and expand image data base, and depth convolutional neural networks are trained using image data base is expanded
The quantity for arriving crop disease identification model, adding image in image data base so that the training to convolutional neural networks is more
Add comprehensively, improve the accuracy rate of crop disease identification model, while also improve the applicability of crop disease identification model;Separately
On the one hand, by being inputted leaf image into crop disease identification model and carrying out image classification to characteristic vector, further according to
Class label obtains the Damage Types of crop to be identified with image classification result, improves to the Damage Types of crop to be identified
Recognition speed so that plant personnel and scientific research personnel timely can effect a radical cure according to Damage Types to the disease of crop,
Improve economic benefit.
It should be appreciated that the general description and following detailed description of the above are only exemplary and explanatory, not
The disclosure can be limited.
Brief description of the drawings
Accompanying drawing herein is merged in specification and forms the part of this specification, shows the implementation for meeting the disclosure
Example, and be used to together with specification to explain the principle of the disclosure.It should be evident that drawings in the following description are only the disclosure
Some embodiments, for those of ordinary skill in the art, on the premise of not paying creative work, can also basis
These accompanying drawings obtain other accompanying drawings.
Fig. 1 schematically shows a kind of flow chart of the crop leaf disease recognition method based on depth convolutional neural networks.
Fig. 2 schematically shows a kind of method flow diagram expanded raw image database.
Fig. 3 schematically shows a kind of method flow diagram for building depth convolutional neural networks.
Embodiment
Example embodiment is described more fully with referring now to accompanying drawing.However, example embodiment can be with a variety of shapes
Formula is implemented, and is not understood as limited to example set forth herein;On the contrary, these embodiments are provided so that the disclosure will more
Fully and completely, and by the design of example embodiment comprehensively it is communicated to those skilled in the art.Described feature, knot
Structure or characteristic can be incorporated in one or more embodiments in any suitable manner.In the following description, there is provided permitted
More details fully understand so as to provide to embodiment of the present disclosure.It will be appreciated, however, by one skilled in the art that can
Omitted with putting into practice the technical scheme of the disclosure one or more in the specific detail, or others side can be used
Method, constituent element, device, step etc..In other cases, be not shown in detail or describe known solution a presumptuous guest usurps the role of the host to avoid and
So that each side of the disclosure thickens.
In addition, accompanying drawing is only the schematic illustrations of the disclosure, it is not necessarily drawn to scale.Identical accompanying drawing mark in figure
Note represents same or similar part, thus will omit repetition thereof.Some block diagrams shown in accompanying drawing are work(
Can entity, not necessarily must be corresponding with physically or logically independent entity.
In recent years, convolutional neural networks achieve breakthrough achievement in terms of image recognition.Due to convolutional neural networks not
Traditional machine learning is same as, convolutional neural networks can learn more abstract high-level characteristic automatically from the image of input,
The characteristic extraction procedure of complexity is avoided, can solve the problems such as image classification of complexity and crop disease identification.
Further, depth convolutional neural networks are a kind of deep learning networks for having supervision, in crop disease blade
Preliminary research and application are obtained in image recognition.For example, Sladojevic et al., which proposes one kind, is based on depth convolutional Neural
The plant leaf portion disease geo-radar image recognition methods of network, achieves higher recognition correct rate;Qin Feng et al. is based on convolutional Neural net
Network extracts scab characteristics of image, establishes disease recognition model.The studies above achievement using convolutional neural networks it can be shown that carried out
Crop leaf disease recognition is feasible, but is also apparent from the defects of above-mentioned algorithm;Such as:The parameter that network is related to compares
It is more, training that time is long and the problems such as the universality difference of network.
A kind of crop leaf disease recognition based on depth convolutional neural networks is provide firstly in this example embodiment
Method.With reference to shown in figure 1, the crop leaf disease recognition method based on depth convolutional neural networks of being somebody's turn to do can include following step
Suddenly:
Step S110. is expanded to obtain to raw image database expands image data base.
Step S120. builds depth convolutional neural networks, and using the expansion image data base to the depth convolution
Neutral net is trained to obtain crop disease identification model.
Step S130. the leaf image of crop to be identified is inputted into the crop disease identification model obtain feature to
Amount, and image classification is carried out to the characteristic vector.
