CN109558787A - A kind of Bamboo insect pests recognition methods based on convolutional neural networks model - Google Patents
A kind of Bamboo insect pests recognition methods based on convolutional neural networks model Download PDFInfo
- Publication number
- CN109558787A CN109558787A CN201811138103.9A CN201811138103A CN109558787A CN 109558787 A CN109558787 A CN 109558787A CN 201811138103 A CN201811138103 A CN 201811138103A CN 109558787 A CN109558787 A CN 109558787A
- Authority
- CN
- China
- Prior art keywords
- neural networks
- convolutional neural
- insect pests
- networks model
- model
- 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
- G06V20/10—Terrestrial scenes
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Catching Or Destruction (AREA)
Abstract
The invention discloses a kind of Bamboo insect pests recognition methods based on convolutional neural networks model, belong to agriculture and forestry injurious insect identification and sorting technique field, method includes the following steps: acquisition Bamboo insect pests sample image data;Bamboo insect pests sample image data is handled, the Bamboo insect pests sample image data collection that obtains that treated;It is optimized using VGG convolutional neural networks model, and to VGG convolutional neural networks model;Target pest identification classification is carried out using the VGG convolutional neural networks model after optimization;Based on VGGNet network structure develop it is a set of suitable for Bamboo insect pests method for identifying and classifying, it can be achieved that the quick identification and classification of common Bamboo insect pests are, it can be achieved that automatic identification and classification, recognition accuracy height.
Description
Technical field
The invention belongs to agriculture and forestry injurious insect identification and sorting technique fields, are specifically related to a kind of based on convolutional neural networks mould
The Bamboo insect pests recognition methods of type.
Background technique
China is a large agricultural country, can all encounter invading for variety classes pest every year in the cultivating process of crops
Evil, so that the decline that crop is different degrees of in yield and quality, when disaster is serious, results even in the big face of crops
Product total crop failure.Accurately and effectively to insect carry out taxonomic identification and identification be can be unfolded in time pest control, avoid it is huge
Crops economic loss an important prerequisite.Insect is the animal of most species in natural environment, and form is changeable, line
The abundant identification of reason gets up to have very big difficulty.Traditional classification and identification of insects works mainly by insect expert or classification of insect people
Member carries out according to professional knowledge and research experience or reference literature data identification identification, knowledge even with profession and
Experience abundant is also difficult the case where avoiding type from obscuring, and therefore, develops a kind of quickly and effectively classification for pest
Identification system will be helpful to the prevention and treatment of crop pest, to promote agricultural development, reduce economic loss.
China is the most abundant country of bamboo resource in the world.With further going deep into for intensive management, Bamboo insect pests
It gets worse, threatens the yield and quality of bamboo shoots bamboo wood.In numerous pests, is most contained, made with atrachea vulgaris harm
At the largest loss, in woodss such as the propinquity bamboo in the provinces such as Zhejiang, Guangdong, red shell bamboo, tea stalk bamboos (Pseudosasa amablis)
Interior, the hazard ratio of atrachea vulgaris is up to 60% or more, notably up to 100%.China's atrachea vulgaris type is more, structure is complicated, Ji Zhufan
Enclose that wide, harm is hidden, difficulty of prevention and cure is big, and to rely primarily on classification of insect personnel special according to form for traditional taxonomic identification work
Sign is identified that the identification mark of moth class insect mainly has wing, including vein, wing linkage mode, finned and finned surface speckle,
Feature used in it includes feeler, mouthpart (mainly must and beak), foot etc..Moth class classification of insect is strongly professional, available data
Based on mostly being described with character property, conventional method classification is mainly that expert with eyes observes its structure feature and color characteristic etc., is needed
There are sturdy professional knowledge and experience abundant, and type is obscured and happens occasionally[2].Therefore, in time for particular kind of
Preventing and controlling are unfolded in pest, avoid huge economy of forestry loss, rapidly and accurately judge that crop pests type seems
It is extremely important.
With the rise of neural network filter, the automatic identification technology based on image has become ongoing research area
Hot spot, the research of insect image identification identification such as carry out feature extraction to the shape of insect image identification and color to obtain greater advance
Afterwards, it completes to classify using radial base neural net, accuracy rate is up to 96%;BP neural network model is established to identify Ponkan disease pest
Evil, Average Accuracy is up to 92.67%;The grain storage pest image-recognizing method based on depth convolutional neural networks is introduced, is identified
Up to 97.61% in terms of accuracy.But the above method still has defect, and such as data sample amount is insufficient, and data prediction is multiple
Miscellaneous, feature extraction is insufficient, and model-fitting degree fluctuation is larger.In addition, for it is some exclusively for data set (include 1000 classifications,
More than 1000000 width images) propose deep layer network model, with regard to the particularity of pest and complexity come to model progress
It optimizes and revises.
VGG convolutional neural networks model is developed by Oxford University's visual geometric group (Visual Geometry Group)
The convolutional neural networks structure of development, it was in ILSVRC (ImageNet Large Scale Visual in 2014
Recognition Challenge) achievement of second place is obtained in contest.The network structure of VGGNet initial design and each layer
Concrete configuration as shown in Figure 1, being mainly made of input layer, convolutional layer, pond layer, full articulamentum, Softmax layers.VGGNet
5 layer groups, including Conv1 convolutional layer group 1, Conv2 convolutional layer group 2, Conv3 convolutional layer group 3, Conv4 convolutional layer group can be divided into
4 and Conv5 convolutional layer group 5, full articulamentum include the full articulamentum 6 of FC6 and the full articulamentum 7 of FC7.Every group of convolutional layer group contain 1~
4 convolutional layers, the size of convolution kernel are unified for 3x3, and each layer group is separated with a pond layer, the size of maximum pond layer
It is 2 × 2.In Fig. 1, the convolutional layer number of plies of Conv1, Conv2 are 1, and the number of convolution kernel is respectively 64,128;Conv3,
The convolutional layer number of plies of Conv4, Conv5 are that the number of the convolutional layer convolution kernel of 2, Conv3 is the convolutional layer of 256, Conv4
The number of convolution kernel is 512, the number of the convolutional layer convolution kernel of Conv5 is 512;The neuron number of FC6 and FC7 is equal
It is 4096.
Oxford University's visual geometric group gives 6 different network structures, respectively A, A- in the paper of VGGNet
LRN, B, C, D, E, their depth gradually deepen 19 layers of (structure D and structure E i.e. VGG16 to structure E from 11 layers of structure A
And VGG19), from the point of view of the result of paper, the effect of structure A is good not as good as structure B, C, D, E, illustrate the number of plies of network it is more deep more
It is good.
In neural network structure, convolutional layer is mainly used for feature extraction, what each convolutional layer was learnt by multi-kernel convolution
Different characteristic extracts original image feature.In convolution process, the weight of convolution kernel immobilizes, after being converted by activation primitive
Corresponding characteristic pattern is obtained, calculating process can be expressed as
As shown in formula (1), the output characteristic pattern of l layers of convolutional layerIt is by upper one layer of output characteristic patternThrough 3 × 3
Convolution kernelBias term is added after convolutionThen it is obtained by activation primitive f (x).
