CN108537777A - A kind of crop disease recognition methods based on neural network - Google Patents
A kind of crop disease recognition methods based on neural network Download PDFInfo
- Publication number
- CN108537777A CN108537777A CN201810230177.9A CN201810230177A CN108537777A CN 108537777 A CN108537777 A CN 108537777A CN 201810230177 A CN201810230177 A CN 201810230177A CN 108537777 A CN108537777 A CN 108537777A
- Authority
- CN
- China
- Prior art keywords
- image
- layer
- hidden layer
- hidden
- disease
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/90—Determination of colour characteristics
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30181—Earth observation
- G06T2207/30188—Vegetation; Agriculture
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Evolutionary Biology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of crop disease recognition methods based on neural network, estimates the classification of disease leaf image including the use of the adaptive global pool convolutional neural networks of structure；The adaptive global pool convolutional neural networks are sequentially connected and are constituted by 1 input layer, 1 batch of normalization layer, 6 hidden layers and a classification output layer；Preceding 4 every layer of hidden layers in the hidden layer include convolution operation, activation operation, maximum pondization operation and local acknowledgement's normalization operation, and the 5th hidden layer includes convolution sum activation operation, and the 6th hidden layer operates for global poolization；The present invention can greatly shorten the training required time using autoadapted learning rate, overcome the problem of owing study caused by fixed learning rate and being absorbed in local optimum, improve convergence rate, generalization ability and the stability of network.
Description
Technical field
The present invention relates to crop disease identification technology field more particularly to a kind of crop disease identifications based on neural network
Method.
Background technology
Crop disease has seriously affected the yield and quality of crop.Disease control is an important ring in crop production
Section.Controlling disease is wanted, identification disease species are first had to.Disease leaf portion symptom is the Main Basiss of disease recognition.Based on disease leaf
The disease recognition technique study of picture, which is always that one of computer visual angle, image procossing and machine learning field is important, to be ground
Study carefully project.But since crop leaf diseases color of image, shape and texture are ever-changing, and the blade figure of actual acquisition
The background of picture is more complicated, therefore traditional crop disease recognition methods and technology cannot meet the crop for being actually based on Internet of Things
Leaf diseases monitor the demand of system.In recent years, convolutional neural networks (Convolutional NeuralNetworks, CNN)
It has been found to be one of most effective model in image recognition, in recognition of face, plant species identification and plant disease identification etc.
Aspect achieves research outstanding and application achievements, and is obtained in international large-scale image identification contest (LSVRC-14-15)
Achievement outstanding.CNN can be coloured image directly as input, by learning automatic, implicitly study to figure at many levels
As better deep layer abstract characteristics, the subjective blindness of traditional feature extraction and selection method is overcome, and to image
Geometric transformation, deformation, illumination have a degree of invariance, have stronger adaptability.But it is in practical applications, existing
There is training convergence time length in many CNN, more, the fixed learning rate of training parameter is easily absorbed in local optimum and is susceptible to over-fitting
The problems such as, it cannot be efficiently applied in crop disease identification.
Invention content
It is an object of the invention to solve the problems of the above-mentioned prior art, provide a kind of based on convolutional neural networks
Crop disease recognition methods provides technical support for crop leaf diseases identifying system.In order to overcome traditional crop disease recognition
The discrimination of method is low and the convergence rate of existing CNN is slow, the problems such as being easily absorbed in local optimum, and the present invention provides a kind of adaptive
Answer global pool convolutional neural networks (Adaptive global pooling CNN, AGPCNN).The characteristics of network is：Profit
Full articulamentum and autoadapted learning rate is replaced to substitute traditional global learning rate with global pool layer.AGPCNN by 1 input layer,
1 batch of normalization layer, 6 hidden layers and a classification output layer, wherein each layer of preceding 4 hidden layers includes convolution, activation, pond and returns
One changes operation, and the 5th hidden layer includes convolution sum activation operation, and the 6th hidden layer operates for global poolization.Convolution operation is used for image
Feature extraction, pondization operation be used for network parameter yojan, global poolization operation for reduce network parameter scale and training when
Between, output layer is used for disease blade Classification and Identification.Since softmax functions calculate simply, and value is similar in [0,1]
Probability is adopted so selecting softmax to carry out guidance model study as the loss function of model in output layer during the network operation
Use autoadapted learning rate.
