CN108549910A - A kind of corn seed producing fruit ear image preliminary classification method based on convolutional neural networks - Google Patents
A kind of corn seed producing fruit ear image preliminary classification method based on convolutional neural networks Download PDFInfo
- Publication number
- CN108549910A CN108549910A CN201810344463.8A CN201810344463A CN108549910A CN 108549910 A CN108549910 A CN 108549910A CN 201810344463 A CN201810344463 A CN 201810344463A CN 108549910 A CN108549910 A CN 108549910A
- Authority
- CN
- China
- Prior art keywords
- corn ear
- image
- convolutional neural
- neural networks
- corn
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
Abstract
The present invention provides a kind of corn seed producing fruit ear image preliminary classification method based on convolutional neural networks, including：Obtain corn ear image original training sample collection and test sample collection；By transfer learning method, using AlexNet convolutional neural networks, type belonging to the affiliated type of each fruit ear test sample and the reality of each test sample is obtained according to differentiating, determines the accuracy rate for differentiating result；If accuracy rate outside preset range, according to amplification fruit ear training sample set, optimizes AlexNet convolutional neural networks, obtains the second convolutional neural networks, and differentiates that the test sample concentrates each affiliated type of test sample again.The present invention from hidden layer autonomous learning fruit ear image by the low layers such as color, side to angle point, shape contour level feature by way of, avoid artificial extraction fruit ear characteristics of image it is cumbersome with it is unilateral, make convolutional neural networks that there is autonomous selection characteristics of image and learnt, recognition capability, method is provided for the automation fringe choosing of corn seed producing fruit ear primary.
Description
Technical field
The present invention relates to area of pattern recognition, more particularly, to a kind of corn seed producing fruit based on convolutional neural networks
Fringe image preliminary classification method.
Background technology
Seeds company typically first carries out fringe and selects work, then carry out grain-by-grain seed selection finishing during corn seed producing.In fringe
When selecting, the good normal corn ear of phenotypic characteristic is picked out, recycling band bract corn ear is gone again at bract
Reason, discard damage by worms, go mouldy, the corn ear of kernel abortion and xenogenesis, for improve seed quality and purity, increase corn
Relationship with Yield is close.The fruit ears such as the fruit ear of damaging by worms for including in other corn ears, the fruit ear that goes mouldy, xenogenesis fruit ear, kernel abortion, body
On present phenotypic characteristic, shape, size, grain shape, grain color, cob color, texture of corn ear etc. and normal corn ear
Feature difference is larger.And the bract of corn ear is obviously many compared with corn ear in color and shape feature.Conventional method master
Band bract corn ear and other corn ears are filtered out by manpower, it is slow that there are breakneck accelerations, expends a large amount of manpowers and wealth
The problem of power.
Currently with computer vision technique, characteristics of image is extracted, is identified and corn ear of classifying, instead of biography
The hand picking of system carries out fine digital assay and processing, although avoiding with image processing techniques to fruit ear image
The shortcomings that consuming of manpower and materials, but this method needs artificial progress feature extraction and design, calculates complicated and domestic practical
Mostly in outdoor progress, production of hybrid seeds site environment is complicated and changeable for production of hybrid seeds processing factory, captured image recognition poor robustness, it is difficult to push away
To practical production of hybrid seeds assembly line.
Invention content
It is refreshing based on convolution that the present invention provides a kind of one kind for overcoming the above problem or solving the above problems at least partly
Corn seed producing fruit ear image preliminary classification method through network.
According to an aspect of the present invention, fraction at the beginning of a kind of corn seed producing fruit ear image based on convolutional neural networks is provided
Class method, including：
S1, corn ear original training sample collection and corn ear test sample collection, the original instruction of corn ear are obtained
It includes the Two-dimensional Color Image of normal corn ear, with bract corn fruit to practice sample set and the corn ear test sample collection
The Two-dimensional Color Image of fringe and the Two-dimensional Color Image of other corn ears；
S2, by the first convolutional neural networks of AlexNet, differentiate that the corn ear test sample concentrates each test specimens
The affiliated type of this affiliated type, each test sample is normal corn ear, band bract corn ear or other corn fruits
Fringe；
S3, according to the reality for differentiating the obtained affiliated type of each corn ear test sample and each test sample
Affiliated type determines the accuracy rate for differentiating result；
If S4, the accuracy rate for differentiating result are except preset range, right according to amplification corn ear training sample set
The first convolutional neural networks of the AlexNet optimize training, obtain the second convolutional neural networks, utilize second convolution
Neural network differentiates that the corn ear test sample concentrates the affiliated type of each test sample, the amplification corn fruit again
Fringe training sample set concentrates the quantity of original training sample to obtain by increasing the corn ear original training sample.