Pond layer is used for Feature Mapping, and the characteristic pattern exported by down-sampling to convolutional layer carries out compression integration, in this way
The purpose done is to retain main feature, reduce parameter and calculation amount, raising model generalization ability.The calculating of pond layer characteristic pattern
Process is as follows
In formula (2),Weight corresponding to characteristic pattern is exported for jth,For corresponding bias, down (x) is
Down-sampling function, down-sampling mode include maximum pond, average pond and random pool etc..Using maximum pond in vgg16 network
Change and carry out compressed data, classification layer uses softmax classifier [11], makes penalty values by stochastic gradient descent method when training
Reach minimum, loss function is as follows
In formula (3), whole parameter matrixs is represented with θ;sicWhen i-th of sample belongs to C class, logic judgment
Value is very, to sicIt is assigned a value of 1, it is false then be assigned a value of 0;It is that output probability value is done into normalized.VGGNet network
Structure is initially designed by ImageNet data set, to be mainly used in large-scale data classification identification, if directly taken
To carry out recognition training to data set of the invention, it is easy to occur to train receipts because of data volume and the particularity of data category
The problems such as holding back difficult and over-fitting.
How to develop a set of technology for identifying and classifying suitable for Bamboo insect pests based on VGGNet network structure has weight
The research significance wanted.
Summary of the invention
Goal of the invention: in order to overcome the deficiencies in the prior art, data sample amount in the method that the prior art is taken
Deficiency, data prediction is complicated, and feature extraction is insufficient, and model-fitting degree fluctuation is larger, not with regard to the particularity of pest and again
Polygamy to model optimizes adjustment, and it is low to result in recognition accuracy.The present invention provides a kind of based on convolutional neural networks mould
The Bamboo insect pests recognition methods of type, based on VGGNet network structure develop it is a set of be suitable for Bamboo insect pests method for identifying and classifying,
The quick identification and classification that common Bamboo insect pests can be achieved are, it can be achieved that automatic identification and classification, recognition accuracy are high.
Technical solution: to achieve the above object, the Bamboo insect pests method for identifying and classifying of the invention based on deep learning should
Method the following steps are included:
S1 acquires Bamboo insect pests sample image data;
S2 handles Bamboo insect pests sample image data, the Bamboo insect pests sample image data collection that obtains that treated;
S3 is optimized using VGG convolutional neural networks model, and to VGG convolutional neural networks model;
S4 carries out target pest identification classification using the VGG convolutional neural networks model after optimization.
Further, in the step S1, the image data of acquisition Bamboo insect pests different growth stages.Such as acquisition evil
Worm, to the image data of adult stage, is used for training pattern from ovum to larva, avoids the posture due to the pest different growth stages
There are errors for the larger bring model inspection of feature difference, further improve model inspection accuracy rate.
Further, in the step S2, comprising the following steps:
S21 handles mistake and duplicate data in Bamboo insect pests sample image data;
S22 expands Bamboo insect pests sample image data by the way of data enhancing.
Further, the step S3 the following steps are included:
S300 selects the structure A of VGG convolutional neural networks model as basis VGG convolutional neural networks model;
S301, from treated, Bamboo insect pests sample image data concentrates selection training set and verifying collection, utilizes training set
Training VGG convolutional neural networks model;
S302, by adjusting the convolution of each convolutional layer in each layer of group in the basis VGG convolutional neural networks model
Nucleus number mesh is trained optimization to the basis VGG convolutional neural networks model;
S303 selects full convolutional layer to replace the full articulamentum in the VGG convolutional neural networks model of basis, convolutional layer with
Last classification work is completed on the basis of the extracted characteristic of pond layer;
S304 carries out again the basis VGG convolutional neural networks model by adjusting the width of the full convolutional layer
Training optimization;
S305 selects the softmax cross entropy cost function of multi-class classification to construct classifier.
Further, it in the step S302, adjusts in the basis VGG convolutional neural networks model in each layer of group
The convolution kernel number of each convolutional layer obtains four kinds of basis VGG convolutional neural networks models, is respectively as follows: Conv1, and 2,3,4,5:
64-128-256-256-256 model structure, Conv1,2,3,4,5:64-128-128-128-128 model structure, Conv1,2,
3,4,5:64-128-64-64-128 model structure and Conv1,2,3,4,5:64-64-64-64-64 model structure;By this four
The structure A of kind basis VGG convolutional neural networks model and original VGG convolutional neural networks model is run respectively, thus to base
Plinth VGG convolutional neural networks model is trained optimization.
Further, the convolution of each convolutional layer in each layer of group in the basis VGG convolutional neural networks model is adjusted
Nucleus number mesh, obtains basic VGG convolutional neural networks model: Conv1, and 2,3,4,5:64-128-64-64-128 model structures are made
For basic VGG convolutional neural networks model.
Further, comprising the following steps:
Inhibit overfitting problem to be trained optimization to the VGG convolutional neural networks model after optimization using dropout, selects
Different optimization algorithms and different dropout probability are selected to be trained experiment, analysis optimization algorithm and dropout probability
With the relationship between identification model accuracy rate, select suitable optimization algorithm and dropout probability to VGG convolutional neural networks
Model carries out suboptimization again.
Further, the optimization algorithm includes two kinds of optimization algorithms of RMSPro and Adam.
Further, different optimization algorithms and following three kinds of dropout probability are selected be trained experiment, it is described
Dropout probability takes 0.5,0.7 and 0.9.
The utility model has the advantages that the present invention compared with the prior art, has the advantage, that
1, the present invention provides a kind of based on the Bamboo insect pests recognition methods based on convolutional neural networks model, is based on VGGNet
Network structure develop it is a set of suitable for Bamboo insect pests method for identifying and classifying, it can be achieved that common Bamboo insect pests it is quick identification with
Classification is, it can be achieved that automatic identification and classification, recognition accuracy are high;
2, the present invention further optimizes and improves to the structure and training parameter of VGG convolutional neural networks model, optimizes
Model accuracy rate afterwards reaches 99.02%, improves 4.04 percentage points than the network structure before optimization;
3, the structure A of the invention with VGG convolutional neural networks model is basis VGG convolutional neural networks model, according to number
According to redundancy and over-fitting degree, its structure and training parameter are further optimized and improved, it is more applicable compared to other models
It is identified in bamboo grove common insect pests;
4, the present invention inhibits overfitting problem using dropout to complete to train test experiments, and it is normal to further increase bamboo grove
See pest recognition accuracy;
5, the present invention carries out tuning to by the parameter to convolution nuclear volume and full convolutional layer width etc., further increases bamboo
Woods common insect pests recognition accuracy;
6, the present invention, can be for different phase in pest developmental process when acquiring bamboo grove common insect pests sample image data
Image Acquisition is carried out, such as acquisition pest avoids to the image data of adult stage for training pattern from ovum to larva
Since the aspectual character of pest different growth stages differs greatly, there are errors for bring model inspection, further improve model
Detection accuracy;
7, the present invention is by the bamboo grove common insect pests sample image data of acquisition in such a way that this kind of data enhance
The data volume of every kind of insect is expanded by carrying out the operation such as distortion cutting and color adjustment to sample image, improves net by reason
Network is to pest classification accuracy, the problems such as keeping the performance of network more preferable, prevent over-fitting, further improves model inspection standard
True rate.
