CN109344699A - Winter jujube disease recognition method based on depth of seam division convolutional neural networks - Google Patents
Winter jujube disease recognition method based on depth of seam division convolutional neural networks Download PDFInfo
- Publication number
- CN109344699A CN109344699A CN201810957206.1A CN201810957206A CN109344699A CN 109344699 A CN109344699 A CN 109344699A CN 201810957206 A CN201810957206 A CN 201810957206A CN 109344699 A CN109344699 A CN 109344699A
- Authority
- CN
- China
- Prior art keywords
- image
- dcnn
- layer
- winter jujube
- layering
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- 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
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/56—Extraction of image or video features relating to colour
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/60—Type of objects
- G06V20/68—Food, e.g. fruit or vegetables
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Multimedia (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The winter jujube disease recognition method based on depth of seam division convolutional neural networks that the present invention relates to a kind of, includes the following steps: to be pre-processed to obtain scab image to winter jujube disease geo-radar image；Three color components of pretreated scab image are inputted into three layering DCNN respectively；Each layering DCNN is trained；The full articulamentum for establishing each layering DCNN, the neuron of full articulamentum is connect with upper one layer of all neurons；It is layered the characteristic of division that DCNN obtains three component images of scab image by three, is classified using SVM classifier to disease geo-radar image.The present invention obtains the natural characteristic of original sample using depth convolutional neural networks, directly using leaf image sample as training set, can automatically it learn from winter jujube disease geo-radar image to effective feature, the process for extracting characteristic of division from leaf image is eliminated, classical plant classification algorithm is efficiently solved and extracts and selection sort feature problem.
Description
Technical field
It is especially a kind of based on depth of seam division convolutional neural networks the invention belongs to computer image processing technology field
Winter jujube disease recognition method.
Background technique
Winter jujube is usually planted in plastic greenhouse, is had the characteristics that temperature is high, humidity is big in greenhouse, is occurred for winter jujube disease
Suitable environmental condition is provided with sprawling, disease species are more, generation is frequent, and difficulty of prevention and cure is big.Because most of orchard worker lacks
The real time information of weary disease and scientific prevention and cure guidance, so regardless of the fruit tree of oneself is ill disease-free or needs not needing, in 1 year
The four seasons periodically give more than jujube tree sprinkling 20 different pesticides, and especially some orchard worker's increasing amounts are sprayed insecticide more times, to ensure fruit
Disease can never occur for tree.Jujube tree disease can quickly and efficiently be prevented really by doing so, but cause residue of pesticide excessive and
The serious problems such as environmental pollution.
Although there are many winter jujube Defect inspection and diagnostic method, and many methods are combined with map and written form
It is discussed in detail and how to prevent and treat winter jujube pest and disease damage, but these methods are differentiated and examine by the mode of artificial eye observation, comparison
Disconnected disease occurs and Damage Types, the growing environment of the winter jujube of especially different greenhouses differ greatly and each greenhouse in the winter
The form of expression of jujube disease is complicated and changeable, so many orchard workers are difficult correctly to carry out Defect inspection using these methods and type is examined
It is disconnected.There are many crop disease intelligent identification Methods for being based on crop disease (blade or fruit) image at present, most of method is all
It is to extract the characteristic of division being artificially arranged from crop disease image, a classifier is then trained according to the feature extracted,
Test data is recycled to carry out method validation.The validity of these methods, which depends greatly on, artificially to be selected to be characterized in
It is no reasonable.Due to the highly complex diversity of crop disease image, so that the feature that different methods is extracted is different, utilization is existing
Method hundred kinds of features above can be extracted from a width disease geo-radar image, but be difficult to determine which feature is more preferable, thus these
The subjective of method, discrimination be not high, generalization ability is poor.
Deep learning is the new research direction in one, machine learning field, in recent years in computer vision, image and video
The application of the numerous areas such as analysis, speech recognition, multimedia retrieval achieves breakthrough.Deep learning has multilayer non-thread
Property the deep structure penetrated, can pass through and layer-by-layer eigentransformation is carried out to original image, the character representation by image in former space becomes
New feature space is changed to, the character representation of stratification is automatically learned, to obtain the feature for being more advantageous to classification, is overcome
The deficiency of special characteristic is manually extracted in traditional crop disease recognition methods.
How deep learning is applied to freeze disaster disease recognition is problem in the urgent need to address at present.