Preferably, further include before step S4：The original training sample that the corn ear original training sample is concentrated
Translated, rotated respectively, mirror image, change original training sample brightness and fuzzy original training sample, obtain the amplification
Corn ear training sample set.
Preferably, further include before step S4：It is masked according to the probability value of setting by Dropout methods described
The partial nerve member of the first convolutional neural networks of AlexNet.
Preferably, step S4 further comprises：According to the amplification corn ear training sample set, by under stochastic gradient
Drop algorithm optimizes training to the first convolutional neural networks of the AlexNet, obtains second convolutional neural networks.
Preferably, further include before step S2：According to the original corn ear training sample set, to initial convolutional Neural
Network carries out pre-training, obtains the first convolutional neural networks of the AlexNet.
Preferably, the initial convolutional neural networks include 5 convolutional layers, 3 maximum pond layers and 3 full articulamentums,
For 5 convolutional layers, the size of convolution kernel is respectively 11*11,5*5,3*3,3*3,3*3, convolution step-length is respectively 4,1,
1、1、1。
Preferably, the learning rate for adjusting the initial convolutional neural networks into Mobile state by Adam algorithms, to obtain
State the first convolutional neural networks.
Preferably, step S1 is specifically included：
S11, all corn ears of extraction Two-dimensional Color Image in area-of-interest；
S12, the area-of-interest is pre-processed, pretreated Two-dimensional Color Image is obtained, after the pretreatment
Two-dimensional Color Image include pretreated area-of-interest；
S13, the Two-dimensional Color Image that the pretreated Two-dimensional Color Image is divided into normal corn ear, band bract
The Two-dimensional Color Image of corn ear and the Two-dimensional Color Image of other corn ears, and according to 4:1 preset ratio respectively with
Machine matches normal corn ear image, the corn ear image with bract and other corn ear images, to obtain the corn
Fruit ear original training sample collection and the corn ear test sample collection.
Preferably, it is specifically included in step S11：
S111, collected two-dimensional color corn ear image is transformed into hsv color space by RGB color；
S112, medium filtering denoising is carried out for tone channel；
S113, using the corn ear image feature different from background color, for the tone channel after medium filtering
Image carries out binary conversion treatment；
S114, morphology opening operation processing is carried out to bianry image, removes broken fritter；
S115, setting area threshold, exclude the smaller connected region detected；
S116, the bulk region detected retained are corn ear region.
Preferably, it is specifically included in step S12：By sample-by-sample mean value abatement to the areas ROI of the Two-dimensional Color Image
Domain is pre-processed, and the Two-dimensional Color Image is made to normalize；
The present invention proposes a kind of corn seed producing fruit ear image preliminary classification method based on convolutional neural networks, at the beginning of fruit ear
Grade fringe selects the stage that can identify that normal corn ear, the method with bract corn ear and other corn ears, this method pass through
Fruit ear Image Acquisition frame is built using holder by side on a moving belt, and carrying out region of interesting extraction to the image collected goes forward side by side
Row characteristics of image normalized, classification marker corn ear image pattern constitute sample set.The present invention utilizes AlexNet convolution
Neural network model is finely adjusted training to corn ear training sample set, and the adaptive of learning rate is completed using Adam methods.
After the test of corn ear test sample collection, to improve Network Recognition precision, by expanding corn ear training sample,
The partial nerve member that suitable Dropout parameters mask convolutional neural networks is set, network over-fitting is prevented, using boarding steps
Degree optimization training algorithm completes the optimization of network.This corn ear image of independently being chosen layer by layer from hidden layer is by color, side
Equal low layers to angle point, the mode of shape contour level feature, avoid manually extract corn ear characteristics of image it is cumbersome with it is unilateral,
Make network that there is the autonomous ability for choosing corn ear characteristics of image and being learnt and being identified.This method meets during the production of hybrid seeds
Fringe choosing accuracy requirement, reduce the consuming of manpower and financial resources, precision is high, for fruit ear fringe choosing method provide it is a kind of it is new on the way
Diameter substantially increases the efficiency of the corn ear production of hybrid seeds.
Description of the drawings
Fig. 1 is a kind of corn seed producing fruit ear image preliminary classification method based on convolutional neural networks of the embodiment of the present invention
Flow chart；
Fig. 2 is a kind of corn seed producing fruit ear image preliminary classification method based on convolutional neural networks of the embodiment of the present invention
Convolutional neural networks fundamental diagram.
Specific implementation mode
With reference to the accompanying drawings and examples, the specific implementation mode of the present invention is described in further detail.Implement below
Example is not limited to the scope of the present invention for illustrating the present invention.