Detailed description of the invention
Fig. 1 is VGG convolutional neural networks prototype network structure chart.
Fig. 2 is the bamboo grove common insect pests recognition methods flow chart the present invention is based on deep learning.
Fig. 3 be in embodiment six Bamboo insect pests bamboo as sample image;
Fig. 4 is the image that Fig. 3 sample image obtains after cutting;
Fig. 5 is the image that Fig. 3 sample image obtains after distortion;
Fig. 6 is the image that Fig. 3 sample image obtains after color adjusts;
Fig. 7 is impact effect figure of the different convolution kernel numbers to convolutional neural networks verifying accuracy rate;
Fig. 8 is impact effect figure of the width to N4 model discrimination of full convolutional layer;
Fig. 9 is impact effect figure of the width to VGG_A model discrimination of full convolutional layer;
Figure 10 (a) be use RMSProp optimization algorithm, dropout probability for 0.5 experimental result picture;
Figure 10 (b) be use RMSProp optimization algorithm, dropout probability for 0.7 experimental result picture;
Figure 10 (c) be use RMSProp optimization algorithm, dropout probability for 0.9 experimental result picture;
Figure 10 (d) be use Adam optimization algorithm, dropout probability for 0.5 experimental result picture;
Figure 10 (e) be use Adam optimization algorithm, dropout probability for 0.7 experimental result picture;
Figure 10 (f) be use Adam optimization algorithm, dropout probability for 0.7 experimental result picture;
Icon:
1, convolutional layer group 1;2, convolutional layer group 2;3, convolutional layer group 3;4, convolutional layer group 4;5, convolutional layer group 5;6, full articulamentum 1;
7, full articulamentum 2.
Specific embodiment
The present invention will be further explained with reference to the accompanying drawing.
Embodiment one:
A kind of Bamboo insect pests recognition methods based on convolutional neural networks model of the present embodiment, reference Fig. 2, including it is following
Step:
Acquire Bamboo insect pests sample image data;
Bamboo insect pests sample image data is handled, the Bamboo insect pests sample image data collection that obtains that treated;
It is optimized using VGG convolutional neural networks model and to VGG convolutional neural networks model:
VGG convolutional neural networks model referring to Fig.1, mainly by input layer, convolutional layer, pond layer, full articulamentum,
Softmax layers of composition.VGGNet can be divided into 5 layer groups, including Conv1 convolutional layer group 1, Conv2 convolutional layer group 2, Conv3 volumes
Lamination group 3, Conv4 convolutional layer group 4 and Conv5 convolutional layer group 5, full articulamentum include the full articulamentum 6 of FC6 and the full articulamentum of FC7
7.Every group of layer group contains 1~4 convolutional layer, and the size of convolution kernel is unified for 3x3, and each layer group is separated with a pond layer,
The size of maximum pond layer is 2 × 2, and the convolutional layer number of plies of Conv1, Conv2 are 1 in Fig. 1, and the number of convolution kernel is respectively
64,128;The convolutional layer number of plies of Conv3, Conv4, Conv5 are that the number of the convolutional layer convolution kernel of 2, Conv3 is 256,
The number of the convolutional layer convolution kernel of Conv4 is 512, the number of the convolutional layer convolution kernel of Conv5 is 512;FC6's and FC7
Neuron number is 4096.
Convolutional layer is mainly used for feature extraction, and each convolutional layer extracts original image by the different characteristic that multi-kernel convolution learns
Feature.
Pond layer is used for Feature Mapping, and the characteristic pattern exported by down-sampling to convolutional layer carries out compression integration, in this way
The purpose done is to retain main feature, reduce parameter and calculation amount, raising model generalization ability.
Full articulamentum is placed in the decline in convolutional neural networks, its effect is to complete classification work.
Softmax layers are completed final classification work to construct classifier.
VGG convolutional neural networks model is optimized, carries out mesh using the VGG convolutional neural networks model after optimization
Mark pest identification classification.
Embodiment two:
A kind of Bamboo insect pests recognition methods based on convolutional neural networks model of the present embodiment is based on embodiment one, adopts
Collect the image data of Bamboo insect pests different growth stage, for example, acquisition pest from ovum to larva to the image data of adult stage,
For training pattern, avoid since the aspectual character of pest different growth stages differs greatly bring model inspection in the presence of mistake
Difference, certain pest species are different, and the growth stage of pest is different.Furthermore optionally retain Bamboo insect pests sample image color, because
It is not only had differences on shape, texture for pest, color is also an important factor for influencing recognition accuracy, therefore, in data
In treatment process, input image data is not converted into gray level image, but retains tri- Color Channels of RGB of image
Value.
Embodiment three:
A kind of Bamboo insect pests recognition methods based on convolutional neural networks model of the present embodiment is based on embodiment two, right
Bamboo insect pests sample image data is handled, the Bamboo insect pests sample image data collection that obtains that treated, including following step
It is rapid:
Handle mistake and duplicate data in Bamboo insect pests sample image data;
Expand Bamboo insect pests sample image data by the way of data enhancing.
Expand Bamboo insect pests sample image data by the way of data enhancing: after Screening Treatment, every kind of insect sample
This quantity is different, the problems such as in order to improve the classification accuracy of network, keep the performance of network more preferable, prevent over-fitting, this reality
Example is applied by the way of data enhancing to expand above-mentioned Bamboo insect pests sample image data amount, by above-mentioned Bamboo insect pests sample
This image data is operated, for example, operating in the present embodiment to the larva sample image data of above-mentioned Bamboo insect pests one
Expand, the larva sample image of above-mentioned Bamboo insect pests one is cut, distort and color adjustment, by above-mentioned Bamboo insect pests one
Larva sample image data amount amplification.
Example IV:
A kind of Bamboo insect pests recognition methods based on convolutional neural networks model of the present embodiment is based on embodiment three, adopts
With VGG convolutional neural networks model, VGG convolutional neural networks model is optimized, comprising the following steps:
Select the structure A of VGG convolutional neural networks model as basis VGG convolutional neural networks model;
From treated, Bamboo insect pests sample image data concentrates selection training set and verifying collection, utilizes training set training
VGG convolutional neural networks model;
By adjusting the convolution nucleus number of each convolutional layer in each layer of group in the basis VGG convolutional neural networks model
Mesh is trained optimization to the basis VGG convolutional neural networks model:
Convolutional layer in VGG convolutional neural networks structure is responsible for extracting low-level image feature, is the portion of most critical in whole network
Point, other than the size of convolution kernel, the quantity of convolution can also be impacted discrimination, as used in the present invention
The data set ImageNet data set that is far from it is huge, therefore the structure A in preferential selection VGGNet is tested, and is being tested
In the process, the convolution kernel number of each convolutional layer can be also gradually reduced to be trained and test, and investigate the number of local receptor field
Influence of the mesh to model performance.In addition, the present embodiment uses Gaussian Profile to the initialization selection of weight, mean value is set as 0,
Standard deviation is set as 0.01.