Step S140. obtains the crop to be identified according to the class label of the crop to be identified with image classification result
Damage Types.
The above-mentioned crop leaf disease recognition method based on depth convolutional neural networks, on the one hand, by original image
Database, which is expanded to obtain, expands image data base, and depth convolutional neural networks are instructed using image data base is expanded
The experienced quantity for arriving crop disease identification model, adding image in image data base so that the training to convolutional neural networks
More comprehensively, the accuracy rate of crop disease identification model is improved, while also improve the applicability of crop disease identification model;
On the other hand, by being inputted leaf image into crop disease identification model and image classification, then root being carried out to characteristic vector
The Damage Types of crop to be identified are obtained according to class label and image classification result, improve the Damage Types to crop to be identified
Recognition speed so that plant personnel and scientific research personnel timely can carry out root according to Damage Types to the disease of crop
Control, improve economic benefit.
Below, by the above-mentioned crop leaf disease recognition side based on depth convolutional neural networks in this example embodiment
Each step in method carries out detailed explanation and explanation.
In step s 110, raw image database is expanded to obtain and expands image data base.Wherein, with reference to figure 2
Shown, expansion is carried out to raw image database can include step S1102 and step S1104.Wherein:
In step S1102, rotation process is carried out to each history leaf image in raw image database, translation is grasped
Make, size change over operations and noise processed obtain multiple new leaf images.Specifically:
First, multiple predetermined angles are set, and according to each predetermined angle to each in the raw image database
History leaf image is rotated to obtain multiple new leaf images.For example:
First, multiple predetermined angles are set, such as can be ± 3 degree and ± 5 degree or other angles, such as
It it is ± 8 degree and ± 10 degree etc., this example is not done specifically limited to this;Secondly, can be according to each predetermined angle, by original graph
As each history leaf image in database is rotated to obtain multiple new leaf images;For example, leaf image A is carried out respectively
Positive hour hands rotate 3 degree and 5 degree, obtain image A1 and A2;Then again by 3 degree and 5 degree of leaf image A rotate counterclockwises,
Obtain image A3 and A4.
Secondly, presetted pixel number is set, and according to the presetted pixel number in the raw image database
Each history leaf image carries out translation and obtains multiple new leaf images.For example:
First, presetted pixel number is set, such as can be 10 pixels or 12 pixels or 20 pixels
Etc., this example is not done specifically limited to this;Secondly, can be according to presetted pixel number, will be each in raw image database
History leaf image translates the presetted pixel number and obtains multiple new leaf images;For example, leaf image A is put down to the left respectively
Move, to right translation, translate up and translate 10 each pixels downwards, obtain leaf image A5, A6, A7 and A8 etc..
Again, set to preset and cut out number of pixels, and number of pixels is cut out to raw image database according to described preset
In each history leaf image be cut out to obtain multiple new leaf images.For example:
First, set to preset and cut out number of pixels, such as can be 10 pixels or 12 pixels or 20
Pixel etc., this example are not done specifically limited to this;Secondly, number of pixels can be cut out according to default, by raw image data
The number of pixels of each history leaf image in storehouse punctures this and default cut out number of pixels and obtain multiple new leaf images respectively;
For example, leaf image A is respectively cut out into 10 pixels up and down, leaf image A9, A10, A11 and A12 etc. can be obtained
Deng.
Finally, salt-pepper noise and white Gaussian noise are set;And according to the salt-pepper noise and white Gaussian noise to institute
State the progress of each history leaf image in raw image database noise processed and obtain multiple new leaf images.For example:
First, salt-pepper noise and white Gaussian noise are set;Wherein, the density of salt-pepper noise can be 0.02 and
0.03, or color density, such as can be 0.04 and 0.06 etc., this example is not done specifically limited to this;Gauss
The average of white noise can be 0, and variance can be 0.01 and 0.02, can also set other averages and variance, this example
This is not done specifically limited;Secondly, can be according to salt-pepper noise and white Gaussian noise to respectively going through in raw image database
History leaf image carries out noise processed and obtains multiple new leaf images;For example, image A is increased respectively density for 0.02 and
0.03 salt-pepper noise, obtain leaf image A13 and A14;Increase image A average respectively again as 0, variance be 0.01 and
0.02 white Gaussian noise, obtain leaf image A15 and A16 etc..