Fig. 1 is a kind of corn seed producing fruit ear image preliminary classification method based on convolutional neural networks of the embodiment of the present invention
Flow chart, as shown in Figure 1, this method includes：
S1, corn ear original training sample collection and corn ear test sample collection, the original instruction of corn ear are obtained
It includes the Two-dimensional Color Image of normal corn ear, with bract corn fruit to practice sample set and the corn ear test sample collection
The Two-dimensional Color Image of fringe and the Two-dimensional Color Image of other corn ears；
S2, by the first convolutional neural networks of AlexNet, differentiate that the corn ear test sample concentrates each test specimens
The affiliated type of this affiliated type, each test sample is normal corn ear, band bract corn ear or other corn fruits
Fringe；
S3, according to the reality for differentiating the obtained affiliated type of each corn ear test sample and each test sample
Affiliated type determines the accuracy rate for differentiating result；
If S4, the accuracy rate for differentiating result are except preset range, right according to amplification corn ear training sample set
The first convolutional neural networks of the AlexNet optimize training, obtain the second convolutional neural networks, utilize second convolution
Neural network differentiates that the corn ear test sample concentrates the affiliated type of each test sample, the amplification corn fruit again
Fringe training sample set concentrates the quantity of original training sample to obtain by increasing the corn ear original training sample.
First according to the Two-dimensional Color Image of fruit ear, fruit ear here includes normal corn ear, band bract corn ear
It is color to choose a certain number of two dimensions respectively from sample using the Two-dimensional Color Image of fruit ear as sample with other corn ears
Color image respectively constitutes corn ear original training sample collection and test sample collection.Corn ear original training sample is concentrated
Input of the Two-dimensional Color Image as initial convolutional neural networks, pre-training is carried out to initial convolutional neural networks, is fitted
Together in the first convolutional neural networks of AlexNet of detection corn seed producing fruit ear.Using the first convolutional neural networks of AlexNet to jade
The Two-dimensional Color Image that rice cracker fringe test sample is concentrated, differentiates that the corn ear test sample concentrates the institute of each test sample
Belong to type, the affiliated type of each test sample is normal corn ear, band bract corn ear and other corn ears；Due to
Certainly there is error in calculating process, the Two-dimensional Color Image with bract fruit ear may be divided into normal corn ear class, or
The Two-dimensional Color Image of normal corn ear is divided into other corn ear classes by person, therefore, it is desirable to statistical classification result it is accurate
Rate determines and sentences according to type belonging to the obtained affiliated type of each test sample and the reality of each test sample is differentiated
The accuracy rate of other result.
If differentiating the accuracy rate of result within preset range, illustrate to train AlexNet the first convolution nerve nets come
Network is relatively adapted to detect for normal corn ear, band bract corn ear and other corn ears in this environment, utilizes AlexNet
The test sample that first convolutional neural networks concentrate the corn ear test sample is classified, and normal corn ear is obtained
Class, band bract corn ear class and other corn ear classes.
If the accuracy rate of classification results except preset range, illustrates to train AlexNet the first convolution nerve nets come
Network does not reach the required precision of project needs.The corn ear original training sample used when neural network initial due to training
Overall number is fewer and the sample size of other corn ears with bract corn ear is considerably less than Mormal ear sample
Quantity, there are lacks of uniformity, trained result can also certain guidance quality between sample, in order to make finally obtained convolution
Neural network is most suitably adapted for the classification of the corn seed producing fruit ear image under detection actual conditions, increases original training sample
Each Two-dimensional Color Image that original training sample is concentrated is done various deformation by number, such as to each two-dimensional color corn fruit
Fringe image translated, rotation, mirror image, the brightness for changing corn ear image and blurred picture, each Two-dimensional Color Image
Corn ear original image, translated after image, carry out rotation process after image, carry out mirror image after image, change
Become the image after original corn ear brightness of image and carries out the image construction expansion after fuzzy operation to original corn ear image
Increase corn ear training sample set, to expand the image of corn ear training sample concentration as the first convolutional Neurals of AlexNet
The input of network is trained the first convolutional neural networks of AlexNet, obtains the second convolutional neural networks, utilize volume Two
The test sample that product neural network concentrates test sample is classified, and normal corn ear class, band bract corn ear are obtained
Class and other corn ear classes.
The present invention provides a kind of corn seed producing fruit ear image preliminary classification method based on convolutional neural networks, autonomous to choose
Corn ear image avoids artificial extraction corn ear by the low layers such as side, color to angle point, the mode of shape contour level feature
Characteristics of image it is cumbersome with it is unilateral, so that convolutional neural networks is had and autonomous choose corn ear characteristics of image and learnt and known
Other ability provides a method the automatically screening of corn seed producing fruit ear primary fringe choosing.