Summary of the invention
It is an object of the invention to overcome the deficiencies of the prior art and provide a kind of based on depth of seam division convolutional neural networks
Winter jujube disease recognition method, solving that existing winter jujube disease recognition is subjective, discrimination is not high, generalization ability is poor etc. asks
Topic.
The present invention solves its technical problem and adopts the following technical solutions to achieve:
A kind of winter jujube disease recognition method based on depth of seam division convolutional neural networks, comprising the following steps:
Step 1 pre-processes winter jujube disease geo-radar image to obtain scab image；
Three color components of pretreated scab image are inputted three layering DCNN by step 2 respectively；
Step 3 is trained each layering DCNN；
Step 4, the full articulamentum for establishing each layering DCNN, by the neuron of full articulamentum and upper one layer of all nerves
Member connection；
Step 5 is layered the characteristic of division that DCNN obtains three component images of scab image by three, using svm classifier
Device classifies to disease geo-radar image.
Further, the processing method of the step 1 are as follows: winter jujube disease geo-radar image is carried out using vision significance algorithm multiple
Feature channel and multiple dimensioned decomposition, then be filtered to obtain characteristic pattern, characteristic pattern is merged, visual saliency map is generated；
It determines the center of significant graph region and as center, a rectangular area is intercepted in original disease geo-radar image, recycles K- equal
Value clustering procedure clusters salient region, obtains scab image.
Further, the processing method of the step 3 are as follows: when objective function is closer to optimal value using smaller than front
Habit rate is trained；Gradient is calculated using back transfer, the parameter of optimization layering DCNN obtains more optimal sorting category feature；It selects non-
Linear ReLU function is activation primitive, is trained to the data after convolutional Neural layer；It, will be every using batch regularization method
One layer of probability distribution is converted to standardized normal distribution and optimizes；It acquires maximum pond algorithm the feature of acquisition adopt
Sample；Each convolutional layer for being layered DCNN is revised as convolution autocoder, is layered DCNN's using reconstructed error adjustment is minimized
Weight.
Further, in step 4, in conjunction with winter jujube disease fruit image data set, by the output of the full articulamentum of the last layer
Setting 32.
Further, the processing method of the step 5 are as follows:
Firstly, each component image for setting scab image is then rolled up at first as the two dimensional character figure of 64 × 64 sizes
Lamination obtains the two dimensional character of 12 60 × 60 sizes by the picture of 64 × 64 size of convolution nuclear convolution of 5 × 5 sizes
Figure；In first pond layer, the characteristic pattern of 12 30 × 30 sizes is obtained by activation primitive；The contracting that each sub-sampling layer uses
Putting the factor is 2, each sub-sampling characteristic pattern needs 2 parameters of training；
Secondly, the convolution kernel that second convolutional layer is 5 × 5 using size, then the characteristic pattern size obtained is 26 × 26；The
Each characteristic pattern in two convolutional layers when making convolution, is combined by several characteristic patterns in the first pond layer or whole characteristic patterns
Input carries out convolution again and obtains；Remaining convolutional layer and the data handling procedure of sub-sampling layer and the layer of front are essentially identical；Through
The convolution sum down-sampling for crossing multilayer, extracts the characteristic of division of image；
Finally, the characteristic of division dimension that three layering DCNN are obtained is 32 × 3=96, thus the SVM of training classification layer divides
Class device classifies to disease geo-radar image.
The advantages and positive effects of the present invention are:
1, the present invention obtains the natural characteristic of original sample using depth convolutional neural networks, directly by leaf image sample
As training set, it can automatically learn from winter jujube disease geo-radar image to effective feature, eliminate and extracted from leaf image
The process of characteristic of division efficiently solves classical plant classification algorithm and extracts and selection sort feature problem.
2, for the present invention on the basis of DCNN, using the winter jujube disease recognition method based on layering DCNN, this method can
Directly carry out winter jujube disease recognition using colored winter jujube disease fruit image, solve existing winter jujube disease recognition it is subjective,
The problems such as discrimination is not high, generalization ability is poor.