Full convolutional layer is selected to replace the full articulamentum in the VGG convolutional neural networks model of basis, in convolutional layer and pond layer
Last classification work is completed on the basis of extracted characteristic:
Full articulamentum is placed in the decline in convolutional neural networks, its effect is to complete classification work, due to complete
It is fine and close connection between articulamentum, joins high number, although in this way model learning ability can be reinforced with lift scheme complexity,
But this is also easy to that over-fitting occurs, so that operation time, inefficiency, consuming storage.For this purpose, the present embodiment selects
It replaces connecting entirely using full convolution Fully Convolutional Networks (FCN), between full articulamentum and convolutional layer
Unique difference is exactly that the neuron in convolutional layer is only connect with a regional area in input data, and is arranged in convolution
In neuron shared parameter.However in two class layers, neuron is all to calculate dot product, so their functional form is one
Sample, therefore the two can mutually convert.Convolutional network can be allowed sliding on the input picture of a Zhang Geng great using full convolutional layer
It is dynamic, multiple outputs are obtained, influence of the location information to classification is reduced, improves the robustness of model, reduce characteristic extraction procedure
In, influence of the location information to classification improves the robustness of model, by adjusting the full convolutional layer width to the base
Plinth VGG convolutional neural networks model carries out training optimization again;
The softmax cross entropy cost function of multi-class classification is selected to construct classifier:
The present embodiment constructs classifier using the softmax cross entropy cost function for being more suitable for multi-class classification, for
The initial value of learning rate learing_rate is set as 0.001, and after 1000 step of every repetitive exercise, learning rate is reduced to
Half originally.In addition, train epochs are set as 10000 by the present embodiment in order to enable model to be trained up,
Bach_size is set as 64.
In the present embodiment will based on VGG convolutional neural networks model, according to data redundancy and over-fitting degree,
Its structure and training parameter are further optimized and improved, so that the model after improvement is more suitable for the knowledge of bamboo grove common insect pests
Not,
Data enhancing has been carried out by image cropping and color adjustment first on the image data set of arrangement;Then to 11
The VGG convolutional neural networks model of layer carries out tuning by the parameter to convolution nuclear volume and full convolutional layer width etc., and uses
Dropout inhibits overfitting problem to complete to train test experiments.
The experimental results showed that identifying pest using the model after optimization, accuracy rate reaches 99.02%, before optimization
Network structure accuracy rate improves 4.04 percentage points.
Embodiment five:
A kind of Bamboo insect pests recognition methods based on convolutional neural networks model of the present embodiment is based on example IV, adopts
Inhibit overfitting problem to be trained optimization to the VGG convolutional neural networks model after optimization with dropout, utilizes neuron
The random update for inhibiting (Dropout) to inhibit partial parameters in network, finds optimal parameter in 0.5~0.9 range.
Dropout refers to allows the weight of the certain hidden layer nodes of network to be not used at random in model training, does not use
Those of node can temporarily not think be network structure a part, but its weight must remain, because of next sample
It may be needed to be used when this input.After using Dropout, some variations will occur for the training of neural network, corresponding
Formula variation is as follows.
There is no the neural network of Dropout:
There is the neural network of Dropout:
Dropout can regard that a kind of model is average as, and so-called model is average, be exactly from different moulds as its name suggests
The estimation of type is predicted through certain power in the training process of each batch, due to the hidden layer section ignored at random every time
Point is all different, thus makes the network trained every time all and be different, and each training can singly do the mould of one " new "
Type;In addition, implicit node is all to occur at random with certain probability, therefore cannot be guaranteed that every 2 implicit nodes go out simultaneously every time
Existing, the update of such weight is no longer dependent on the collective effect that fixed relationship implies node, and certain features is prevented only to exist
Situation just effective under other special characteristics.Dropout process is exactly that a very effective neural network model is flat in this way
Equal method carrys out consensus forecast probability by a large amount of different network of training.
Select different optimization algorithms and different dropout probability be trained experiment, analysis optimization algorithm and
Relationship between dropout probability and identification model accuracy rate selects suitable optimization algorithm and dropout probability to VGG volumes
Product neural network model carries out suboptimization again.Selection uses Adam optimizer, because Adam optimization algorithm introduces quadratic power ladder
Correction, is conducive to the searching of globe optimum, it faster than stochastic gradient descent method, while also reducing and falling into local optimum
Risk.
Embodiment six:
A kind of Bamboo insect pests recognition methods based on convolutional neural networks model of the present embodiment is based on embodiment five, this
Using Bamboo insect pests as test object in embodiment, 12 kinds of pests of bamboo grove are chosen, acquire 12 kinds of Bamboo insect pests sample graphs
As data, and the image data of Bamboo insect pests different growth stages is acquired, and optionally retain Bamboo insect pests sample image face
Color because pest not only has differences on shape, texture, color be also influence recognition accuracy an important factor for, therefore,
In data processing, input image data is not converted into gray level image, but retains tri- face of RGB of image
The value of chrominance channel;The image data for acquiring Bamboo insect pests different growth stage, as Bamboo insect pests from ovum to larva to adult respectively at
The image data in long stage avoids the aspectual character due to the insect different growth stage from differing greatly bring model inspection effect
Difference, Bamboo insect pests sample image data is mainly obtained by modes such as indoor shot and search engine retrievings in the present embodiment,
The Bamboo insect pests sample image data of acquisition is as shown in table 1,
Table 1
In above-mentioned table 1, the Bamboo insect pests sample arranged has 12 kinds altogether, to 12 kinds of Bamboo insect pests samples of above-mentioned acquisition
Image data handled.
Handle Bamboo insect pests sample image data in mistake and duplicate data: in the present embodiment by manual sort's screening Lai
Avoid the repetition and mistake of data;
Expand Bamboo insect pests sample image data by the way of data enhancing: after Screening Treatment, every kind of insect sample
This quantity is different, the problems such as in order to improve the classification accuracy of network, keep the performance of network more preferable, prevent over-fitting, this reality
Example is applied by the way of data enhancing to expand above-mentioned Bamboo insect pests sample image data amount, by above-mentioned Bamboo insect pests sample
This image data is operated, and carries out operation expansion to the larva sample image data of 12 kinds of pests of above-mentioned bamboo grove in the present embodiment
Fill, the larva sample image of 12 kinds of pests of above-mentioned bamboo grove is cut, distort and color adjustment, such as scheme, by above-mentioned bamboo grove 12
The larva sample image data amount of kind pest is expanded to 500;It is operated with this, in the present embodiment, by above-mentioned Bamboo insect pests sample
This image data amount extends to 500.
The different growth stage image datas of collecting part pest, acquire ovum, larva and the adult of pest in the present embodiment
Stage image data, therefore, the Bamboo insect pests sample arranged have 24 kinds altogether, after Screening Treatment, every kind of insect sample
Quantity is different, the problems such as in order to improve the classification accuracy of network, keep the performance of network more preferable, prevent over-fitting, uses herein
The mode of data enhancing carrys out expanding data amount, by carrying out the operation such as distortion cutting and color adjustment to image, by every kind of insect
Data volume expand to 1000.The Bamboo insect pests sample arranged in the present embodiment has 24 kinds, totally 24000 width Bamboo insect pests altogether
Sample image randomly selects 16800 width as training set to wherein each classification, and the sample size of 72600 width is as verifying collection.