On the basis of the above embodiments, it is preferable that further include before step S4：Trained sample original to the corn ear
The original training sample of this concentration translated, rotated respectively, mirror image, the brightness for changing original training sample and obscures original instruction
Practice sample, obtains the amplification corn ear training sample set.
On the basis of the above embodiments, it is preferable that further include before step S4：By Dropout methods according to setting
Shielding probability mask the partial nerve members of the first convolutional neural networks of the AlexNet at random, the shielding probability of this paper sets
It is set to 0.5.
On the basis of the above embodiments, specifically, further include before step S4：It is trained according to the amplification corn ear
Sample set optimizes training, described in acquisition by stochastic gradient descent algorithm to the first convolutional neural networks of the AlexNet
Second convolutional neural networks.
Fig. 2 is a kind of corn seed producing fruit ear image preliminary classification method based on convolutional neural networks of the embodiment of the present invention
Convolutional neural networks fundamental diagram, as shown in Fig. 2, after obtaining amplification corn ear training sample set, amplification corn ear
Input of the image that training sample is concentrated as the second convolutional neural networks, is trained the second convolutional neural networks, uses
Dropout methods mask the partial nerve of the first convolutional neural networks of the AlexNet according to the shielding probability of setting at random
Member.Dropout methods refer to, for certain neurons, being incited somebody to action according to certain probability in the training process of convolutional neural networks
It is temporarily abandoned from convolutional neural networks, i.e. the weights of these neurons are set as 0.Each neuron of output layer is according to general
As a result, the weights of the neuron masked are set as 0, formula is for rate p outputs：
R=m.a (Wv),
Wherein v is the column vector of n*1 dimensions, and W is the matrix of d*n, and m is 01 column vector of a d*1, and a (x) is one and meets a
(0)=0 excitation function form.N is the number of neuron, and W is the connection weight between network, and d is weights number, and v is defeated
Enter neuron, r is output.
Then training is optimized to the first convolutional neural networks of the AlexNet by stochastic gradient descent algorithm, obtained
Second convolutional neural networks are taken, when due to being trained to the second convolutional neural networks, training sample is enough and balanced
, the second obtained convolutional neural networks are more suitable for classifying to corn ear.
It should be noted that stochastic gradient descent algorithm, in given sample set M, the random copy N that takes out replaces original
Beginning, sample M was used as complete or collected works, was trained to model.This training is due to being extraction section data, so there is larger probability
It obtains, a locally optimal solution.But an apparent benefit is, if in sampling OK range, both to ask
Go out as a result, and speed it is also fast.
The present invention carries out various deformation on the basis of first time training, to original corn ear training sample, expands beautiful
The number of rice cracker fringe training sample makes the convolutional neural networks tested corn ear test sample collection have better Shandong
Stick, the partial nerve member masked at random according to shielding probability using Dropout methods, effectively prevents convolutional neural networks
Over-fitting is optimized using the second convolutional neural networks of stochastic gradient descent algorithm pair, accelerates convolutional neural networks to jade
The speed that rice cracker fringe test sample is classified.
On the basis of the above embodiments, specifically, the initial convolutional neural networks are AlexNet convolutional neural networks
Model, the AlexNet convolutional neural networks model include 5 convolutional layers, 3 maximum pond layers and 3 full articulamentums.
AlexNet convolutional neural networks models are a kind of common convolutional neural networks models, first provide the one of AlexNet
A little parameters and structure chart, 5 layers of convolutional layer, 3 layers of full articulamentum, 8 layers, number of parameters 60M, neuron number 650k of depth, classification
1000 class of number, it is output softmax that the last one full articulamentum neuron number, which is 1000, in AlexNet convolutional neural networks
Classification number, the embodiment of the present invention by fruit ear be divided into normal corn ear, band bract corn ear and other corn ear three classes,
Therefore output number need to be changed to the classification number 3 of object set.
On the basis of the above embodiments, specifically, for 5 convolutional layers, the size of convolution kernel is respectively 11*
11,5*5,3*3,3*3,3*3, convolution step-length are respectively 4,1,1,1,1.The number of 5 convolutional layers output characteristic images is respectively
96、256、384、384、256。
Initial convolutional neural networks are AlexNet volumes containing 5 convolutional layers, 3 maximum pond layers and 3 full articulamentums
Product neural network model, the original training sample concentrated using original corn ear training sample and original corn ear train sample
Input of the sample label as initial convolutional neural networks in this, and it is converted into the specified number of initial convolutional neural networks
According to type, and using the input data as the input of first convolutional layer of AlexNet, the size of the convolution kernel of first convolutional layer
For 11*11, convolution step-length is 4 pixels, realizes the function to input data multilayer convolution, and sample to the output after convolution
(pond), normalization, the input as next layer of convolutional layer.Similarly until the 5th convolutional layer.In initial convolutional neural networks most
What neuron exported in the full articulamentum of the latter is fruit ear classification number, if fruit ear to be only divided into normal corn ear, band bract
Output number need to be changed to the classification number 3 of corn ear by corn ear and 3 classifications of other corn ears.