Detailed description of the invention
Fig. 1 is that the present invention arranges process flow diagram；
Fig. 2 a is the image instance that winter jujube suffers from anthracnose；
Fig. 2 b is the image instance that winter jujube suffers from decayed fruit disease；
Fig. 2 c is the image instance that winter jujube suffers from fruit-shrink disease；
Fig. 2 d is the image instance that winter jujube suffers from diplostomiasis；
Fig. 2 e is the multiple image by a sub-picture through rotating and disturbance treatment obtains；
Fig. 2 f is the multiple image handled by a sub-picture through brightness change；
Fig. 3 a is the original image before process of convolution；
Fig. 3 b is the channel first time R trellis diagram；
Fig. 3 c is first time convolution composite diagram；
Fig. 3 d is the third secondary volume collection figure in the channel R；
Fig. 3 e is third secondary volume collection figure；
Fig. 4 is the winter jujube disease recognition procedure chart based on layering DCNN.
Specific embodiment
The embodiment of the present invention is further described below in conjunction with attached drawing:
The present invention carries out winter jujube disease recognition using depth convolutional Neural network technology.Layering DCNN is implicitly from training number
Feature is differentiated according to middle study, and the neuron weight on same Feature Mapping face is identical, so as to collateral learning, avoids
Explicit characteristic extraction procedure in traditional images identification.The weight sharing policy of layering DCNN reduces the complexity of network
Degree.Meanwhile the same characteristic pattern uses identical convolution kernel in feature extraction, different characteristic patterns uses different convolution
Core.Convolutional layer preserves different local features, so that the feature extracted is provided with rotation, translation invariance.Layering
The training process of DCNN, which is divided into, to be propagated and two stages of back-propagation forward.In propagation stage forward, DCNN is layered from sample set
It is middle to extract a label as YmSample X input network, information from input layer by step by step transformation is transmitted to feature output layer, count
Calculate corresponding reality output Om:
Om=fn(...(f2(f1(XW1)W2)...)Wn) (1)
Wherein, f () is an activation primitive, Wi(i=1,2 ..., n) is trained mapping weight matrix.
This process is the process that network is executed when operating normally after completing training.In the process, model carries out defeated
Enter the weight matrix phase dot product with every layer, obtains output result to the end.In actual operation, convolution kernel is used for image block
Convolution is carried out, needs to increase a bias term, is then in the convolution algorithm of l layers of output
xl=f (Wlxl-1+bl) (2)
Wherein, l is the model number of plies, WlExpression has trained the mapping weight matrix of current layer, blFor the additivity of "current" model
Biasing.
In the back-propagation stage, the difference of reality output and ideal output is calculated, and reversed by the method for minimization error
Propagate adjustment weight matrix.
It is layered DCNN and independent down-sampling operation is generally carried out to each characteristic pattern using average pond or maximum pond algorithm.
Mean value of the average pondization according to pixel in the neighborhood window calculation particular range of definition, neighborhood window translating step are greater than 1 and (are less than
Equal to the size of pond window), and mean value is then replaced with most value and is output to the next stage by maximum pondization.It is defeated after pondization operation
The resolution ratio of characteristic pattern reduces out, but the feature that can preferably keep high-resolution features figure to describe.In practical applications, it is layered
DCNN can remove down-sampling process, reach reduction resolution greater than 1 by the way that convolution kernel window sliding step-length is arranged in the convolution stage
The purpose of rate.
Layering DCNN is adjusted output by activation primitive after completing convolution mapping, the feature that convolutional layer is extracted
It is inputted as function, carries out Nonlinear Mapping.Common activation primitive has the saturation nonlinearities function such as sigmoid and tanh.Closely
Unsaturated nonlinear function ReLU is widely used within several years.In training gradient decline, ReLU is than traditional saturation nonlinearity function
Convergence rate faster, therefore training whole network when, training speed is more very fast than traditional method.
Based on above description, a kind of winter jujube disease recognition method based on depth of seam division convolutional neural networks of the invention,
As shown in Figure 1, comprising the following steps:
Step 1 pre-processes winter jujube disease geo-radar image to obtain scab image.
In this step, since the collected disease geo-radar image of internet of things sensors has stronger noise jamming, Wo Menli
With simple vision significance algorithm (Itti algorithm), multiple feature channels and multiple dimensioned point are carried out to winter jujube disease geo-radar image
Solution, then be filtered to obtain characteristic pattern, characteristic pattern is merged, visual saliency map is generated.Determine the center of significant graph region,
The point centered on the center of significant graph region intercepts a rectangular area in original disease geo-radar image, recycles K- mean cluster
Method clusters salient region, obtains scab image.
Due to the complicated variety of disease geo-radar image, so that the preprocessing process of image is more complicated, time-consuming, so we taste
An area-of-interest of disease geo-radar image is extracted in examination, as the input of recognition methods, omits background separation and scab cutting procedure.