Further, in order that obtaining each leaf image standardization, depth convolutional neural networks can be more applicable for, are also needed
Average value processing is carried out to each new leaf image and history leaf image, can specifically included:To each history leaf
Picture and new leaf image carry out average value processing.For example:
Average value processing is carried out to history leaf image A and new leaf image A1~A16, then again by each leaf image
Scaled is 128 × 128, or other sizes, such as can be 256 × 256 etc., it is special that this example is not done to this
Limitation.
In step S1104, the expansion image data base is formed using each new leaf image.Specifically:
When passing through rotation process, translation, size change over operations and noise respectively to above-mentioned each history leaf image
After processing, each history leaf image and new leaf image are stored into database respectively to obtain expanding image data base;
Further, it is also necessary to according to the class label of history leaf image, new leaf image is marked, looked into facilitating
Ask.By using the mode of above-mentioned expansion raw image database, the amount of images in raw image database is expanded into original
16 times (more than or) come so that depth convolutional neural networks in the next step may learn more images, increase
Add the universality of the depth convolutional neural networks.
In the step s 120, depth convolutional neural networks are built, and using the expansion image data base to the depth
Convolutional neural networks are trained to obtain crop disease identification model.Wherein, with reference to shown in figure 3, depth convolutional Neural net is built
Network can include step S10- steps S70.Wherein:
Step S10, build the level 0 of the depth convolutional neural networks;Wherein, level 0 is input layer;The input
Every width leaf image of the input of layer is the long and wide RGB image for being respectively provided with formed objects.Specifically:
Build the level 0 of depth convolutional neural networks;Level 0 can also be input layer;Wherein, the input layer input
Each leaf image be the long and wide RGB image with identical size;Such as:The size of input picture can be 128*128*3*
Number, that is, input disease leaf image length and it is wide be 128 pixels RGB image, number can represent input leaf
Picture number.
Step S20, build the first layer of the depth convolutional neural networks;Wherein, the first layer includes convolutional layer, most
Great Chiization layer and batch normalization operation;The convolutional layer is used to obtain multiple fisrt feature to RGB image progress convolution
Figure;The maximum pond layer is used to carry out dimension-reduction treatment to the fisrt feature figure;Described batch of normalization operation is used for dimensionality reduction
Fisrt feature figure after processing carries out batch normalized and obtains multiple second feature figures.Specifically:
Build the first layer of depth convolutional neural networks;Wherein, first layer can use convolutional layer+maximum pond layer+criticize
Normalized pattern:Further, convolution, maximum pond and batch normalization operation can be carried out successively in first layer.Wherein, roll up
Lamination can be used for carrying out convolution to RGB image (128*128*1 input data) and obtain the fisrt feature of the RGB image
Figure;Further, this layer has the convolution kernel that 128 window sizes are 9*9, and adjacent local receptor field centre distance can be 4, can
To export 96 fisrt feature figures;And then, dimension-reduction treatment can be carried out to each fisrt feature figure by a maximum pond;Its
In, pond core window size is 2*2, and the centre distance of adjacent local receptor field can be 2;Finally, to each after dimension-reduction treatment
Fisrt feature figure carries out batch normalized, and 128 obtained dimensions are 60*60 second feature figure, then by each second feature
Figure is input to the second layer of convolutional neural networks.It should be added that, receptive field is responsible for can be directly from original defeated herein
Enter image zooming-out lowermost layer feature;Also, the ability of different size convolution kernel extraction low-level features is different, the core size of receptive field
Can be identical with convolution kernel size.
Below, traveling one is entered to the process of convolution, the processing of maximum pondization and batch normalization operation that are mentioned in step S20
The explanation of step and explanation.