On the basis of the above embodiments, specifically, the initial convolutional Neural is adjusted into Mobile state by Adam algorithms
The learning rate of network, to obtain the first convolutional neural networks of AlexNet after training.
It should be noted that Adam algorithms refer to adaptive moments estimation.Square is meant that in probability theory：If one random
Variable X obeys some distribution, and the first moment of X is E (X), that is, sample mean, and the second moment of X is exactly E (X^2), that is,
The average value of sample square.Adam algorithms are according to loss function to the single order moments estimation and second order moments estimation of the gradient of each parameter
Dynamic adjustment is directed to the learning rate of each parameter.The method that Adam algorithms are also based on gradient decline, but each iteration
The Learning Step of parameter all there are one the range determined, will not lead to prodigious Learning Step because of prodigious gradient, parameter
It is worth more stable.
On the basis of the above embodiments, it is preferable that step S1 is specifically included：S11, the Two-dimensional Color Image for extracting fruit ear
In area-of-interest；S12, the area-of-interest is pre-processed, obtains pretreated Two-dimensional Color Image, it is described pre-
Corn ear Two-dimensional Color Image that treated includes pretreated area-of-interest；S13, by the pretreated corn
Fruit ear Two-dimensional Color Image is divided into the Two-dimensional Color Image of normal corn ear and the Two-dimensional Color Image with bract corn ear,
And according to 4:1 preset ratio matches normal corn ear image, band bract corn ear image and other corn ears respectively
Image, to obtain the corn ear original training sample collection and the corn ear test sample collection.
It should be noted that area-of-interest (regionofinterest, abbreviation ROI), machine vision, image procossing
In, region to be treated is sketched the contours of in a manner of box, circle, ellipse, irregular polygon etc. from processed image, is referred to as felt
Interest region.In image processing field, ROI is the image-region selected from image, this region is your image point
Analyse emphasis of interest.The region is drawn a circle to approve to be further processed.Want the target read using ROI delineations, it is possible to reduce place
The time is managed, precision is increased.
After the Two-dimensional Color Image for extracting fruit ear, ROI region is extracted to each Two-dimensional Color Image, and color to each two dimension
The ROI region of color image is pre-processed, since sample at this time is fewer, the Two-dimensional Color Image by fruit ear that can be artificial
Label is divided into the Two-dimensional Color Image, the Two-dimensional Color Image with bract corn ear and other corn ears of normal corn ear
Two-dimensional Color Image, it is color to randomly select the Two-dimensional Color Image of normal corn ear, the two dimension with bract corn ear respectively
The two dimension match color image of color image and other corn ears forms corn ear original training sample collection according to 80% ratio, remains
Remaining 20% normal corn ear Two-dimensional Color Image, the Two-dimensional Color Image with bract corn ear and other corn ears
And be coloured image composition test sample collection.
On the basis of the above embodiments, specifically, it is specifically included in step S11：S111, by collected two-dimensional color
Corn ear image is transformed into hsv color space by RGB color；S111 carries out medium filtering for H (tone) channel and goes
It makes an uproar processing；S113 is using the corn ear image feature different from background color, for H (tone) channel after medium filtering
Image carries out OTSU binaryzations；S114 carries out morphology opening operation processing, the more broken fritter of removal to bianry image；
S115 sets area threshold, excludes the smaller connected region detected, and the area threshold being arranged herein is 10000；S116 is stayed
It is corn ear region to leave the bulk region detected come,.
On the basis of the above embodiments, specifically, it is specifically included in step S12：S121 is cut down by sample-by-sample mean value
The ROI region of the corn ear Two-dimensional Color Image is pre-processed, the corn ear Two-dimensional Color Image normalizing is made
Change.
It should be noted that sample-by-sample mean value abatement is one of method of data normalization, if the data of processing are flat
Steady (i.e. the statistics of each dimension of data obeys same distribution), it may be considered that subtract the statistics of data on each sample
Average value (sample-by-sample calculating).For example, for image, this normalization can remove the average brightness value of image
(intensity).To the illumination of image and lose interest in many cases, and increasingly focus on its content, at this moment to every number
The mean value that strong point removes pixel is meaningful.
Specifically, sample-by-sample mean value cut down, be exactly the assembly average for subtracting these data on each sample, by by
Sample average abatement pre-processes the ROI region of the corn ear Two-dimensional Color Image, makes the Two-dimensional Color Image
Normalization.With P={ p(1),p(2),…,p(n)Come indicate extraction ROI image region, p(i)For the pixel in ROI image region
Value, X (i) are the sample after sample-by-sample mean value abatement, and mean (p (i)) is to average, and var (p (i)) is to seek variance, specific public
Formula is：
Wherein ε is a correction amount, and it is zero that can prevent denominator, also there is certain effect for inhibiting noise.