If the size of Bamboo insect pests sample image is 224 × 224 × 3, the corresponding Bamboo insect pests sample image pair of training set is utilized
VGG convolutional neural networks model is trained.Referring to Fig. 3, Fig. 4, Fig. 5 and Fig. 6, Fig. 3 is Bamboo insect pests bamboo as sample image,
Fig. 4 is the image that Fig. 3 sample image obtains after cutting, and Fig. 5 is the image that Fig. 3 sample image obtains after distortion, figure
6 be the image that Fig. 3 sample image obtains after color adjusts.Other pest images expansions and so on.
Embodiment seven:
A kind of Bamboo insect pests recognition methods based on convolutional neural networks model of the present embodiment is based on embodiment six, this
Desktop computer in embodiment using TensorFlow deep learning frame and equipped with GPU graphics processor is completed to test,
Specific running environment is as shown in table 2,
Table 2
Embodiment eight:
A kind of Bamboo insect pests recognition methods based on convolutional neural networks model of the present embodiment is based on embodiment seven, leads to
The convolution kernel number for adjusting each convolutional layer in each layer of group in the basis VGG convolutional neural networks model is crossed, to described
Basic VGG convolutional neural networks model is trained optimization, convolutional layer as the floor portions in convolutional neural networks structure,
Feature extraction directly is carried out to input picture, therefore is considered as most sensitive parameter, the present embodiment is first by adjusting convolution
The number of core is tested, and the relationship between convolution kernel number and recognition accuracy is analyzed.Network structure adjusted, fortune
Row is time-consuming, verify accuracy rate and loss value is as shown in table 3, influence of the convolution kernel number to convolutional neural networks verifying accuracy rate
As shown in Figure 7.
Table 3
As can be seen from Table 2, adjusting the volume of each convolutional layer in each layer of group in the basis VGG convolutional neural networks model
Product nucleus number mesh, obtains four kinds of basis VGG convolutional neural networks models, is respectively as follows: N:2:Conv1, and 2,3,4,5:64-128-
256-256-256 model structure, N3:Conv1,2,3,4,5:64-128-128-128-128 model structure, N4:Conv1,2,
3,4,5:64-128-64-64-128 model structure and N5:Conv1,2,3,4,5:64-64-64-64-64 model structure;It should
The structure A-N1 of four kinds of basis VGG convolutional neural networks models and original VGG convolutional neural networks model is run respectively, from
And optimization is trained to basic VGG convolutional neural networks model,
It can be seen that N4:Conv1 from table 3 and Fig. 7, the verifying of 2,3,4,5:64-128-64-64-128 model structures is quasi-
True rate highest, penalty values are minimum, and time consumption for training is relatively also relatively low, thus by optimizing to model training after, select N4:
Basis VGG convolutional neural networks model based on Conv1,2,3,4,5:64-128-64-64-128 model structure.
For class categories are not king-sized common insect pests data set, the quantity of characteristic pattern can be made greatly very much instead
At in model parameter redundancy, computing resource waste and operational efficiency are low the problems such as.The reduction network architecture reduces convolution
The number of core can not only improve discrimination and the time consumed by training can also be effectively reduced.It should be noted that not being
The fewer the number of convolution kernel the better, if convolution kernel number is very few, as feature extraction is insufficient and leads to classification results not
It is ideal.
Embodiment nine:
A kind of Bamboo insect pests recognition methods based on convolutional neural networks model of the present embodiment is based on embodiment eight, choosing
The full articulamentum in full convolutional layer replacement basis VGG convolutional neural networks model is selected, it is extracted in convolutional layer and pond layer
Last classification work is completed on the basis of characteristic, by adjusting the full convolutional layer width to the basis VGG convolution
Neural network model carries out training optimization again, and full convolutional layer is complete on the basis of convolutional layer and pond layer extracted feature
At last classification work, the present embodiment select in embodiment eight the N4 model that behaves oneself best and original VGG_A model come into
Row experiment, influence of the width of full convolutional layer to it is as shown in Figure 8 and Figure 9, and Fig. 8 is that the width of full convolutional layer knows N4 model
The not impact effect figure of rate, Fig. 9 are impact effect figure of the width to VGG_A model discrimination of full convolutional layer, can be with from Fig. 8
Find out, for N4 model structure, the higher but wide full convolutional layer of the complete wider accuracy rate of convolutional layer is easy to instruct
White silk does not restrain, expends the problems such as time;And for more complicated VGG_A model, as shown in figure 9, classifying not in training sample
In the case where more, the width for suitably reducing full convolutional layer also can achieve the purpose for improving recognition accuracy.Generally speaking, right
In the classification problem of data set, full convolutional layer has an optimal width to reach better recognition effect.
Embodiment ten:
A kind of Bamboo insect pests recognition methods based on convolutional neural networks model of the present embodiment is based on embodiment nine, adopts
Inhibit overfitting problem to be trained optimization to the VGG convolutional neural networks model after optimization with dropout, selects different
Optimization algorithm and different dropout probability are trained experiment, analysis optimization algorithm and dropout probability and identification mould
Relationship between type accuracy rate selects suitable optimization algorithm and dropout probability to carry out VGG convolutional neural networks model
Suboptimization again.Optimization algorithm includes that two kinds of optimizations of RMSProRMSProp (rootmean squareprop) [16] and Adam are calculated
Method.As shown in Figure 10 (a), 10 (b), 10 (c), 10 (d), 10 (e) and 10 (f), 10 (a), 10 (b) and 10 (c) are used
RMSProp method, 10 (d), 10 (e) and 10 (f) use Adam method, and 10 (a) and 10 (d) dropout is 0.5,10
It (b) is 0.9 with 10 (e) the dropout dropout for being 0.7,10 (c) and 10 (f), experimental result is shown, two kinds of optimization sides
The resulting result of method is not much different, the difference is that Adam method ratio RMSProp method has preferably in overfitting problem
Performance, and RMSProp method ratio Adam method convergence speed will be with fast.In addition, for Adam method, dropout
Probability setting is too small to be made training fluctuation very big and extend the training time;For RMSProp method then need by
The smaller overfitting problem to inhibit it is arranged in dropout probability.
RMSProp allows the every bout of r to decay certain proportion by introducing an attenuation coefficient.It, which is implemented, needs to mention
For global learning rate ∈, initial parameter θ, numerical stability amount δ, rate of decay ρ, gradient cumulative amount r (being initialized as 0) etc. ginseng
Number, every step iteration randomly select the sample { x that a batch volume is m from training set1..., xmAnd relevant output yi。
Its gradient calculates as follows:
Accumulative squared gradient is as follows:
r←ρr+(1-ρ)g⊙g
Calculating parameter updates as follows:
θ←θ+Δθ
The problem of terminating too early in RMSProp optimization algorithm very good solution deep learning, is suitble to processing non-stationary mesh
Mark, but due to introducing new super ginseng, attenuation coefficient ρ, still dependent on global learning rate again.