Three color components of pretreated scab image are inputted three layering DCNN by step 2 respectively.
Step 3 is trained each layering DCNN.
In the training process, it is trained, is led to using the learning rate smaller than front when objective function is closer to optimal value
Cross the classification performance that adjusting parameter improves layering DCNN.Gradient is calculated using back transfer, optimization is layered the parameters of DCNN,
Obtain more optimal sorting category feature.Selecting nonlinear ReLU function is activation primitive, is trained to the data after convolutional Neural layer.
Convolution kernel appropriate is set, the performance of layering DCNN is improved.The size of convolution kernel determines the size of a neuron receptive field, if
Convolution kernel is too small, then cannot extract effective local feature；But if convolution kernel is excessive, and the complexity for the feature extracted may be remote
Far more than the expression ability of convolution kernel.Using batch regularization method, each layer of probability distribution is converted into standard normal point
Cloth optimizes.It acquires maximum pond algorithm and down-sampling is carried out to the feature of acquisition.Each convolutional layer for being layered DCNN is revised as
Convolution autocoder utilizes the weight for minimizing reconstructed error adjustment layering DCNN.Other parameters are chosen for Matlab
Default parameter value in 2017a in deep learning tool box (Deep Learn Tool Box).
Step 4, the full articulamentum for establishing each layering DCNN.
The neuron of full articulamentum is connect with upper one layer of all neurons.It can be generated due to connecting layer parameter entirely excessively
Fitting.In order to solve this problem, in conjunction with winter jujube disease fruit image data set, the output for adjusting the full articulamentum of the last layer is
32, the node parameter of full articulamentum can be reduced in this way.Formula (3) is recognition methods error:
Wherein, n, m are respectively the total sample number and classification number of training sample, and f is m × 1 obtained by excitation function
Output matrix, flabelIt is the two values matrix of m × 1 for training sample label.
The superiority and inferiority of network model training is mainly determined by loss function, and the lower model training of penalty values with test is trained
It is better to obtain, and whole network is restrained in the training stage.
Step 5 classifies to disease geo-radar image:
In step, it is layered the characteristic of division that DCNN obtains three component images of scab image by three, in classification layer
It is a feature vector by these feature integrations, then is classified using the SVM classifier of comparative maturity to disease geo-radar image.
The convolution nuclear structure in convolutional layer of the present invention is analyzed below:
Assuming that each component image of scab image is the two dimensional character figure of 64 × 64 sizes, then in first convolutional layer
(C1) by the picture of 64 × 64 size of convolution nuclear convolution of 5 × 5 sizes, the two dimensional character of 12 60 × 60 sizes is obtained
Figure.Wherein, 5 × 5 convolution kernel that the same characteristic pattern uses is identical.It is 12 × (5 × 5+1) that C1, which needs the number of parameters of training,
=312, and the connection number of input layer and C1 are 312 × (60 × 60)=1123200.In first pond layer (S1), lead to
The activation primitive for crossing formula (1) obtains the characteristic pattern of 12 30 × 30 sizes, i.e., by by 2 × 2 not overlapped all in C1
Sub-block x summation, multiplied by a weight w, in addition what a bias term b was obtained.Because characteristic pattern size is 60 × 60 in C1, institute
The feature subgraph for being 30 × 30 with obtained sub-sampling result.Then, the zoom factor that each sub-sampling layer uses is 2, mesh
Control scaling decline speed because scaling be exponential scaling, the speed of diminution also implies that very much extraction characteristics of image fastly
It is more coarse, it will to lose more image detail features.Each sub-sampling characteristic pattern needs 2 parameters of training, so S1 is needed
Train 12 × 2=24 parameter.
It is similar with C1 in second convolutional layer (C2).It is 5 × 5 convolution kernel, the then characteristic pattern obtained that C2, which also uses size,
Size is 26 × 26.After C1 and S1, the receptive field of each neuron covering of S1 is equivalent to the 10 × 10 of original image
(i.e. the convolution kernel of C1 is 5 × 5, and the sampling sub-block size of second pond layer (S2) is 2 × 2, then 5 × 5 × 2 × 2=10 ×
10).Convolution kernel of the C2 Jing Guo 5 × 5 sizes extracts the feature of S1, its receptive field further expansion is equivalent to original image
50×50.C1 obtains 12 mapped plans by one picture of input layer, and present C2 needs are mapped out from 12 characteristic patterns of S1
24 characteristic patterns, need exist for certain skill.Each characteristic pattern in C2 when making convolution, be by characteristic patterns several in S1 or
Whole characteristic patterns are combined into input, then carry out convolution and obtain.The data handling procedure of remaining convolutional layer and sub-sampling layer is with before
The layer in face is essentially identical.Convolution sum down-sampling by multilayer, the feature of extraction is more abstract, also has more ability to express.