A kind of crop disease recognition methods based on neural network, including the use of the adaptive global pool convolution god of structure
Estimate the classification of disease leaf image through network；
The adaptive global pool convolutional neural networks are by 1 input layer, 1 batch of normalization layer, 6 hidden layers and one
Classification output layer is sequentially connected composition；Preceding 4 every layer of hidden layers in the hidden layer include convolution operation, activation operation, maximum
Pondization operates and local acknowledgement's normalization operation, and the 5th hidden layer includes convolution sum activation operation, and the 6th hidden layer is global pool
Operation；
The convolution operation is used for the feature extraction of image, and pondization operation is used for network parameter yojan, global poolization operation
For reducing network parameter scale and training time, output layer is used for disease blade Classification and Identification.
Further, the crop disease recognition methods based on neural network as described above, the place of described batch of normalization layer
Reason process is：The mean μ and variances sigma of each batch n width images are calculated, i.e., Thus
Each image is normalized：
Wherein, ε is variance, xiFor the i-th width training image,For xiNormalized image, thenMean value be 0, variance is
1；
The normalized imageRestore the feature distribution of original image by following reconstructed operation：
Wherein, yiFor to xiImage after batch normalization, βiAnd γiIn the training process respectively use batch mean value and
Batch variance uses the mean value and variance of all samples respectively during the test.
Further, the crop disease recognition methods based on neural network as described above, first 5 in the hidden layer are hidden
Layer convolution operation process be：The output of convolution operation is expressed as in first of hidden layer
xl=f (Wlxl-1+bl) (3)
Wherein, xl-1For the output of convolutional layer in the l-1 hidden layer, xlFor the output of convolutional layer in first of hidden layer, WlIt indicates
The mapping weight matrix of first of hidden layer, blFor the biasing of first of hidden layer, f () is ReLU activation primitives.
Further, the crop disease recognition methods based on neural network as described above, first 4 in the hidden layer are hidden
Layer local acknowledgement's normalization operation process be：If ai(x, y) is the ith feature figure of maximum pond layer generation in the position (x, y)
Set the excitation at place, oi(x, y) is the response after normalization：
Wherein, n indicates that the number of the same space position adjacent feature figure, N indicate total feature map number, the row of characteristic pattern
Row sequence be arbitrary, this before training pattern it needs to be determined that, parameter k, n, α, β from a series of tests verification concentrate determination,
Default value is k=2, n=5, α=10-4, β=0.75.
Further, the crop disease recognition methods based on neural network as described above, the adaptive global pool
The update mechanism of the autoadapted learning rate of convolutional neural networks is
Wherein, ηt-1For the learning rate of the t-1 times iteration, particularly, as t=1, η0For the initial value of learning rate, setting
It is 0.01.λ is a constant, there is many experiments decision；gtFor gradient of the network in the t times iteration, there is stochastic gradient descent
Method calculates.
Further, the crop disease recognition methods based on neural network as described above is utilizing the adaptive of structure
Global pool convolutional neural networks are come before the classification to estimate disease leaf image, it is necessary first to the disease geo-radar image of acquisition into
Row pretreatment, the pretreated method are：The disease geo-radar image of reading is converted into hsv color pattern, all images are carried out big
Small and brightness normalized, obtains pretreated image；
Then, pretreated image is converted to obtain several new images, then by pretreated image and
Several new images are obtained after transformation and are all normalized to an equal amount of HSV images, are then trained using these images adaptive
Answer global pool convolutional neural networks.