First, the process of convolution in step S20 specifically may comprise steps of:
Step S201,128 sizes of random initializtion are 9*9*3 convolution mask;Wherein, the convolution mask can represent
81 weight parameters;
Step S202, connection input neuron and convolutional layer neuron, then by the convolution mask from the upper left corner of image with
The length that step-length is 4 is slided on image;Wherein, a position is often slided into, it is necessary to input neuron and corresponding weight
Parameter seeks the sum of products;Further, by activation primitive, the characteristic pattern that 128 sizes are 120*120*3, Mei Gete are obtained
The corresponding convolution mask of sign figure.Further, corresponding to the value of each pixel can include on the convolution characteristic pattern of l layers
Convolution mask and the image block in corresponding position receptive field region carry out the result of convolution, can specifically be shown below:
Wherein, l is the network number of plies,For the output of l-1 layers, some value of i-th of convolution output matrix is represented,For volume
Product core, MjFor the receptive field of input layer,For a bias of each output figure, the numbering of output matrix corresponding to j expressions,
0 is represented sequentially as from left to right and arrives m, and m represents the number of convolution output matrix, and f represents nonlinear function, used here
Sigmoid activation primitives:
Secondly, the dimension-reduction treatment in step S20 can include:Fisrt feature figure is entered using the method for local maximizing
Row dimension-reduction treatment;Wherein, can be specifically:
Wherein,For weight coefficient,To ask all values in receptive fieldMaximum.
Finally, the dimension-reduction treatment in step S20 can include:Using batch normalization algorithm to adjacent the after dimension-reduction treatment
Excitation in one characteristic pattern is normalized, and obtains multiple second feature figures.Can be specifically:Calculated using batch normalization
Excitation in adjacent feature figure is normalized method, so as to lift the generalization ability of network.The process of algorithm can include:
First, it is assumed that each batch has p feature x1,x2,...,xq, calculate its average is respectively with variance:
Wherein, μ is batch average, and σ is variance;
Secondly, by data normalization, the data that average is 0, variance is 1 are obtained
Wherein, ε is to prevent variance for the invalid small normal amount set of 0 time division type.
Step S30, build the second layer of the depth convolutional neural networks;Wherein, the second layer is used for described the
Two characteristic patterns carry out convolution, maximum pond and batch normalized and obtain multiple third feature figures.Specifically:
Build the second layer of depth convolutional neural networks;Wherein, the second layer can use convolutional layer+maximum pond layer+criticize
Normalized pattern:Further, convolution, maximum pond and batch normalization operation can be carried out successively in the layer, to first layer
The second feature figure that 128 obtained dimensions are 60*60 carries out convolution, maximum pond layer and batch normalized, processing procedure
It is similar with first layer;Further, this layer can include the convolution kernel that 256 window sizes are 5*5, adjacent local experiences Yezhong
Heart distance, which is set, to be 4;Pond core window size is 2*2, and the centre distance of adjacent local receptor field can be 2, obtains criticizing and returns
The third feature figure that 256 dimensions after one change are 28*28, is then input to depth convolutional Neural again by the third feature figure again
The third layer of network.
Step S40, build the third layer of the depth convolutional neural networks;Wherein, the third layer is used for described the
Three characteristic patterns carry out convolution, maximum pond and batch normalized and obtain multiple fourth feature figures.Specifically:
Build the third layer of depth convolutional neural networks;Wherein, third layer can use convolutional layer+maximum pond layer+criticize
Normalized pattern:Further, convolution, maximum pond and batch normalization operation can be being carried out successively in the layer, to second
The third feature figure that 256 dimensions that layer obtains are 28*28 carries out convolution, maximum pond layer and batch normalized, treats
Journey is similar with first layer;Further, this layer can include the convolution kernel that 512 window sizes are 3*3, adjacent local receptor field
Centre distance can be 4;Pond core window size is 2*2, and the centre distance of adjacent local receptor field can be 2, obtains criticizing and returns
512 dimensions after one change are 13*13 fourth feature figure, then the fourth feature figure is input into depth convolutional neural networks
4th layer.
Step S50, build the 4th layer of the depth convolutional neural networks;Wherein, described 4th layer is used for described the
Four characteristic patterns carry out convolution and batch normalized obtains multiple fifth feature figures.Specifically:
Build the 4th layer of depth convolutional neural networks;Wherein, 512 dimensions that this layer can obtain to third layer are
13*13 fourth feature figure carries out convolution+batch normalized, and processing procedure is similar with first layer;Further, the volume of this layer
Product core size can be 3*3, and this layer can include the convolution kernel that 1024 window sizes are 3*3, adjacent local receptor field center
Distance can be 4, obtain the fifth feature figure that 1024 dimensions after batch normalization are 11*11, then the fifth feature figure is defeated
Enter the layer 5 to depth convolutional neural networks.