In order to test the nicety of grading whether the first convolutional neural networks reach requirement, AlexNet the first convolution god is used
The test sample concentrated to corn ear test sample through network is classified, and obtains the first convolutional neural networks of AlexNet
Nicety of grading.
On the basis of the above embodiments, further include before step S1：S0 arranges the Two-dimensional Color Image acquisition ring of fruit ear
Border acquires the Two-dimensional Color Image of fruit ear.Specifically, arrange that the Two-dimensional Color Image acquisition environment of fruit ear is specially：First, exist
The hardware device of Image Acquisition is built above conveyer belt, includes erection and the light source of industrial CCD camera (resolution ratio is 2,000,000)
Arrangement, chooses a small amount of corn ear and marker (blank sheet of paper)；Then, the camera determined under imaging system static environment is tested
Highly, so that camera shooting breadth is covered whole fringe fruit ear, adjust the white balance of camera；Secondly, it tests and determines quantity of light source, light
Source height, angle and light source light filling intensity etc. determine stable shooting light for fruit ear Image Acquisition；Finally, dynamic ring is determined
The speed of the picking rate of camera and conveyer belt under border, adjustment time for exposure, focal length and aperture size avoid image streaking existing
As.
The present invention provides a kind of method for selecting the stage that can accurately identify corn seed producing fruit ear in corn ear primary fringe,
This method builds fruit ear Image Acquisition frame by side on a moving belt using holder, is carried out to collected corn ear image
ROI extracts and carries out characteristics of image normalized, and classification marker corn ear sample constitutes sample set.The present invention utilizes
AlexNet convolutional neural networks models are finely adjusted training to training sample set, and the adaptive of learning rate is completed using Adam methods
It answers.After the test by test sample collection, to improve Network Recognition precision, by expanding corn ear training sample, setting
Suitable Dropout parameters mask the partial nerve member of convolutional neural networks, prevent network over-fitting, excellent using stochastic gradient
Change the optimization that training algorithm completes network.It is this from hidden layer layer by layer independently choose image by the low layers such as side, color to angle point,
The mode of shape contour level feature, avoid artificial extraction corn ear characteristics of image it is cumbersome with it is unilateral, make network that there is oneself
The main ability for choosing corn ear characteristics of image and being learnt and being identified.This method meets the accurate of the choosing of the fringe during the production of hybrid seeds
Property require, reduce the consuming of manpower and financial resources, precision is high, provides a kind of new way for fruit ear fringe choosing method, greatly improves
The efficiency of the fruit ear production of hybrid seeds.
Finally, method of the invention is only preferable embodiment, is not intended to limit the scope of the present invention.It is all
Within the spirit and principles in the present invention, any modification, equivalent replacement, improvement and so on should be included in the protection of the present invention
Within the scope of.
Claims (10)
1. a kind of corn seed producing fruit ear image preliminary classification method based on convolutional neural networks, which is characterized in that including：
S1, corn ear original training sample collection and corn ear test sample collection, the original trained sample of corn ear are obtained
This collection and the corn ear test sample collection include the Two-dimensional Color Image of normal corn ear, with bract corn ear
The Two-dimensional Color Image of Two-dimensional Color Image and other corn ears；
S2, by the first convolutional neural networks of AlexNet, differentiate that the corn ear test sample concentrates each test sample
The affiliated type of affiliated type, each test sample is normal corn ear, band bract corn ear or other corn ears；
S3, according to differentiating belonging to the obtained affiliated type of each corn ear test sample and the reality of each test sample
Type determines the accuracy rate for differentiating result；
If S4, the accuracy rate for differentiating result are except preset range, according to amplification corn ear training sample set, to described
The first convolutional neural networks of AlexNet optimize training, obtain the second convolutional neural networks, utilize second convolutional Neural
Network differentiates that the corn ear test sample concentrates the affiliated type of each test sample, the amplification corn ear instruction again
Practice sample set and concentrates the quantity of original training sample to obtain by increasing the corn ear original training sample.
2. method according to claim 1, which is characterized in that further include before step S4：
The original training sample that the corn ear original training sample is concentrated is translated respectively, is rotated, mirror image, changes original
The brightness of beginning training sample and fuzzy original training sample obtain the amplification corn ear training sample set.
3. method according to claim 1, which is characterized in that further include before step S4：
The part god of the first convolutional neural networks of the AlexNet is masked according to the probability value of setting by Dropout methods
Through member.