Further, the crop disease recognition methods based on neural network, the method for the transformation include as described above：
(1) Random-Rotation image certain angle；(2) along horizontal or vertical direction flipped image；(3) according to certain ratio
Example zooms in or out image；(4) image is translated in a certain way on the image plane, changes the position of picture material；
(5) size or fog-level of picture material are changed to image filtering using specified scale factor；(6) in the HSV face of image
The colour space changes saturation degree S and V luminance component, keeps tone H constant, increases illumination variation；(7) to each pixel of image
Carry out random perturbation；(8) singular value decomposition is carried out to image, then utilize the corresponding feature of a certain number of larger characteristic values to
Amount restores image；(9) it is weighted by the training image that randomly selects that treated and average generates new training image.
Advantageous effect：
(1) many traditional crop leaf diseases recognition methods generally comprise three steps, and leaf image drops in (a)
The pretreatments such as make an uproar, enhance and filter；(b) designed characteristic of division is extracted from every width leaf image；(c) grader is utilized to know
Other Damage Types.Compared with traditional crop disease recognition methods, AGPCNN proposed by the present invention can be from a large amount of training diseases
Automatically learn effective identification feature in leaf image, so as to avoid image preprocessing and manual features extraction process, overcome
The blindness and limitation of traditional artificial extraction identification feature；
(2) many existing disease recognition methods based on deep learning are stacked by convolutional layer, pond layer, full articulamentum
The network to get up.Convolutional layer carries out linear convolution operation by linear filter, is then generated by nonlinear activation function special
Sign figure.Compared with these existing methods, the AGPCNN that the present invention designs will criticize normalization, local acknowledgement's normalization and global pool
Change be combined can accelerate network convergence, Enhanced feature figure and classification relationship, prevent over-fitting, Dropout parameters avoided to seek
The advantages that excellent, enhancing model generalization ability；
(3) traditional convolutional neural networks are all to accelerate network convergence using fixed learning rate.The present invention uses certainly
Adaptive learning rate can greatly shorten the training required time, overcome and owe study caused by fixed learning rate and be absorbed in part
Optimal problem improves convergence rate, generalization ability and the stability of network.
Description of the drawings
The present invention is based on the crop disease recognition methods schematic diagrams of neural network for the positions Fig. 1.
Specific implementation mode
To make the object, technical solutions and advantages of the present invention clearer, the technical solution below in the present invention carries out clear
Chu is fully described by, it is clear that described embodiments are some of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, those of ordinary skill in the art are obtained every other without creative efforts
Embodiment shall fall within the protection scope of the present invention.
As shown in Figure 1, method provided by the invention includes the following steps：
Embodiment：
Step 1：Image preprocessing.Image is read, and the image of reading is converted into hsv color pattern, to all images
Size and brightness normalized are carried out, to eliminate the influence of illumination and scale size to follow-up recognition methods；Described image is
Crop leaf or other images with disease position.
Step 2：It is converted to obtain several new images per piece image using step 1 is pretreated；The variation
Method include：
(1) Random-Rotation image certain angle；(2) along horizontal or vertical direction flipped image；(3) according to certain ratio
Example zooms in or out image；(4) image is translated in a certain way on the image plane, changes the position of picture material；
(5) size or fog-level of picture material are changed to image filtering using specified scale factor；(6) in the HSV face of image
The colour space changes saturation degree S and V luminance component, keeps tone H constant, increases illumination variation；(7) to each pixel of image
Carry out random perturbation；(8) singular value decomposition is carried out to image, then utilize the corresponding feature of a certain number of larger characteristic values to
Amount restores image；(9) it is weighted by the training image that randomly selects that treated and average generates new training image.
Since convolutional neural networks need great amount of samples to be trained, and the disease leaf image actually obtained is limited, institute
To need to expand existing disease leaf image.Finally, 9 kinds above all of the images and process step 1 obtained
It converts obtained all new images and is all normalized to size as 128 × 128 × 3 HSV images, then one is formed by these images
Larger expansion training image collection.
Step 3：Adaptive global pool convolutional neural networks AGPCNN is built using the expansion training image collection.
The AGPCNN is by 1 input layer, 1 batch of normalization layer, 6 hidden layers and a classification output layer, wherein first 4
Each layer of hidden layer includes convolution, activation, pond and normalization operation, and the 5th hidden layer includes convolution sum activation operation, the 6th hidden layer
It is operated for global poolization.