Step S60, build the layer 5 of the depth convolutional neural networks;Wherein, the layer 5 is used for described the
Five characteristic patterns carry out global average pondization operation and obtain the characteristic value of the fifth feature figure;Wherein, the dimension of the characteristic value
Equal to the quantity of fifth feature figure.For example:
Assuming that the characteristic pattern size of last convolutional layer is m × n (m can be with equal with n value);Wherein, then pass through
After the average pondization operation of the overall situation, the characteristic value of the corresponding output of l width characteristic pattern can be:
Wherein, ylFor the characteristic value of the corresponding output of l width characteristic pattern;
For the characteristic value of l width characteristic patterns.
Step S70, build the layer 6 of the depth convolutional neural networks;Wherein, the layer 6 is used for the spy
Value indicative is classified.Specifically:
Build the layer 6 of depth convolutional neural networks;Wherein, the layer can be used for more uses after the completion of feature extraction
Excitation function of the Softmax regression models of full-mesh as this layer, and the output of each neuron is interpreted as each defeated
Enter the probability of the affiliated species of leaf image, carry out image classification;Further, output node number is equal to class number.Further
, after above-mentioned depth convolutional neural networks structure is completed, using above-mentioned expansion image data base to depth convolutional Neural net
Network is trained to obtain crop disease identification model.
In step s 130, the leaf image of crop to be identified is inputted and obtains spy into the crop disease identification model
Sign vector, and image classification is carried out to the characteristic vector.Specifically:
, can directly will be to be identified when above-mentioned depth convolutional neural networks are after training obtains crop disease identification model
The leaf image of crop is inputted to the crop disease identification model, and obtain the feature of the leaf image of the crop to be identified to
Amount;Then the classification results that obtain leaf image are classified to this feature vector again.
In step S140, obtained according to the class label of the crop to be identified with image classification result described to be identified
The Damage Types of crop.Specifically:
After the classification results of leaf image are obtained, according to the class label and figure of the leaf image of the crop to be identified
As classification results obtain the Damage Types of the crop to be identified so that crop cultivate personnel can be according to the Damage Types to the work
Thing timely treat and handle, in order to avoid the long range diffusion of disease.
In addition, although describing each step of method in the disclosure with particular order in the accompanying drawings, still, this does not really want
These steps must be performed according to the particular order by asking or implying, or the step having to carry out shown in whole could be realized
Desired result.It is additional or alternative, it is convenient to omit some steps, multiple steps are merged into a step and performed, and/
Or a step is decomposed into execution of multiple steps etc..
Those skilled in the art will readily occur to the disclosure its after considering specification and putting into practice invention disclosed herein
Its embodiment.The application is intended to any modification, purposes or the adaptations of the disclosure, these modifications, purposes or
Person's adaptations follow the general principle of the disclosure and including the undocumented common knowledges in the art of the disclosure
Or conventional techniques.Description and embodiments are considered only as exemplary, and the true scope of the disclosure and spirit are by appended
Claim is pointed out.
Claims (10)
- A kind of 1. crop leaf disease recognition method based on depth convolutional neural networks, it is characterised in that including:Raw image database is expanded to obtain and expands image data base;Depth convolutional neural networks are built, and the depth convolutional neural networks are instructed using the expansion image data base Get crop disease identification model;The leaf image of crop to be identified is inputted into the crop disease identification model and obtains characteristic vector, and to the spy Sign vector carries out image classification;The Damage Types of the crop to be identified are obtained according to the class label of the crop to be identified and image classification result.
- 2. crop leaf disease recognition method according to claim 1, it is characterised in that carried out to raw image database Expansion, which obtains expansion image data base, to be included:To in raw image database each history leaf image carry out rotation process, translation, size change over operations and Noise processed obtains multiple new leaf images;The expansion image data base is formed using each new leaf image.
- 3. crop leaf disease recognition method according to claim 2, it is characterised in that in raw image database Each history leaf image progress rotation process, which obtains multiple new leaf images, to be included:Multiple predetermined angles are set, and according to each predetermined angle to each history blade figure in the raw image database As being rotated to obtain multiple new leaf images.