Adam is substantially the RMSprop with momentum term, it is dynamic using the single order moments estimation and second order moments estimation of gradient
State adjusts the learning rate of each parameter.The advantages of Adam, essentially consists in after bias correction, and iterative learning rate has each time
A determining range, so that parameter is more steady.It, which is implemented, needs step value ∈, initial parameter θ, numerical stability amount δ, and one
Rank momentum attenuation coefficient ρ1, second order momentum attenuation coefficient ρ2, wherein the value of three parameters is general are as follows: δ=10-8, ρ1=0.9,
ρ1=0.999.Single order momentum s, second order momentum r, is initialized to 0.Every step iteration randomly selects a batch volume from training set
For the sample { x of m1..., xmAnd relevant output yi。
Its gradient calculates as follows:
t←t+1
Inclined first moment is estimated as follows:
s←ρ1r+(1-ρ1)g
Inclined second moment is estimated as follows:
r←ρ2+(1-ρ2)g⊙g
Correct the deviation of first moment:
Correct the deviation of second moment:
Calculating parameter updates as follows:
θ←θ+Δθ
Adam optimization algorithm combines RMSprop and is good at the advantages of handling non-stationary target, calculates not for different parameters
Same autoadapted learning rate, memory requirements is smaller, is suitable for most non-convex optimization and large data sets and higher dimensional space.
Gradient and error are calculated, r is updated, further according to r and gradient calculating parameter renewal amount.
After the completion of model training, result is shown on the visualization tool TensorBoard that TensorFlow is carried.And it will
Trained bamboo shoots pest identification network model saves as the file of ckpt format, is needing to carry out bamboo shoots pest Classification and Identification
When, pest image to be sorted is input in trained model, the recognition result then fed back to.
Embodiment 11:
A kind of Bamboo insect pests recognition methods based on convolutional neural networks model of the present embodiment is based on embodiment ten, this
Invent the bamboo grove common insect pests recognition methods based on deep learning proposed, referring to Fig. 2, windows operating system,
The training and test of convolutional neural networks are carried out in the environment of TensorFlow frame and eclipse programming platform.Test iteration
Number is set as 4000 times, and learning rate is set as 0.002, with the increase of frequency of training, is learnt after 2000 iteration of every completion
Rate reduces 2 times automatically.
To its pond layer of Artificial Neural Network Structures described above using average Downsapling method, gradient descent method
Using AdamOptimizer descent method, classifier selection Softmax return method come test CNN recognition accuracy and
Penalty values loss.
After the completion of model training, result is shown on the visualization tool TensorBoard that TensorFlow is carried.And it will
Trained bamboo grove common insect pests identification network model saves as the file of ckpt format, is needing to carry out bamboo grove common insect pests
When Classification and Identification, pest image to be sorted is input in trained model, the recognition result then fed back to.
Model step is specific as follows:
Input layer: using 224 × 224 × 3 Bamboo insect pests image pixel matrix as input;
C1:C1 layers of convolutional layer be by multiple and different filters to input matrix carry out convolution algorithm obtain it is multiple and different
Original signal characteristic extract layer.This paper Selection utilization 96 are rolled up the image of input layer having a size of 11 × 11 convolution kernel
Product operation, convolution results carry out be calculated in C1 layers 64 differences by an activation primitive plus after 6 bias vectors
Characteristic pattern;
S1:S1 layers of down-sampling layer carry out down-sampling to upper one layer of C1 and handle to obtain 96 new characteristic patterns, and sampling window is big
Small is 2 × 2.The resolution ratio of multiple characteristic patterns in C1 layers is reduced here by the mode of simple scalability, to reduce net below
The quantity of input weight parameter in network, and weaken output for the sensitivity for being displaced and deforming, make CNN to image scaling, sky
Between displacement, rotation etc. influences have preferable robustness;
Convolutional layer C2: S1 layers of multiple characteristic patterns progress convolution algorithm is obtained by the way of being similar to C1 layer more
A feature extraction figure, C2 layer choosing, which is selected, carries out feature extraction with 5 × 5 convolution kernel, and the amplification of convolution nuclear volume is 256;
Pond layer S2: down-sampling is carried out using 2 × 2 windows identical with S1;
The convolution kernel that convolutional layer C3:C3 layer choosing selects 3 × 3 carries out feature extraction, convolution kernel to S2 layers of multiple characteristic patterns
Quantity is still 256;
The convolution kernel that pond layer C4:C4 layer choosing selects 3 × 3 carries out feature extraction, and convolution nuclear volume is 256;
The convolution kernel that pond layer C5:C5 layer choosing selects 3 × 3 carries out feature extraction, and convolution nuclear volume is 256;
Pond layer S5: down-sampling is carried out using 2 × 2 windows;
Full articulamentum F6: similar to common BP hidden layer, selection here carries out full connection behaviour using 2048 neurons
Make;
Full articulamentum F7: full attended operation is carried out using 2048 neurons identical with F6;
Output layer Output: due to needing the food leaf class pest common to bamboo grove, tree borer, root-feeding insect, thorn herein
Inhaling class pest etc., totally 33 class pests carry out Classification and Identification, so output layer neuron number is 33.
The training of convolutional neural networks model is divided into two stages, specific as follows:
Propagation stage forward: (1) from taking out sample in sample set and be input in network model;(2) it calculates corresponding
Export Op, in this stage, information is ultimately delivered to output layer by transformation step by step from input layer, and in the process, network executes
Be input and every layer of weight matrix phase dot product:
Op=Fn(...(F2(F1(XpW(1))W(2))...)W(n))
The back-propagation stage: (1) reality output O is calculatedpWith corresponding ideal output YpDifference;(2) minimization error is pressed
Method backpropagation adjust weight matrix.
Activation primitive corrects linear unit function using ReLU, it calculates f (x)=max (0, x) to input, in other words,
Be exactly input data less than 0 when, forced boil down to 0, on the contrary then data remain unchanged;
Last full articulamentum uses dropout to reduce over-fitting, and the value of dropout is set as 0.5;By
The repetitive exercise of 4000 steps, final recognition effect can achieve 90% or more accuracy rate.
The above is only a preferred embodiment of the present invention, it should be pointed out that: those skilled in the art are come
It says, various improvements and modifications may be made without departing from the principle of the present invention, these improvements and modifications should also regard
For protection scope of the present invention.
Claims (9)
1. a kind of Bamboo insect pests recognition methods based on convolutional neural networks model, it is characterised in that: this method includes following step
It is rapid:
S1 acquires Bamboo insect pests sample image data;
S2 handles Bamboo insect pests sample image data, the Bamboo insect pests sample image data collection that obtains that treated;
S3 is optimized using VGG convolutional neural networks model, and to VGG convolutional neural networks model;
S4 carries out target pest identification classification using the VGG convolutional neural networks model after optimization.
2. the Bamboo insect pests recognition methods according to claim 1 based on convolutional neural networks model, it is characterised in that: institute
It states in step S1, the image data of acquisition Bamboo insect pests different growth stages.
3. the Bamboo insect pests recognition methods according to claim 1 based on convolutional neural networks model, it is characterised in that: institute
It states in step S2, comprising the following steps:
S21 handles mistake and duplicate data in Bamboo insect pests sample image data;
S22 expands Bamboo insect pests sample image data by the way of data enhancing.