Wherein, the convolution operation obtains the non-of image for the feature extraction of image then by ReLU excitation functions
Linear character；Pondization operation is used for network parameter yojan, and global poolization is operated for reducing network parameter scale and training time,
Output layer is used for disease blade Classification and Identification, and there are one independent softmax loss functions to come guidance model study, network for the layer
Autoadapted learning rate is used in operational process.The nonlinear characteristic that first four convolution operation obtains is grasped using maximum pondization respectively
Make to carry out compressive features figure, thus simplify the complexity of network, the feature that e-learning arrives is made to have certain rotation, scaling constant
Property.In order to improve the universality of network, local acknowledgement's normalization operation is carried out to each maximum pond result.5th is rolled up
The result of product operation carries out global pool operation, and global pool result input grader is then carried out disease recognition.AGPCNN
Concrete operation step it is as follows：
Step 3.1：It is 128 × 128 × 3 in the size of each image of input layer input, i.e., input picture is that length and width are equal
For the HSV images of 128 pixels；
Step 3.2：Criticizing normalized process is：The mean μ and variances sigma of each batch n width images are calculated, i.e.,Thus each image is normalized：
Wherein, xiFor the i-th width training image,For xiNormalized image.ThenMean value be 0, variance 1.Because simple
Single may destroy data normalization the feature distribution of image, so needing to restore original image by following reconstructed operation
Feature distribution：
Wherein, yiFor to xiImage after batch normalization, βiAnd γiIn the training process respectively use batch mean value and
Batch variance uses the mean value and variance of all samples respectively during the test.
Step 3.3：Training image after the normalization of batch that hidden layer 1 obtains step 3.2 carry out successively convolution operation,
Activation operation, maximum pondization operation and local acknowledgement's normalization operation.Convolution kernel size is 9 × 9, maximum pond window for 2 ×
2, step-length 2, then convolution output characteristic pattern size is 120 × 120 × 3, and maximum pond feature sizes are 60 × 60 × 3, are obtained
32 × 3 characteristic patterns.
Step 3.4：Convolution operation, activation operation, maximum pond are carried out successively to the characteristic pattern that step 3.3 obtains in hidden layer 2
Change operation and local acknowledgement's normalization operation.Convolution kernel size is 9 × 9, and maximum pond window is 2 × 2, step-length 2, then convolution
It is 52 × 52 × 3 to export characteristic pattern size, and maximum pond feature sizes are 26 × 26 × 3, obtain 64 × 3 characteristic patterns.
Step 3.5：Convolution operation, activation operation, maximum pond are carried out successively to the characteristic pattern that step 3.4 obtains in hidden layer 3
Change operation and local acknowledgement's normalization operation.Convolution kernel size is 7 × 7, and maximum pond window is 2 × 2, step-length 2, then convolution
It is 20 × 20 × 3 to export characteristic pattern size, and maximum pond feature sizes are 10 × 10 × 3, obtain 128 × 3 characteristic patterns.
Step 3.6：Convolution operation, activation operation, maximum pond are carried out successively to the characteristic pattern that step 3.5 obtains in hidden layer 4
Change operation and local acknowledgement's normalization operation.Convolution kernel size is 5 × 5, and maximum pond window is 2 × 2, step-length 2, then convolution
It is 6 × 6 × 3 to export characteristic pattern size, and maximum pond feature sizes are 3 × 3 × 3, obtain 256 × 3 characteristic patterns.
Step 3.7：Convolution operation is carried out to the characteristic pattern that step 3.6 obtains in hidden layer 5 and activation operates.Convolution kernel size
It is 3 × 3, then convolution output characteristic pattern size is 1 × 1 × 3, obtains 512 × 1 × 3 characteristic patterns.
Step 3.8：The characteristic pattern obtained to step 3.7 in hidden layer 6 carries out maximum global poolization operation, operating process be from
Maximum value is selected to form a feature vector in every width characteristic pattern that step 3.7 obtains.
Step 3.9：Training image collection after expansion is input to training in AGPCNN structures, finally obtains trained net
Network model.