- 4. crop leaf disease recognition method according to claim 2, it is characterised in that in raw image database Each history leaf image progress translation, which obtains multiple new leaf images, to be included:Presetted pixel number is set, and according to the presetted pixel number to each history blade in the raw image database Image carries out translation and obtains multiple new leaf images.
- 5. crop leaf disease recognition method according to claim 2, it is characterised in that in raw image database Each history leaf image progress size change over operations, which obtains multiple new leaf images, to be included:Set to preset and cut out number of pixels, and number of pixels is cut out to each history in raw image database according to described preset Leaf image is cut out to obtain multiple new leaf images.
- 6. crop leaf disease recognition method according to claim 2, it is characterised in that in raw image database Each history leaf image progress noise processed, which obtains multiple new leaf images, to be included:Salt-pepper noise and white Gaussian noise are set;And according to the salt-pepper noise and white Gaussian noise to the original image Each history leaf image in database carries out noise processed and obtains multiple new leaf images.
- 7. the crop leaf disease recognition method according to any one of claim 3~6, it is characterised in that the crop leaf Piece disease recognition method also includes:Average value processing is carried out to each history leaf image and new leaf image.
- 8. crop leaf disease recognition method according to claim 1, it is characterised in that structure depth convolutional neural networks Including:Step S10, build the level 0 of the depth convolutional neural networks;Wherein, level 0 is input layer;The input layer is defeated Each leaf image entered is the long and wide RGB image with identical size;Step S20, build the first layer of the depth convolutional neural networks;Wherein, the first layer includes convolutional layer, maximum pond Change layer and batch normalization operation;The convolutional layer is used to obtain multiple fisrt feature figures to RGB image progress convolution;The maximum pond layer is used for Dimension-reduction treatment is carried out to the fisrt feature figure;Described batch of normalization operation is used to carry out the fisrt feature figure after dimension-reduction treatment Criticize normalized and obtain multiple second feature figures;Step S30, build the second layer of the depth convolutional neural networks;Wherein, the second layer is used for special to described second Sign figure carries out convolution, maximum pond and batch normalized and obtains multiple third feature figures;Step S40, build the third layer of the depth convolutional neural networks;Wherein, the third layer is used for special to the described 3rd Sign figure carries out convolution, maximum pond and batch normalized and obtains multiple fourth feature figures;Step S50, build the 4th layer of the depth convolutional neural networks;Wherein, described 4th layer is used for the described 4th spy Sign figure carries out convolution and batch normalized obtains multiple fifth feature figures;Step S60, build the layer 5 of the depth convolutional neural networks;Wherein, the layer 5 is used for special to the described 5th Sign figure carries out global average pondization operation and obtains the characteristic value of the fifth feature figure;Wherein, the dimension of the characteristic value is equal to The quantity of fifth feature figure;Step S70, build the layer 6 of the depth convolutional neural networks;Wherein, the layer 6 is used for the characteristic value Classified.
- 9. crop leaf disease recognition method according to claim 8, it is characterised in that in the step S20, to institute Stating the progress dimension-reduction treatment of fisrt feature figure includes:Dimension-reduction treatment is carried out to the fisrt feature figure using the method for local maximizing.