4. method according to claim 1, which is characterized in that step S4 further comprises：
According to the amplification corn ear training sample set, by stochastic gradient descent algorithm to the first convolution of the AlexNet
Neural network optimizes training, obtains second convolutional neural networks.
5. method according to claim 1, which is characterized in that further include before step S2：
According to the original corn ear training sample set, pre-training carried out to initial convolutional neural networks, described in acquisition
The first convolutional neural networks of AlexNet.
6. method according to claim 5, which is characterized in that the initial convolutional neural networks include 5 convolutional layers, 3
Maximum pond layer and 3 full articulamentums, for 5 convolutional layers, the size of convolution kernel is respectively 11*11,5*5,3*3,3*
3,3*3, convolution step-length are respectively 4,1,1,1,1.
7. method according to claim 5, which is characterized in that adjust the initial convolution god into Mobile state by Adam algorithms
Learning rate through network, to obtain the first convolutional neural networks of the AlexNet.
8. method according to claim 1, which is characterized in that step S1 is specifically included：
S11, all corn ears of extraction Two-dimensional Color Image in area-of-interest；
S12, the area-of-interest is pre-processed, obtains pretreated Two-dimensional Color Image, described pretreated two
It includes pretreated area-of-interest to tie up coloured image；
S13, the Two-dimensional Color Image that the pretreated Two-dimensional Color Image is divided into normal corn ear, band bract corn
The Two-dimensional Color Image of fruit ear and the Two-dimensional Color Image of other corn ears, and according to 4:1 preset ratio is matched at random respectively
Than normal corn ear image, the corn ear image with bract and other corn ear images, to obtain the corn ear
Original training sample collection and the corn ear test sample collection.
9. method according to claim 8, which is characterized in that specifically included in step S11：
S111, collected two-dimensional color corn ear image is transformed into hsv color space by RGB color；
S112, medium filtering denoising is carried out for tone channel；
S113, using the corn ear image feature different from background color, for the image in the tone channel after medium filtering
Carry out binary conversion treatment；
S114, morphology opening operation processing is carried out to bianry image, removes broken fritter；
S115, setting area threshold, exclude the smaller connected region detected；
S116, the bulk region detected retained are corn ear region.
10. method according to claim 8, which is characterized in that specifically included in step S12：
The ROI region of the Two-dimensional Color Image is pre-processed by sample-by-sample mean value abatement, makes the two-dimensional color figure
As normalization.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810344463.8A CN108549910A (en) | 2018-04-17 | 2018-04-17 | A kind of corn seed producing fruit ear image preliminary classification method based on convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810344463.8A CN108549910A (en) | 2018-04-17 | 2018-04-17 | A kind of corn seed producing fruit ear image preliminary classification method based on convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108549910A true CN108549910A (en) | 2018-09-18 |
Family
ID=63515468
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810344463.8A Pending CN108549910A (en) | 2018-04-17 | 2018-04-17 | A kind of corn seed producing fruit ear image preliminary classification method based on convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108549910A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109635795A (en) * | 2018-10-22 | 2019-04-16 | 小黄狗环保科技有限公司 | A kind of Intelligent supplemental lighting method improving Bottle & Can discrimination based on VGG16 network model |
| CN110267024A (en) * | 2019-07-26 | 2019-09-20 | 惠州视维新技术有限公司 | Adjust the method, apparatus and computer readable storage medium of TV white balance value |
| CN110766053A (en) * | 2019-09-23 | 2020-02-07 | 中国农业大学 | Corn seed production ear screening method and device based on double-path convolutional neural network |
| CN110991511A (en) * | 2019-11-26 | 2020-04-10 | 中原工学院 | Sunflower crop seed sorting method based on deep convolutional neural network |
| CN112102257A (en) * | 2020-08-26 | 2020-12-18 | 电子科技大学 | Automatic human body excrement identification method based on convolutional neural network |
| CN112990198A (en) * | 2021-03-22 | 2021-06-18 | 华南理工大学 | Detection and identification method and system for water meter reading and storage medium |
| CN114724005A (en) * | 2022-03-25 | 2022-07-08 | 清华大学 | Solid brewing granular raw material doping identification method based on deep neural network |
| CN115713763A (en) * | 2022-11-25 | 2023-02-24 | 青海卓旺智慧信息科技有限公司 | Potato image recognition system based on deep learning |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104517122A (en) * | 2014-12-12 | 2015-04-15 | 浙江大学 | Image target recognition method based on optimized convolution architecture |