Step 4：Disease leaf image to be identified is read, renormalization is the HSV images that size is 128 × 128 × 3,
Then obtained AGPCNN is trained to estimate the classification of disease leaf image using step 3.
Convolution operation process in the step 3.3 to step 3.7 is
The output of convolution operation is expressed as in first of hidden layer
xl=f (Wlxl-1+bl) (3)
Wherein, xl-1For the output of convolutional layer in the l-1 hidden layer, xlFor the output of convolutional layer in first of hidden layer, WlIt indicates
The mapping weight matrix of first of hidden layer, blFor the biasing of first of hidden layer, f () is ReLU activation primitives.
Maximum pond operating process in the step 3.3 to step 3.6 is the volume in step 3.3 to step 3.6
With step-length it is 2 to take maximum value in 2 × 2 regions, composition characteristic figure successively on the characteristic pattern extracted by activation after lamination.
The operation can reduce last layer to the number of next layer of input neuron, can also reduce over-fitting, increase the spy to be extracted
Levy dimension.
Local acknowledgement's normalization operation process in the step 3.3 to step 3.6 is：If ai(x, y) is maximum pond
Change excitation of the ith feature figure of layer generation at the position (x, y), oi(x, y) is the response after normalization：
Wherein, n indicates that the number of the same space position adjacent feature figure, N indicate total feature map number, the row of characteristic pattern
Row sequence be arbitrary, this before training pattern it needs to be determined that, parameter k, n, α, β from a series of tests verification concentrate determination,
Default value is k=2, n=5, α=10-4, β=0.75.The main purpose of the operation is that hidden layer is inhibited to export big excitation, is carried
Rise the generalization ability of network model.
The update mechanism of the autoadapted learning rate of the AGPCNN is
Wherein, ηt-1For the learning rate of the t-1 times iteration, particularly, η0For the initial value of the learning rate as t=1, if
It is set to 0.01.λ is a constant, there is many experiments decision.gtFor gradient of the network in the t times iteration, have under stochastic gradient
Drop method calculates.This method can be adjusted correspondingly update by iterative gradient to learning rate so that learning rate was being trained
It is adaptively adjusted in journey.
The beneficial effects of the invention are as follows：
The present invention utilizes the depth convolutional neural networks in deep learning field, improves the crop disease based on disease leaf image
Evil recognition methods efficiently solves the segmentation of disease leaf image and feature extraction problem, improves based on disease leaf image
The discrimination of crop disease recognition methods provides technical support for control of crop disease system；The improvement convolution god of structure simultaneously
In the training process through network, any restrictions are not used and it is assumed that being with good expansibility and higher generalization ability, are had
Conducive to the extensive processing of the magnanimity disease leaf image of Internet of Things acquisition.
Training set and test set and experiment condition under the same conditions, with the newest work based on disease leaf image
Object disease recognition method is compared, experimental result such as the following table 1.It can thus be seen that advantages of the present invention is that discrimination is high, when training
Between and recognition time it is short.
Table 1
Finally it should be noted that：The above embodiments are merely illustrative of the technical solutions of the present invention, rather than its limitations；Although
Present invention has been described in detail with reference to the aforementioned embodiments, it will be understood by those of ordinary skill in the art that：It still may be used
With technical scheme described in the above embodiments is modified or equivalent replacement of some of the technical features；
And these modifications or replacements, various embodiments of the present invention technical solution that it does not separate the essence of the corresponding technical solution spirit and
Range.
Claims (7)
1. a kind of crop disease recognition methods based on neural network, which is characterized in that including the use of the adaptive overall situation of structure
Pond convolutional neural networks estimate the classification of disease leaf image；
The adaptive global pool convolutional neural networks are by 1 input layer, 1 batch of normalization layer, 6 hidden layers and a classification
Output layer is sequentially connected composition；Preceding 4 every layer of hidden layers in the hidden layer include convolution operation, activation operation, maximum pond
Operation and local acknowledgement's normalization operation, the 5th hidden layer include convolution sum activation operation, and the 6th hidden layer operates for global poolization；
The convolution operation is used for the feature extraction of image, and pondization operation is used for network parameter yojan, and global poolization operation is used for
Network parameter scale and training time are reduced, output layer is used for disease blade Classification and Identification.
