CN109508650A - A kind of wood recognition method based on transfer learning - Google Patents
A kind of wood recognition method based on transfer learning Download PDFInfo
- Publication number
- CN109508650A CN109508650A CN201811234969.XA CN201811234969A CN109508650A CN 109508650 A CN109508650 A CN 109508650A CN 201811234969 A CN201811234969 A CN 201811234969A CN 109508650 A CN109508650 A CN 109508650A
- Authority
- CN
- China
- Prior art keywords
- training
- model
- picture
- image
- layer
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The wood recognition method based on transfer learning that the invention discloses a kind of, comprising the following steps: S1 collects Tree image, makes tree species image data set；S2 carries out data enhancing to raw data set image, to expand the quantity of picture；S3 obtains more than one pre-training models based on convolutional neural networks trained on large-scale image data set, S4 is trained pre-training model with tree species image data set, optimize one or several full articulamentums in pre-training model in the training process, to train multiple classifiers based on convolutional neural networks；The accuracy rate for testing each classifier selects accuracy rate highest classifier；S5 carries out wood recognition with the classifier that step S4 is selected, to obtain the result of identification.The present invention is based on the modes of transfer learning to identify to tree species, so that can also train the good model of generalization ability in the limited situation of tree species sample size, and substantially increases the wood recognition rate under complex background.
Description
Technical field
The present invention relates to a kind of wood recognition methods, are specifically related to a kind of wood recognition method based on transfer learning.
Background technique
Tree is one kind of plant, refers to xylophyta.Newest research indicates, there is 60065 kinds of trees in the world, and
This number is also being continuously updated, because all can newly name about 2000 kinds of trees every year.Type is very various, wherein
Chinese about 8000 kinds of trees.In face of a lot of tree species information, depends merely on and manually go identification clearly unreasonable, no
But arduously it is also possible to the inaccuracy of identification.And researcher is needed to have comparable experience by human brain identification.It is therefore desirable to mention
A kind of method out can automatically identify the type of the tree by the image of trees.This construction to Forestry informationlization, intelligence
Certain positive effect is built up in intelligent forestry.
Wood recognition technology, it will usually tree species be identified according to the image of trees.Zhang Shuai et al. is " based on layering volume
Product deep learning system plant leaf blade Study of recognition " in provide a kind of wood recognition technology, by leaf image into
Then the identification of row tree species is recycled first with convolutional neural networks autonomous learning leaf color, shape, Texture eigenvalue
The classifiers such as SVM, Softmax classify to plant.But this method needs the more demanding of leaf image data set
Each picture is pre-processed, blade-section is separated from background.Therefore, blade data are substantially increased
Collect the difficulty obtained.Also, this mode is low to the trees discrimination under complex background；It is additionally based on depth convolutional Neural net
The plants identification method of network depends on data-driven, it is desirable that the enough effects being likely to of trained data more.
Notification number is the Chinese patent " image-recognizing method based on deep learning and transfer learning " of CN106991439A,
This method comprises the following steps: one, the preparation stage: reading pre-training model, and reads picture directory, divides training set, verifying
Collection and test set；Two, the training stage: full Connection Neural Network classifier is constructed, and using pictures as the pre-training mould
The input of type updates the full Connection Neural Network classifier using the output with the pre-training model；Three, memory phase:
Storage model result.This method provide based on deep learning with the image-recognizing method of transfer learning ining conjunction with deep learning and
The application of transfer learning, to be provided relatively on the basis of extremely limited training time, training samples number for user
Accurate image recognition result.But this method still has certain defect, this method is in the case where limited sample size
Training classifier, obtained classifier is used directly to carry out image recognition, it cannot be guaranteed that the standard of the reliability of classifier and identification
True property.
Summary of the invention
Goal of the invention: in order to overcome the deficiencies in the prior art, the present invention provides a kind of tree based on transfer learning
Kind recognition methods, the image that this method passes through the trees entirety under nature photographed scene carry out wood recognition, and image data set obtains
It takes conveniently；And using the thought of transfer learning, even if the situation limited in training sample, training the model come also has very
Good generalization ability, substantially increases the wood recognition rate under complex background.