- 10. crop leaf disease recognition method according to claim 8, it is characterised in that right in the step S20 Fisrt feature figure after dimension-reduction treatment carries out batch normalized and obtains multiple second feature figures including:The excitation in the adjacent fisrt feature figure after dimension-reduction treatment is normalized using batch normalization algorithm, obtained more Individual second feature figure.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711195140.9A CN107798356A (en) | 2017-11-24 | 2017-11-24 | Crop leaf disease recognition method based on depth convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711195140.9A CN107798356A (en) | 2017-11-24 | 2017-11-24 | Crop leaf disease recognition method based on depth convolutional neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107798356A true CN107798356A (en) | 2018-03-13 |
Family
ID=61535872
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711195140.9A Pending CN107798356A (en) | 2017-11-24 | 2017-11-24 | Crop leaf disease recognition method based on depth convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107798356A (en) |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108537777A (en) * | 2018-03-20 | 2018-09-14 | 西京学院 | A kind of crop disease recognition methods based on neural network |
CN108960310A (en) * | 2018-06-25 | 2018-12-07 | 北京普惠三农科技有限公司 | A kind of agricultural pest recognition methods based on artificial intelligence |
CN109002853A (en) * | 2018-07-18 | 2018-12-14 | 宁夏智启连山科技有限公司 | Crop specie combines recognition methods and device with pest and disease damage type |
CN109086799A (en) * | 2018-07-04 | 2018-12-25 | 江苏大学 | A kind of crop leaf disease recognition method based on improvement convolutional neural networks model AlexNet |
CN109360192A (en) * | 2018-09-25 | 2019-02-19 | 郑州大学西亚斯国际学院 | A kind of Internet of Things field crop leaf diseases detection method based on full convolutional network |
CN110245720A (en) * | 2019-06-22 | 2019-09-17 | 中南林业科技大学 | A kind of citrus pest and disease damage intelligent diagnosing method and system based on deep learning |
CN110457982A (en) * | 2018-12-28 | 2019-11-15 | 中国科学院合肥物质科学研究院 | A kind of crop disease image-recognizing method based on feature transfer learning |
WO2019238072A1 (en) * | 2018-06-13 | 2019-12-19 | 深圳市商汤科技有限公司 | Normalization method, apparatus and device for deep neural network, and storage medium |
CN111122453A (en) * | 2018-11-01 | 2020-05-08 | 阿里巴巴集团控股有限公司 | Information processing method, device and system |
CN112883873A (en) * | 2021-02-22 | 2021-06-01 | 暨南大学 | Leaf disease identification method and device, electronic device and storage medium |
CN116310844A (en) * | 2023-05-18 | 2023-06-23 | 四川凯普顿信息技术股份有限公司 | Agricultural crop growth monitoring system |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106295139A (en) * | 2016-07-29 | 2017-01-04 | 姹ゅ钩 | A kind of tongue body autodiagnosis health cloud service system based on degree of depth convolutional neural networks |
CN106650721A (en) * | 2016-12-28 | 2017-05-10 | 吴晓军 | Industrial character identification method based on convolution neural network |
CN106682704A (en) * | 2017-01-20 | 2017-05-17 | 中国科学院合肥物质科学研究院 | Method of disease image identification based on hybrid convolutional neural network fused with context information |
CN106971160A (en) * | 2017-03-23 | 2017-07-21 | 西京学院 | Winter jujube disease recognition method based on depth convolutional neural networks and disease geo-radar image |
CN107274381A (en) * | 2017-07-20 | 2017-10-20 | 深圳市唯特视科技有限公司 | A kind of dynamic texture synthetic method based on double identification stream convolutional networks |
-
2017
- 2017-11-24 CN CN201711195140.9A patent/CN107798356A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106295139A (en) * | 2016-07-29 | 2017-01-04 | 姹ゅ钩 | A kind of tongue body autodiagnosis health cloud service system based on degree of depth convolutional neural networks |
CN106650721A (en) * | 2016-12-28 | 2017-05-10 | 吴晓军 | Industrial character identification method based on convolution neural network |
CN106682704A (en) * | 2017-01-20 | 2017-05-17 | 中国科学院合肥物质科学研究院 | Method of disease image identification based on hybrid convolutional neural network fused with context information |
CN106971160A (en) * | 2017-03-23 | 2017-07-21 | 西京学院 | Winter jujube disease recognition method based on depth convolutional neural networks and disease geo-radar image |
CN107274381A (en) * | 2017-07-20 | 2017-10-20 | 深圳市唯特视科技有限公司 | A kind of dynamic texture synthetic method based on double identification stream convolutional networks |
Non-Patent Citations (2)
Title |
---|
孙俊 等: "基于改进卷积神经网络的多种植物叶片病害识别", 《农业工程学报》 * |