2. the crop disease recognition methods according to claim 1 based on neural network, which is characterized in that described batch of normalizing
Change layer processing procedure be：The mean μ and variances sigma of each batch n width images are calculated, i.e.,
Thus each image is normalized：
Wherein, ε is variance, xiFor the i-th width training image,For xiNormalized image, thenMean value be 0, variance 1；
The normalized imageRestore the feature distribution of original image by following reconstructed operation：
Wherein, yiFor to xiImage after batch normalization, βiAnd γiUse batch mean value and batch side respectively in the training process
Difference uses the mean value and variance of all samples respectively during the test.
3. requiring method described in 1 according to power, which is characterized in that the convolution operation process of preceding 5 hidden layers in the hidden layer is：
The output of convolution operation is expressed as in first of hidden layer
xl=f (Wlxl-1+bl) (3)
Wherein, xl-1For the output of convolutional layer in the l-1 hidden layer, xlFor the output of convolutional layer in first of hidden layer, WlIndicate l
The mapping weight matrix of a hidden layer, blFor the biasing of first of hidden layer, f () is ReLU activation primitives.
4. requiring the method described in 1 according to power, which is characterized in that the local acknowledgement of preceding 4 hidden layers in the hidden layer normalizes
Operating process is：If ai(x, y) is excitation of the ith feature figure of maximum pond layer generation at the position (x, y), oi(x, y) is
Response after normalization：
Wherein, n indicates that the number of the same space position adjacent feature figure, N indicate that total feature map number, the arrangement of characteristic pattern are suitable
Sequence is arbitrary, this before training pattern it needs to be determined that, parameter k, n, α, β from a series of tests verify concentrate determine, acquiescence
Value is k=2, n=5, α=10-4, β=0.75.
5. requiring method described in 1 according to power, which is characterized in that the adaptive global pool convolutional neural networks it is adaptive
The update mechanism of learning rate is
Wherein, ηt-1For the learning rate of the t-1 times iteration, particularly, as t=1, η0For the initial value of learning rate, it is set as
0.01.λ is a constant, there is many experiments decision；gtFor gradient of the network in the t times iteration, there is stochastic gradient descent method
It calculates.
6. requiring any methods of 1-5 according to power, which is characterized in that in the adaptive global pool convolution god using structure
Through network come before the classification to estimate disease leaf image, it is necessary first to be pre-processed to the disease geo-radar image of acquisition, this is pre-
The method of processing is：The disease geo-radar image of reading is converted into hsv color pattern, size is carried out to all images and brightness normalizes
Processing, obtains pretreated image；
Then, pretreated image is converted to obtain several new images, then by pretreated image and transformation
Several new images are obtained afterwards and are all normalized to an equal amount of HSV images, are then trained using these images adaptive complete
Office's pond convolutional neural networks.