4. the Bamboo insect pests recognition methods according to claim 1 based on convolutional neural networks model, it is characterised in that: institute
State step S3 the following steps are included:
S300 selects the structure A of VGG convolutional neural networks model as basis VGG convolutional neural networks model;
S301, from treated, Bamboo insect pests sample image data concentrates selection training set and verifying collection, utilizes training set training
VGG convolutional neural networks model;
S302, by adjusting the convolution nucleus number of each convolutional layer in each layer of group in the basis VGG convolutional neural networks model
Mesh is trained optimization to the basis VGG convolutional neural networks model;
S303 selects full convolutional layer to replace the full articulamentum in the VGG convolutional neural networks model of basis, in convolutional layer and pond layer
Last classification work is completed on the basis of extracted characteristic;
S304 trains the basis VGG convolutional neural networks model by adjusting the width of the full convolutional layer again
Optimization;
S305 selects the softmax cross entropy cost function of multi-class classification to construct classifier.
5. the Bamboo insect pests recognition methods according to claim 4 based on convolutional neural networks model, it is characterised in that: institute
It states in step S302, adjusts the convolution nucleus number of each convolutional layer in each layer of group in the basis VGG convolutional neural networks model
Mesh obtains four kinds of basis VGG convolutional neural networks models, is respectively as follows: Conv1,2,3,4,5:64-128-256-256-256 moulds
Type structure, Conv1,2,3,4,5:64-128-128-128-128 model structure, Conv1,2,3,4,5:64-128-64-64-
128 model structures and Conv1,2,3,4,5:64-64-64-64-64 model structure;By this four kinds basis VGG convolutional neural networks
The structure A of model and original VGG convolutional neural networks model is run respectively, thus to basic VGG convolutional neural networks model
It is trained optimization.
6. the Bamboo insect pests recognition methods according to claim 4 based on convolutional neural networks model, it is characterised in that: adjust
In the whole basis VGG convolutional neural networks model in each layer of group each convolutional layer convolution kernel number, obtain basic VGG volumes
Product neural network model: Conv1,2,3,4,5:64-128-64-64-128 model structures, as basic VGG convolutional neural networks
Model.
7. the Bamboo insect pests recognition methods according to claim 4 based on convolutional neural networks model, it is characterised in that: packet
Include following steps:
Overfitting problem is inhibited to be trained optimization to the VGG convolutional neural networks model after optimization using dropout, selection is not
With optimization algorithm and different dropout probability be trained experiment, analysis optimization algorithm and dropout probability and identification
Relationship between model accuracy rate selects suitable optimization algorithm and dropout probability to carry out VGG convolutional neural networks model
Suboptimization again.
8. the Bamboo insect pests recognition methods according to claim 7 based on convolutional neural networks model, it is characterised in that: institute
Stating optimization algorithm includes two kinds of optimization algorithms of RMSPro and Adam.
9. the Bamboo insect pests recognition methods according to claim 7 based on convolutional neural networks model, it is characterised in that: choosing
Different optimization algorithms and following three kinds of dropout probability are selected to be trained experiment, the dropout probability takes 0.5,0.7
With 0.9.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811138103.9A CN109558787A (en) | 2018-09-28 | 2018-09-28 | A kind of Bamboo insect pests recognition methods based on convolutional neural networks model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811138103.9A CN109558787A (en) | 2018-09-28 | 2018-09-28 | A kind of Bamboo insect pests recognition methods based on convolutional neural networks model |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109558787A true CN109558787A (en) | 2019-04-02 |
Family
ID=65864778
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811138103.9A Pending CN109558787A (en) | 2018-09-28 | 2018-09-28 | A kind of Bamboo insect pests recognition methods based on convolutional neural networks model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109558787A (en) |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110070029A (en) * | 2019-04-17 | 2019-07-30 | 北京易达图灵科技有限公司 | A kind of gait recognition method and device |
CN110082283A (en) * | 2019-05-23 | 2019-08-02 | 山东科技大学 | A kind of Atmospheric particulates SEM image recognition methods and system |
CN110097104A (en) * | 2019-04-22 | 2019-08-06 | 安徽大学 | Apple Leaves disease classification method and its automatic identification equipment based on deep learning |
CN110287574A (en) * | 2019-06-19 | 2019-09-27 | 浙江农林大学 | A kind of multichannel grain storage pest fumigant diffusion rate and insecticide efficiency assess device |
CN110298388A (en) * | 2019-06-10 | 2019-10-01 | 天津大学 | Based on the 5 kinds of damage caused by a drought recognition methods of corn for improving VGG19 network |
CN110378435A (en) * | 2019-07-25 | 2019-10-25 | 安徽工业大学 | A method of the Apple Leaves disease recognition based on convolutional neural networks |
CN110942401A (en) * | 2019-11-21 | 2020-03-31 | 黑龙江电力调度实业有限公司 | Intelligent communication method for power Internet of things |
CN111127423A (en) * | 2019-12-23 | 2020-05-08 | 金陵科技学院 | Rice pest and disease identification method based on CNN-BP neural network algorithm |
CN111428798A (en) * | 2020-03-30 | 2020-07-17 | 北京工业大学 | Plant seedling classification method based on convolutional neural network |
CN111504308A (en) * | 2020-04-21 | 2020-08-07 | 南京比特互动创意科技有限公司 | Intelligent exhibition hall supporting system based on AR technology |
CN111611889A (en) * | 2020-05-12 | 2020-09-01 | 安徽大学 | Miniature insect pest recognition device in farmland based on improved convolutional neural network |
CN112529043A (en) * | 2020-11-19 | 2021-03-19 | 江苏科技大学 | MDRN network-based online grain storage pest identification system and method |
CN112598265A (en) * | 2020-12-18 | 2021-04-02 | 武汉大学 | Decoupling risk estimation-based rapid detection method for hyperspectral pine nematode disease of unmanned aerial vehicle |
CN114092799A (en) * | 2021-10-28 | 2022-02-25 | 长春工业大学 | Forestry pest identification and detection method based on pooling vision Transformer |
CN114494678A (en) * | 2021-12-02 | 2022-05-13 | 国家计算机网络与信息安全管理中心 | Character recognition method and electronic equipment |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105740892A (en) * | 2016-01-27 | 2016-07-06 | 北京工业大学 | High-accuracy human body multi-position identification method based on convolutional neural network |
CN107016405A (en) * | 2017-02-24 | 2017-08-04 | 中国科学院合肥物质科学研究院 | A kind of insect image classification method based on classification prediction convolutional neural networks |
CN107169974A (en) * | 2017-05-26 | 2017-09-15 | 中国科学技术大学 | It is a kind of based on the image partition method for supervising full convolutional neural networks more |
CN107886510A (en) * | 2017-11-27 | 2018-04-06 | 杭州电子科技大学 | A kind of prostate MRI dividing methods based on three-dimensional full convolutional neural networks |