顾月清 等: "《全国普通高等院校生物医学工程规划教材 生物医学工程技术》", 31 July 2017, 中国医药科技出版社 * |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108537777A (en) * | 2018-03-20 | 2018-09-14 | 西京学院 | A kind of crop disease recognition methods based on neural network |
WO2019238072A1 (en) * | 2018-06-13 | 2019-12-19 | 深圳市商汤科技有限公司 | Normalization method, apparatus and device for deep neural network, and storage medium |
CN108960310A (en) * | 2018-06-25 | 2018-12-07 | 北京普惠三农科技有限公司 | A kind of agricultural pest recognition methods based on artificial intelligence |
CN109086799A (en) * | 2018-07-04 | 2018-12-25 | 江苏大学 | A kind of crop leaf disease recognition method based on improvement convolutional neural networks model AlexNet |
CN109002853A (en) * | 2018-07-18 | 2018-12-14 | 宁夏智启连山科技有限公司 | Crop specie combines recognition methods and device with pest and disease damage type |
CN109002853B (en) * | 2018-07-18 | 2021-10-01 | 宁夏智启连山科技有限公司 | Crop type and pest type combined identification method and device |
CN109360192A (en) * | 2018-09-25 | 2019-02-19 | 郑州大学西亚斯国际学院 | A kind of Internet of Things field crop leaf diseases detection method based on full convolutional network |
CN111122453A (en) * | 2018-11-01 | 2020-05-08 | 阿里巴巴集团控股有限公司 | Information processing method, device and system |
CN110457982A (en) * | 2018-12-28 | 2019-11-15 | 中国科学院合肥物质科学研究院 | A kind of crop disease image-recognizing method based on feature transfer learning |
CN110457982B (en) * | 2018-12-28 | 2023-04-11 | 中国科学院合肥物质科学研究院 | Crop disease image identification method based on feature migration learning |
CN110245720A (en) * | 2019-06-22 | 2019-09-17 | 中南林业科技大学 | A kind of citrus pest and disease damage intelligent diagnosing method and system based on deep learning |
CN112883873A (en) * | 2021-02-22 | 2021-06-01 | 暨南大学 | Leaf disease identification method and device, electronic device and storage medium |
CN112883873B (en) * | 2021-02-22 | 2023-08-01 | 暨南大学 | Leaf disease identification method, device, electronic device and storage medium |
CN116310844A (en) * | 2023-05-18 | 2023-06-23 | 四川凯普顿信息技术股份有限公司 | Agricultural crop growth monitoring system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107798356A (en) | Crop leaf disease recognition method based on depth convolutional neural networks | |
CN107392224A (en) | A kind of crop disease recognizer based on triple channel convolutional neural networks | |
CN106570477B (en) | Vehicle cab recognition model building method and model recognizing method based on deep learning | |
CN105138993B (en) | Establish the method and device of human face recognition model | |
CN107220635A (en) | Human face in-vivo detection method based on many fraud modes | |
CN107292250A (en) | A kind of gait recognition method based on deep neural network | |
CN107977609A (en) | A kind of finger vein identity verification method based on CNN | |
CN107480649A (en) | Fingerprint sweat pore extraction method based on full convolution neural network | |
CN104050507B (en) | Hyperspectral image classification method based on multilayer neural network | |
CN108229381A (en) | Face image synthesis method, apparatus, storage medium and computer equipment | |
CN109489576B (en) | Contour detection method based on primary visual path calculation model | |
CN106203363A (en) | Human skeleton motion sequence Activity recognition method | |
CN112070158B (en) | Facial flaw detection method based on convolutional neural network and bilateral filtering | |
Russel et al. | Leaf species and disease classification using multiscale parallel deep CNN architecture | |
CN108537777A (en) | A kind of crop disease recognition methods based on neural network | |
CN106358444A (en) | A method and a system for face verification | |
CN104408435A (en) | Face identification method based on random pooling convolutional neural network | |
CN107358262A (en) | The sorting technique and sorter of a kind of high-definition picture | |
CN103996056A (en) | Tattoo image classification method based on deep learning | |
CN109741268B (en) | Damaged image complement method for wall painting | |
CN106570522A (en) | Object recognition model establishment method and object recognition method | |
Seyyedsalehi et al. | A fast and efficient pre-training method based on layer-by-layer maximum discrimination for deep neural networks | |
CN106599804A (en) | Retina fovea centralis detection method based on multi-feature model | |
CN109858487A (en) | Weakly supervised semantic segmentation method based on watershed algorithm and image category label | |
CN108664994A (en) | A kind of remote sensing image processing model construction system and 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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180313 |
|
RJ01 | Rejection of invention patent application after publication |