7. requiring method described in 6 according to power, which is characterized in that the method for the transformation includes：
(1) Random-Rotation image certain angle；(2) along horizontal or vertical direction flipped image；(3) it puts according to a certain percentage
Big or downscaled images；(4) image is translated in a certain way on the image plane, changes the position of picture material；(5) sharp
With specified scale factor to image filtering, change the size or fog-level of picture material；(6) in the hsv color of image sky
Between, change saturation degree S and V luminance component, keep tone H constant, increases illumination variation；(7) each pixel of image is carried out
Random perturbation；(8) singular value decomposition is carried out to image, then utilizes the corresponding feature vector pair of a certain number of larger characteristic values
Image is restored；(9) it is weighted by the training image that randomly selects that treated and average generates new training image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810230177.9A CN108537777A (en) | 2018-03-20 | 2018-03-20 | A kind of crop disease recognition methods based on neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810230177.9A CN108537777A (en) | 2018-03-20 | 2018-03-20 | A kind of crop disease recognition methods based on neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108537777A true CN108537777A (en) | 2018-09-14 |
Family
ID=63484276
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810230177.9A Pending CN108537777A (en) | 2018-03-20 | 2018-03-20 | A kind of crop disease recognition methods based on neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108537777A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109191455A (en) * | 2018-09-18 | 2019-01-11 | 西京学院 | A kind of field crop pest and disease disasters detection method based on SSD convolutional network |
| CN109446368A (en) * | 2018-09-21 | 2019-03-08 | 华中科技大学 | A kind of birds search method and system based on classification information |
| CN110458802A (en) * | 2019-06-28 | 2019-11-15 | 天津大学 | Based on the projection normalized stereo image quality evaluation method of weight |
| GR1009898B (en) * | 2020-05-12 | 2021-01-08 | Γεωργιος Αθανασιου Φευγας | Method of detection and evaluation of the biotic-abiotic stress in cultivations via thermal photographs and use of artificial intelligence |
| CN112950700A (en) * | 2021-02-25 | 2021-06-11 | 安徽农业大学 | Plant leaf stomatal conductance measuring method based on deep learning and image processing |
| CN113228041A (en) * | 2018-10-24 | 2021-08-06 | 克莱米特公司 | Detection of infection by plant diseases using improved machine learning |
| CN113920587A (en) * | 2021-11-01 | 2022-01-11 | 哈尔滨理工大学 | Human body posture estimation method based on convolutional neural network |
| CN114881928A (en) * | 2022-04-02 | 2022-08-09 | 合肥工业大学 | Wheat frost disease detection method and system based on deep cost sensitive learning |
| CN116740650A (en) * | 2023-08-10 | 2023-09-12 | 青岛农业大学 | Crop breeding monitoring method and system based on deep learning |
| CN117058132A (en) * | 2023-10-11 | 2023-11-14 | 天津大学 | Cultural relic illumination visual comfort quantitative evaluation method and system based on neural network |
| CN117612694A (en) * | 2023-12-04 | 2024-02-27 | 西安好博士医疗科技有限公司 | Data recognition method and system for thermal therapy machine based on data feedback |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100104191A1 (en) * | 2007-03-26 | 2010-04-29 | Mcgwire Kenneth C | Data analysis process |
| CN107798356A (en) * | 2017-11-24 | 2018-03-13 | 郑州大学西亚斯国际学院 | Crop leaf disease recognition method based on depth convolutional neural networks |
-
2018
- 2018-03-20 CN CN201810230177.9A patent/CN108537777A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100104191A1 (en) * | 2007-03-26 | 2010-04-29 | Mcgwire Kenneth C | Data analysis process |
| CN107798356A (en) * | 2017-11-24 | 2018-03-13 | 郑州大学西亚斯国际学院 | Crop leaf disease recognition method based on depth convolutional neural networks |
Non-Patent Citations (3)
| Title |
|---|
| CRAZYVERTIGO: "【深度学习技术】LRN局部响应归一化", 《HTTPS://BLOG.CSDN.NET/HDUXIEJUN/ARTICLE/DETAILS/70570086》 * |
| 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》 * |
| 张善文 等: "卷积神经网络在黄瓜叶部病害识别中的应用", 《江苏农业学报》 * |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109191455A (en) * | 2018-09-18 | 2019-01-11 | 西京学院 | A kind of field crop pest and disease disasters detection method based on SSD convolutional network |
| CN109446368B (en) * | 2018-09-21 | 2020-07-10 | 华中科技大学 | Bird retrieval method and system based on category information |
| CN109446368A (en) * | 2018-09-21 | 2019-03-08 | 华中科技大学 | A kind of birds search method and system based on classification information |