Being layered the characteristic of division dimension that DCNN is obtained by three is 32 × 3=96.Thus the svm classifier of training classification layer
Device obtains trained winter jujube disease recognition method.Then model performance test is carried out using test data.
In order to show the validity of winter jujube disease recognition method proposed by the present invention, tested according to the method described above, into
And examine effect of the invention:
In the common anthracnose of Shaanxi Dali County winter jujube greenhouse garden acquisition winter jujube, decayed fruit disease, fruit-shrink disease, four kinds of diplostomiasis
Each 100 width image of disease fruit, four kinds of disease fruit examples are as shown in Fig. 2 a, Fig. 2 b, Fig. 2 c, Fig. 2 d.Experiment condition is CPU:
Intel Corei3-2120,8G and Windows 64, software platform are Matlab 2017a and Matlab deep learning work
Have case (Deep Learn Toolbox-master).Due to currently without disclosed winter jujube disease geo-radar image database, and depth
Practising model needs a large amount of sample to carry out model training, so we pass through the side such as rotation, color and brightness change, scaled
Each image is extended for 50 width images by formula, thus simulates shooting condition under a variety of environment based on Internet of Things monitor video, then
Disease geo-radar image 4 × 100 × 50=20000 width is obtained, a sense comprising most of scab for then extracting each image is emerging
Interesting region, then each image is cut to 64 × 64 image of size, recycle Itti algorithm to combine with K- means Method
Image partition method divides every width disease geo-radar image, obtains corresponding scab image, Fig. 2 e give by a sub-picture through rotation and
The multiple image that disturbance treatment obtains, Fig. 2 f give the multiple image handled by a sub-picture through brightness change.
The trellis diagram of the layered DCNN of one sub-picture is as shown in Fig. 3 a, Fig. 3 b, Fig. 3 c, Fig. 3 d and Fig. 3 e.From above-mentioned volume collection
Figure is as can be seen that layering DCNN can obtain the essential abstract characteristics of image.
Concentrated in these image datas, arbitrarily selected from every kind of scab image 4000 width totally 16000 width as training set,
Remaining totally 4000 width as test set.Training set is for training DCNN model and SVM classifier, and test set is for testing institute
The performance of the disease recognition method of proposition.Its disease recognition process is as shown in Figure 4.
Three components Rs of the colored RGB scab image after the segmentation of every width, G and B image are separately input to layering DCNN.
The feature vector of one 96 dimension is obtained in full articulamentum.The scab image training DCNN of training set, obtained feature vector are instructed again
Practice SVM classifier.It is inputted after being layered DCNN model stability, then by test set, recognition result is obtained by SVM.Test result is such as
Under:
We are compared with three kinds of fruit disease recognition methods: improved and poor histogram (ISDH) method is based on face
Color, texture and shape feature (CTS) method and be based on image procossing (IP) method.Obtained correct recognition rata and variance such as table 1
It is shown.
For two kinds of situations, repeats in experiment above: (1) making the region of interest area image for the disease geo-radar image that do not divide
For input；(2) image of 1/10th numbers is tested in the database built using us, obtains the results are shown in Table 1.
1. 4 kinds of methods of table are to the scab image of segmentation and the discrimination (%) and variance of original disease geo-radar image
As can be seen from Table 1, the classification results of the present invention and other traditional three kinds of methods compare, in the former two cases
The classification accuracy rate of method proposed by the invention is significantly improved.Due to similar in the class of the winter jujube disease geo-radar image of our construction
Degree is higher, so four kinds of methods are relatively high to the discrimination of the scab image after segmentation.But since the complexity of disease geo-radar image is more
Sample, using general feature extracting method, it is difficult to extract arrive preferable feature, and rotation of traditional recognition methods to image
Turn more sensitive with illumination etc., so the recognition effect of other the three kinds disease recognition methods based on feature extraction is poor.And by
It can learn in deep learning model to useful feature is compared, the recognition effect obtained in the process of the present invention is preferable.But work as
When training sample number is reduced to original 1/10, the recognition effect decline of other three kinds of methods is little, and proposed by the present invention
The recognition effect of recognition methods is decreased obviously, the reason is that cannot train the stable mould of performance without enough training samples
Type, so recognition effect is poor.By being comprehensively compared, on large database, winter jujube disease disaggregated model proposed by the present invention
There is apparent advantage.