| CN106485251A (en) * | 2016-10-08 | 2017-03-08 | 天津工业大学 | Egg embryo classification based on deep learning |
| CN106875361A (en) * | 2017-02-17 | 2017-06-20 | 深圳市唯特视科技有限公司 | A kind of method that poisson noise is removed based on depth convolutional neural networks |
| CN107862300A (en) * | 2017-11-29 | 2018-03-30 | 东华大学 | A kind of descending humanized recognition methods of monitoring scene based on convolutional neural networks |
-
2018
- 2018-04-17 CN CN201810344463.8A patent/CN108549910A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104517122A (en) * | 2014-12-12 | 2015-04-15 | 浙江大学 | Image target recognition method based on optimized convolution architecture |
| CN106485251A (en) * | 2016-10-08 | 2017-03-08 | 天津工业大学 | Egg embryo classification based on deep learning |
| CN106875361A (en) * | 2017-02-17 | 2017-06-20 | 深圳市唯特视科技有限公司 | A kind of method that poisson noise is removed based on depth convolutional neural networks |
| CN107862300A (en) * | 2017-11-29 | 2018-03-30 | 东华大学 | A kind of descending humanized recognition methods of monitoring scene based on convolutional neural networks |
Non-Patent Citations (1)
| Title |
|---|
| 党宇等: ""基于深度学习AlexNet的遥感影像地表覆盖分类评价研究"", 《地球信息科学》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109635795A (en) * | 2018-10-22 | 2019-04-16 | 小黄狗环保科技有限公司 | A kind of Intelligent supplemental lighting method improving Bottle & Can discrimination based on VGG16 network model |
| CN110267024A (en) * | 2019-07-26 | 2019-09-20 | 惠州视维新技术有限公司 | Adjust the method, apparatus and computer readable storage medium of TV white balance value |
| CN110766053A (en) * | 2019-09-23 | 2020-02-07 | 中国农业大学 | Corn seed production ear screening method and device based on double-path convolutional neural network |
| CN110991511A (en) * | 2019-11-26 | 2020-04-10 | 中原工学院 | Sunflower crop seed sorting method based on deep convolutional neural network |
| CN112102257A (en) * | 2020-08-26 | 2020-12-18 | 电子科技大学 | Automatic human body excrement identification method based on convolutional neural network |
| CN112990198A (en) * | 2021-03-22 | 2021-06-18 | 华南理工大学 | Detection and identification method and system for water meter reading and storage medium |
| CN114724005A (en) * | 2022-03-25 | 2022-07-08 | 清华大学 | Solid brewing granular raw material doping identification method based on deep neural network |
| CN115713763A (en) * | 2022-11-25 | 2023-02-24 | 青海卓旺智慧信息科技有限公司 | Potato image recognition system based on deep learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108549910A (en) | A kind of corn seed producing fruit ear image preliminary classification method based on convolutional neural networks | |
| CN109754017B (en) | Hyperspectral image classification method based on separable three-dimensional residual error network and transfer learning | |
| Liu et al. | Identification of rice seed varieties using neural network | |
| CN106485251B (en) | Egg embryo classification based on deep learning | |
| Xiaoxiao et al. | Image recognition of tea leaf diseases based on convolutional neural network | |
| CN109584251A (en) | A kind of tongue body image partition method based on single goal region segmentation | |
| CN108734138B (en) | Melanoma skin disease image classification method based on ensemble learning | |
| CN109241982A (en) | Object detection method based on depth layer convolutional neural networks | |
| CN109961049A (en) | Cigarette brand recognition methods under a kind of complex scene | |
| CN110276386A (en) | A kind of apple grading method and system based on machine vision | |
| CN106845450A (en) | Dark surrounds face identification method based near infrared imaging Yu deep learning | |
| CN101059425A (en) | Method and device for identifying different variety green tea based on multiple spectrum image texture analysis | |
| CN106980852A (en) | Based on Corner Detection and the medicine identifying system matched and its recognition methods | |
| CN110956212A (en) | Threshing quality detection method based on visual feature fusion | |
| CN110046574A (en) | Safety cap based on deep learning wears recognition methods and equipment | |
| Shajahan et al. | Identification and counting of soybean aphids from digital images using shape classification | |
| CN106874929A (en) | A kind of pearl sorting technique based on deep learning | |
| CN109034184A (en) | A kind of grading ring detection recognition method based on deep learning | |
| CN109101926A (en) | Aerial target detection method based on convolutional neural networks | |
| CN107463954A (en) | A kind of template matches recognition methods for obscuring different spectrogram picture | |
| CN105678245A (en) | Target position identification method based on Haar features | |
| CN108416814A (en) | Quick positioning and recognition methods and the system on a kind of pineapple head | |
| CN108629762A (en) | A kind of stone age evaluation and test model reduces the image pre-processing method and system of interference characteristic | |
| CN111652817B (en) | Underwater image sharpening method based on human eye visual perception mechanism | |
| Niu et al. | Automatic localization of optic disc based on deep learning in fundus images |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180918 |