| CN113228041A (en) * | 2018-10-24 | 2021-08-06 | 克莱米特公司 | Detection of infection by plant diseases using improved machine learning |
| CN113228041B (en) * | 2018-10-24 | 2024-04-09 | 克莱米特有限责任公司 | Detection of infection of plant diseases using improved machine learning |
| CN110458802A (en) * | 2019-06-28 | 2019-11-15 | 天津大学 | Based on the projection normalized stereo image quality evaluation method of weight |
| GR1009898B (en) * | 2020-05-12 | 2021-01-08 | Γεωργιος Αθανασιου Φευγας | Method of detection and evaluation of the biotic-abiotic stress in cultivations via thermal photographs and use of artificial intelligence |
| CN112950700A (en) * | 2021-02-25 | 2021-06-11 | 安徽农业大学 | Plant leaf stomatal conductance measuring method based on deep learning and image processing |
| CN113920587A (en) * | 2021-11-01 | 2022-01-11 | 哈尔滨理工大学 | Human body posture estimation method based on convolutional neural network |
| CN114881928A (en) * | 2022-04-02 | 2022-08-09 | 合肥工业大学 | Wheat frost disease detection method and system based on deep cost sensitive learning |
| CN116740650A (en) * | 2023-08-10 | 2023-09-12 | 青岛农业大学 | Crop breeding monitoring method and system based on deep learning |
| CN116740650B (en) * | 2023-08-10 | 2023-10-20 | 青岛农业大学 | Crop breeding monitoring method and system based on deep learning |
| CN117058132A (en) * | 2023-10-11 | 2023-11-14 | 天津大学 | Cultural relic illumination visual comfort quantitative evaluation method and system based on neural network |
| CN117058132B (en) * | 2023-10-11 | 2024-01-23 | 天津大学 | Cultural relic illumination visual comfort quantitative evaluation method and system based on neural network |
| CN117612694A (en) * | 2023-12-04 | 2024-02-27 | 西安好博士医疗科技有限公司 | Data recognition method and system for thermal therapy machine based on data feedback |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108537777A (en) | A kind of crop disease recognition methods based on neural network | |
| CN107145846B (en) | A kind of insulator recognition methods based on deep learning | |
| CN104778448B (en) | A kind of face identification method based on structure adaptive convolutional neural networks | |
| CN108765319A (en) | A kind of image de-noising method based on generation confrontation network | |
| CN108764195A (en) | Handwriting model training method, hand-written character recognizing method, device, equipment and medium | |
| CN111291696B (en) | Handwriting Dongba character recognition method based on convolutional neural network | |
| CN106022392B (en) | A kind of training method that deep neural network sample is accepted or rejected automatically | |
| CN107392224A (en) | A kind of crop disease recognizer based on triple channel convolutional neural networks | |
| CN106407903A (en) | Multiple dimensioned convolution neural network-based real time human body abnormal behavior identification method | |
| CN111563431A (en) | Plant leaf disease and insect pest identification method based on improved convolutional neural network | |
| CN107748895A (en) | UAV Landing landforms image classification method based on DCT CNN models | |
| CN105068644A (en) | Method for detecting P300 electroencephalogram based on convolutional neural network | |
| CN110827260B (en) | Cloth defect classification method based on LBP characteristics and convolutional neural network | |
| CN106960243A (en) | A kind of method for improving convolutional neural networks structure | |
| CN108520202A (en) | Confrontation robustness image characteristic extracting method based on variation spherical projection | |
| CN109034184A (en) | A kind of grading ring detection recognition method based on deep learning | |
| CN110136162B (en) | Unmanned aerial vehicle visual angle remote sensing target tracking method and device | |
| CN113420794B (en) | Binaryzation Faster R-CNN citrus disease and pest identification method based on deep learning | |
| CN107545279A (en) | Image-recognizing method based on convolutional neural networks Yu Weighted Kernel signature analysis | |
| CN109872326B (en) | Contour detection method based on deep reinforced network jump connection | |
| CN114581560A (en) | Attention mechanism-based multi-scale neural network infrared image colorizing method | |
| CN109190666A (en) | Flowers image classification method based on improved deep neural network | |
| CN112580554A (en) | CNN-based MSTAR data noise intensity control classification identification method | |
| CN107749048A (en) | Image flame detection system and method, pseudoisochromatic plate are as correction system and method | |
| CN109344856A (en) | A kind of off-line signature verification method based on multilayer discriminate feature learning |
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: 20180914 |
|
| RJ01 | Rejection of invention patent application after publication |