It is emphasized that embodiment of the present invention be it is illustrative, without being restrictive, therefore packet of the present invention
Include and be not limited to embodiment described in specific embodiment, it is all by those skilled in the art according to the technique and scheme of the present invention
The other embodiments obtained, also belong to the scope of protection of the invention.
Claims (5)
1. a kind of winter jujube disease recognition method based on depth of seam division convolutional neural networks, it is characterised in that the following steps are included:
Step 1 pre-processes winter jujube disease geo-radar image to obtain scab image；
Three color components of pretreated scab image are inputted three layering DCNN by step 2 respectively；
Step 3 is trained each layering DCNN；
Step 4, the full articulamentum for establishing each layering DCNN, the neuron of full articulamentum and upper one layer of all neurons are connected
It connects；
Step 5 is layered the characteristic of division that DCNN obtains three component images of scab image by three, using SVM classifier pair
Disease geo-radar image is classified.
2. the winter jujube disease recognition method according to claim 1 based on depth of seam division convolutional neural networks, feature exist
In the processing method of the step 1 are as follows: carry out multiple feature channels and more to winter jujube disease geo-radar image using vision significance algorithm
The decomposition of scale, then be filtered to obtain characteristic pattern, characteristic pattern is merged, visual saliency map is generated；Determine notable figure area
The center in domain and as center, a rectangular area is intercepted in original disease geo-radar image, recycles K- means Method to aobvious
Work property region is clustered, and scab image is obtained.
3. the winter jujube disease recognition method according to claim 1 based on depth of seam division convolutional neural networks, feature exist
In: the processing method of the step 3 are as follows: instructed when objective function is closer to optimal value using the learning rate smaller than front
Practice；Gradient is calculated using back transfer, the parameter of optimization layering DCNN obtains more optimal sorting category feature；Select nonlinear ReLU
Function is activation primitive, is trained to the data after convolutional Neural layer；Using batch regularization method, by each layer of probability
Distribution is converted to standardized normal distribution and optimizes；It acquires maximum pond algorithm and down-sampling is carried out to the feature of acquisition；It will layering
Each convolutional layer of DCNN is revised as convolution autocoder, utilizes the weight for minimizing reconstructed error adjustment layering DCNN.
4. the winter jujube disease recognition method according to claim 1 based on depth of seam division convolutional neural networks, feature exist
In: in step 4, in conjunction with winter jujube disease fruit image data set, the output of the full articulamentum of the last layer is arranged 32.
5. the winter jujube disease recognition method according to claim 1 based on depth of seam division convolutional neural networks, feature exist
In: the processing method of the step 5 are as follows:
Firstly, setting each component image of scab image as the two dimensional character figure of 64 × 64 sizes, then in first convolutional layer
By the picture of 64 × 64 size of convolution nuclear convolution of 5 × 5 sizes, the two dimensional character figure of 12 60 × 60 sizes is obtained；
In first pond layer, the characteristic pattern of 12 30 × 30 sizes is obtained by activation primitive；The scaling that each sub-sampling layer uses
The factor is 2, each sub-sampling characteristic pattern needs 2 parameters of training；
Secondly, the convolution kernel that second convolutional layer is 5 × 5 using size, then the characteristic pattern size obtained is 26 × 26；Volume Two
Each characteristic pattern in lamination is to be combined into input by several characteristic patterns in the first pond layer or whole characteristic patterns when making convolution
Convolution is carried out again to obtain；Remaining convolutional layer and the data handling procedure of sub-sampling layer and the layer of front are essentially identical；Through excessive
The convolution sum down-sampling of layer, extracts the characteristic of division of image；