CN108009629A (en) * | 2017-11-20 | 2018-05-08 | 天津大学 | A kind of station symbol dividing method based on full convolution station symbol segmentation network |
CN108564589A (en) * | 2018-03-26 | 2018-09-21 | 江苏大学 | A kind of plant leaf blade dividing method based on the full convolutional neural networks of improvement |
-
2018
- 2018-09-28 CN CN201811138103.9A patent/CN109558787A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105740892A (en) * | 2016-01-27 | 2016-07-06 | 北京工业大学 | High-accuracy human body multi-position identification method based on convolutional neural network |
CN107016405A (en) * | 2017-02-24 | 2017-08-04 | 中国科学院合肥物质科学研究院 | A kind of insect image classification method based on classification prediction convolutional neural networks |
CN107169974A (en) * | 2017-05-26 | 2017-09-15 | 中国科学技术大学 | It is a kind of based on the image partition method for supervising full convolutional neural networks more |
CN108009629A (en) * | 2017-11-20 | 2018-05-08 | 天津大学 | A kind of station symbol dividing method based on full convolution station symbol segmentation network |
CN107886510A (en) * | 2017-11-27 | 2018-04-06 | 杭州电子科技大学 | A kind of prostate MRI dividing methods based on three-dimensional full convolutional neural networks |
CN108564589A (en) * | 2018-03-26 | 2018-09-21 | 江苏大学 | A kind of plant leaf blade dividing method based on the full convolutional neural networks of improvement |
Cited By (21)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110070029A (en) * | 2019-04-17 | 2019-07-30 | 北京易达图灵科技有限公司 | A kind of gait recognition method and device |
CN110097104A (en) * | 2019-04-22 | 2019-08-06 | 安徽大学 | Apple Leaves disease classification method and its automatic identification equipment based on deep learning |
CN110082283A (en) * | 2019-05-23 | 2019-08-02 | 山东科技大学 | A kind of Atmospheric particulates SEM image recognition methods and system |
CN110082283B (en) * | 2019-05-23 | 2021-12-14 | 山东科技大学 | Atmospheric particulate SEM image recognition method and system |
CN110298388A (en) * | 2019-06-10 | 2019-10-01 | 天津大学 | Based on the 5 kinds of damage caused by a drought recognition methods of corn for improving VGG19 network |
CN110287574A (en) * | 2019-06-19 | 2019-09-27 | 浙江农林大学 | A kind of multichannel grain storage pest fumigant diffusion rate and insecticide efficiency assess device |
CN110287574B (en) * | 2019-06-19 | 2022-11-01 | 浙江农林大学 | Multichannel grain storage pest fumigant diffusion rate and insecticidal efficiency evaluation device |
CN110378435A (en) * | 2019-07-25 | 2019-10-25 | 安徽工业大学 | A method of the Apple Leaves disease recognition based on convolutional neural networks |
CN110378435B (en) * | 2019-07-25 | 2021-09-10 | 安徽工业大学 | Apple leaf disease identification method based on convolutional neural network |
CN110942401A (en) * | 2019-11-21 | 2020-03-31 | 黑龙江电力调度实业有限公司 | Intelligent communication method for power Internet of things |
CN110942401B (en) * | 2019-11-21 | 2023-12-19 | 黑龙江电力调度实业有限公司 | Intelligent communication method for electric power Internet of things |
CN111127423A (en) * | 2019-12-23 | 2020-05-08 | 金陵科技学院 | Rice pest and disease identification method based on CNN-BP neural network algorithm |
CN111428798A (en) * | 2020-03-30 | 2020-07-17 | 北京工业大学 | Plant seedling classification method based on convolutional neural network |
CN111504308A (en) * | 2020-04-21 | 2020-08-07 | 南京比特互动创意科技有限公司 | Intelligent exhibition hall supporting system based on AR technology |
CN111611889A (en) * | 2020-05-12 | 2020-09-01 | 安徽大学 | Miniature insect pest recognition device in farmland based on improved convolutional neural network |
CN111611889B (en) * | 2020-05-12 | 2023-04-18 | 安徽大学 | Miniature insect pest recognition device in farmland based on improved convolutional neural network |
CN112529043A (en) * | 2020-11-19 | 2021-03-19 | 江苏科技大学 | MDRN network-based online grain storage pest identification system and method |
CN112598265B (en) * | 2020-12-18 | 2022-06-07 | 武汉大学 | Decoupling risk estimation-based rapid detection method for hyperspectral pine nematode disease of unmanned aerial vehicle |
CN112598265A (en) * | 2020-12-18 | 2021-04-02 | 武汉大学 | Decoupling risk estimation-based rapid detection method for hyperspectral pine nematode disease of unmanned aerial vehicle |
CN114092799A (en) * | 2021-10-28 | 2022-02-25 | 长春工业大学 | Forestry pest identification and detection method based on pooling vision Transformer |
CN114494678A (en) * | 2021-12-02 | 2022-05-13 | 国家计算机网络与信息安全管理中心 | Character recognition method and electronic equipment |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109558787A (en) | A kind of Bamboo insect pests recognition methods based on convolutional neural networks model | |
CN110309841A (en) | A kind of hickory nut common insect pests recognition methods based on deep learning | |
CN102013021B (en) | Tea tender shoot segmentation and identification method based on color and region growth | |
CN110210555A (en) | Rail fish scale hurt detection method based on deep learning | |
CN107133960A (en) | Image crack dividing method based on depth convolutional neural networks | |
CN109191455A (en) | A kind of field crop pest and disease disasters detection method based on SSD convolutional network | |
CN109636026A (en) | A kind of wafer yield prediction technique based on deep learning model | |
CN107392130A (en) | Classification of Multispectral Images method based on threshold adaptive and convolutional neural networks | |
CN111696101A (en) | Light-weight solanaceae disease identification method based on SE-Inception | |
CN109344699A (en) | Winter jujube disease recognition method based on depth of seam division convolutional neural networks | |
CN111127423B (en) | Rice pest and disease identification method based on CNN-BP neural network algorithm | |
CN106845497A (en) | Maize in Earlier Stage image damage caused by a drought recognition methods based on multi-feature fusion | |
CN110287777A (en) | A kind of golden monkey body partitioning algorithm under natural scene | |
CN111046793B (en) | Tomato disease identification method based on deep convolutional neural network | |
CN107392251A (en) | A kind of method that target detection network performance is lifted using category images | |
CN107491733A (en) | A kind of chrysanthemum recognition methods based on deep neural network | |
CN114693616A (en) | Rice disease detection method, equipment and medium based on improved target detection model and convolutional neural network | |
Yang et al. | A comparative evaluation of convolutional neural networks, training image sizes, and deep learning optimizers for weed detection in alfalfa | |
CN114266965A (en) | Bitter orange pest and disease identification method based on multitask learning | |
CN114092799A (en) | Forestry pest identification and detection method based on pooling vision Transformer | |
Saputro et al. | Comparison ADAM-optimizer and SGDM for Classification Images of Rice Leaf Disease | |
CN113989536A (en) | Tomato disease identification method based on cuckoo search algorithm | |
CN116310338A (en) | Single litchi red leaf tip segmentation method based on examples and semantic segmentation | |
CN115049920A (en) | GEE combined growth period attribute and object-oriented large-area crop extraction method | |
Ren et al. | Recognition of common pests in agriculture and forestry based on convolutional neural networks |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190402 |
|
RJ01 | Rejection of invention patent application after publication |