Finally, the characteristic of division dimension that three layering DCNN are obtained is 32 × 3=96, the SVM classifier of classification layer is thus trained
Classify to disease geo-radar image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810957206.1A CN109344699A (en) | 2018-08-22 | 2018-08-22 | Winter jujube disease recognition method based on depth of seam division convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810957206.1A CN109344699A (en) | 2018-08-22 | 2018-08-22 | Winter jujube disease recognition method based on depth of seam division convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109344699A true CN109344699A (en) | 2019-02-15 |
Family
ID=65296641
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810957206.1A Pending CN109344699A (en) | 2018-08-22 | 2018-08-22 | Winter jujube disease recognition method based on depth of seam division convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109344699A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109871905A (en) * | 2019-03-14 | 2019-06-11 | 同济大学 | A kind of plant leaf identification method based on attention mechanism depth model |
| CN111506728A (en) * | 2020-04-16 | 2020-08-07 | 太原科技大学 | Hierarchical structure text automatic classification framework based on HD-MSCNN |
| CN111598001A (en) * | 2020-05-18 | 2020-08-28 | 哈尔滨理工大学 | Apple tree pest and disease identification method based on image processing |
| CN111709427A (en) * | 2020-05-26 | 2020-09-25 | 淮阴工学院 | Fruit segmentation method based on sparse convolution kernel |
| CN112001365A (en) * | 2020-09-22 | 2020-11-27 | 四川大学 | High-precision crop disease and insect pest identification method |
| CN112434662A (en) * | 2020-12-09 | 2021-03-02 | 安徽大学 | Tea leaf scab automatic identification algorithm based on multi-scale convolutional neural network |
| CN113112498A (en) * | 2021-05-06 | 2021-07-13 | 东北农业大学 | Grape leaf scab identification method based on fine-grained countermeasure generation network |
| CN113469064A (en) * | 2021-07-05 | 2021-10-01 | 安徽大学 | Method and system for identifying corn leaf disease images in complex environment |
| CN114185431A (en) * | 2021-11-24 | 2022-03-15 | 安徽新华传媒股份有限公司 | Intelligent media interaction method based on MR technology |
| CN114549939A (en) * | 2022-04-26 | 2022-05-27 | 汉谷云智(武汉)科技有限公司 | Tree pest and disease picture classification method based on image feature extraction |
| CN115994907A (en) * | 2023-03-22 | 2023-04-21 | 济南市莱芜区综合检验检测中心 | Intelligent processing system and method for comprehensive information of food detection mechanism |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| CN107248150A (en) * | 2017-07-31 | 2017-10-13 | 杭州电子科技大学 | A kind of Multiscale image fusion methods extracted based on Steerable filter marking area |
| CN107392224A (en) * | 2017-06-12 | 2017-11-24 | 天津科技大学 | A kind of crop disease recognizer based on triple channel convolutional neural networks |
-
2018
- 2018-08-22 CN CN201810957206.1A patent/CN109344699A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| CN107392224A (en) * | 2017-06-12 | 2017-11-24 | 天津科技大学 | A kind of crop disease recognizer based on triple channel convolutional neural networks |
| CN107248150A (en) * | 2017-07-31 | 2017-10-13 | 杭州电子科技大学 | A kind of Multiscale image fusion methods extracted based on Steerable filter marking area |
Non-Patent Citations (2)
| Title |
|---|
| LAURENT ITTI, ET AL.: "A Model of Saliency-Based Visual Attention for Rapid Scene Analysis", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
| 张善文 等: "基于物联网和深度卷积神经网络的冬枣病害识别方法", 《浙江农业学报》 * |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109871905A (en) * | 2019-03-14 | 2019-06-11 | 同济大学 | A kind of plant leaf identification method based on attention mechanism depth model |
| CN111506728A (en) * | 2020-04-16 | 2020-08-07 | 太原科技大学 | Hierarchical structure text automatic classification framework based on HD-MSCNN |
| CN111506728B (en) * | 2020-04-16 | 2023-06-06 | 太原科技大学 | Hierarchical structure text automatic classification method based on HD-MSCNN |
| CN111598001B (en) * | 2020-05-18 | 2023-04-28 | 哈尔滨理工大学 | Identification method for apple tree diseases and insect pests based on image processing |
| CN111598001A (en) * | 2020-05-18 | 2020-08-28 | 哈尔滨理工大学 | Apple tree pest and disease identification method based on image processing |
| CN111709427A (en) * | 2020-05-26 | 2020-09-25 | 淮阴工学院 | Fruit segmentation method based on sparse convolution kernel |
| CN112001365A (en) * | 2020-09-22 | 2020-11-27 | 四川大学 | High-precision crop disease and insect pest identification method |
| CN112434662B (en) * | 2020-12-09 | 2023-04-18 | 安徽大学 | Tea leaf scab automatic identification algorithm based on multi-scale convolutional neural network |
| CN112434662A (en) * | 2020-12-09 | 2021-03-02 | 安徽大学 | Tea leaf scab automatic identification algorithm based on multi-scale convolutional neural network |
| CN113112498A (en) * | 2021-05-06 | 2021-07-13 | 东北农业大学 | Grape leaf scab identification method based on fine-grained countermeasure generation network |
| CN113112498B (en) * | 2021-05-06 | 2024-01-19 | 东北农业大学 | Grape leaf spot identification method based on fine-grained countermeasure generation network |
| CN113469064A (en) * | 2021-07-05 | 2021-10-01 | 安徽大学 | Method and system for identifying corn leaf disease images in complex environment |
| CN113469064B (en) * | 2021-07-05 | 2024-03-29 | 安徽大学 | Identification method and system for corn leaf disease image in complex environment |
| CN114185431A (en) * | 2021-11-24 | 2022-03-15 | 安徽新华传媒股份有限公司 | Intelligent media interaction method based on MR technology |
| CN114185431B (en) * | 2021-11-24 | 2024-04-02 | 安徽新华传媒股份有限公司 | Intelligent media interaction method based on MR technology |
| CN114549939A (en) * | 2022-04-26 | 2022-05-27 | 汉谷云智(武汉)科技有限公司 | Tree pest and disease picture classification method based on image feature extraction |
| CN115994907A (en) * | 2023-03-22 | 2023-04-21 | 济南市莱芜区综合检验检测中心 | Intelligent processing system and method for comprehensive information of food detection mechanism |
| CN115994907B (en) * | 2023-03-22 | 2023-06-06 | 济南市莱芜区综合检验检测中心 | Intelligent processing system and method for comprehensive information of food detection mechanism |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109344699A (en) | Winter jujube disease recognition method based on depth of seam division convolutional neural networks | |
| AU2020102885A4 (en) | Disease recognition method of winter jujube based on deep convolutional neural network and disease image | |
| CN110148120B (en) | Intelligent disease identification method and system based on CNN and transfer learning | |
| CN107016405B (en) | A kind of pest image classification method based on classification prediction convolutional neural networks | |
| CN107657281A (en) | A kind of image-recognizing method based on improved convolutional neural networks | |
| CN107392130A (en) | Classification of Multispectral Images method based on threshold adaptive and convolutional neural networks | |
| CN107330892A (en) | A kind of sunflower disease recognition method based on random forest method | |
| CN106845497B (en) | Corn early-stage image drought identification method based on multi-feature fusion | |
| CN107145889A (en) | Target identification method based on double CNN networks with RoI ponds | |
| CN108734719A (en) | Background automatic division method before a kind of lepidopterous insects image based on full convolutional neural networks | |
| CN107563389A (en) | A kind of corps diseases recognition methods based on deep learning | |
| CN109325495A (en) | A kind of crop image segmentation system and method based on deep neural network modeling | |
| CN111127423B (en) | Rice pest and disease identification method based on CNN-BP neural network algorithm | |
| CN110287882A (en) | A kind of big chrysanthemum kind image-recognizing method based on deep learning | |
| CN110969182A (en) | Convolutional neural network construction method and system based on farmland image | |
| CN114898359B (en) | Litchi plant diseases and insect pests detection method based on improvement EFFICIENTDET | |
| CN115861686A (en) | Litchi key growth period identification and detection method and system based on edge deep learning | |
| CN115631462A (en) | AM-YOLOX-based strawberry disease and pest detection method and system | |
| CN114758132A (en) | Fruit tree pest and disease identification method and system based on convolutional neural network | |
| Miao et al. | Crop weed identification system based on convolutional neural network | |
| Liu et al. | “Is this blueberry ripe?”: a blueberry ripeness detection algorithm for use on picking robots | |
| Dahiya et al. | An effective detection of litchi disease using deep learning | |
| CN115862003A (en) | Lightweight YOLOv 5-based in-vivo apple target detection and classification method | |
| Javeriya | Faster-RCNN based deep learning model for pomegranate diseases detection and classification | |
| CN114511849A (en) | Grape thinning identification method based on graph attention network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190215 |
|
| WD01 | Invention patent application deemed withdrawn after publication |