Technical solution: to achieve the above object, a kind of wood recognition method based on transfer learning of the invention, including with
Lower step:
S1 collects Tree image, makes tree species image data set；
S2 carries out data enhancing to raw data set image, to expand the quantity of picture；
S3 obtains more than one pre-training models based on convolutional neural networks trained on large-scale image data set,
S4 is trained pre-training model with tree species image data set, optimizes in pre-training model in the training process
One or several full articulamentums, to train multiple classifiers based on convolutional neural networks；Test the standard of each classifier
True rate selects accuracy rate highest classifier；
S5 carries out wood recognition with the classifier that step S4 is selected, to obtain the result of identification.
Further, in the step S1, the production method of tree species image data set are as follows: direct by artificial mode
It is shot in natural scene or the image data for crawling related tree species of batch forms picture number on network by crawlers
According to collection.Then classify according to trees to image, gather the image of tree of the same race as one, with the title of each tree species
As the label of corresponding set, data cleansing is then carried out, filters out image and the unmatched data of label, and not by these
Matched data are deleted.
Further, in the step S2, data enhance concrete operations are as follows: pass through the every figure concentrated to initial data
Piece is control/to spin upside down or adjust general brightness, and contrast, the mode of saturation degree convert picture, after transformation
Picture save as new picture, and be stored in corresponding image collection, to expand the quantity of picture.
Further, in the step S3,3 kinds is obtained and trains the base come on ImageNet large size image data set
In the pre-training model of convolutional neural networks.
Further, the step S4 specifically: first by tree species image data set be divided into training set, verifying collection and
The picture of training set is input in pre-training model by test set, and every picture is by each convolutional layer in pre-training model
With multiple characteristic patterns are obtained after the layer of pond, these characteristic patterns obtain prediction result using each full articulamentum, then with intersect
Gap between entropy function evaluation and foreca result and true value all repeats the above steps to every picture in training set, at this
During a, update is optimized with one or several full link layer parameters of the gradient descent method to model every time, it is all
Training picture sample all inputted the update that the network is primary network, and after each network updates, it is accurate to go to calculate with verifying collection
Rate selects the highest model of accuracy rate and obtains a classifier, with similar method, selects different pre-training models, updates
Different full articulamentums obtains multiple classifiers, then the accuracy rate of each classifier is verified with test set, and it is accurate then to select
The highest classifier of rate.
It further, is first fixed size, every picture by the modification of dimension of all pictures when being trained to model
The calculating of multiple convolutional layers, pond layer, full articulamentum, the formula of convolution layer operation can be passed through after being input in pre-training model
It is as follows:
Wherein, LiIndicate i-th layer of convolutional neural networks of characteristic pattern；WiIndicate i-th layer of convolution kernel, WiIt is a tensor；
biIndicate i-th layer of bias vector；Indicate convolution algorithm, convolution is the knot summed after two tensors are multiplied within the scope of certain
Fruit；Act (x) indicates activation primitive；
The whole process of convolution layer operation is (i-1)-th layer of output image and i-th layer of convolution kernel does convolution operation, then plus
Upper bias matrix bi, then calculated by activation primitive, finally obtain i-th layer of characteristic pattern Li；
What pond layer was done is the range in every n × n of image, take out n × n within the scope of it is maximum one value, and by this
A range is translated according to sequence from left to right, from top to bottom on the image, is finally taken out all values and is formed new picture；
The calculation formula of full articulamentum is as follows:
Z=Xi×Wi+bi
In formula, XiIt is characterized each pixel value of figure, WiFor weight parameter, biFor bias, Z is characterized each of figure
Calculated result of a pixel value in full articulamentum；
The last layer of full articulamentum uses softmax function normalization, exports the vector of a n row, and the value of n is tree
The species number of kind.The formula of Softmax function is as follows:
Wherein, xiRepresent the value of i-th of neuron of output layer；
Vector, that is, model output valve of the n row of output, the current line table of the value representative model prediction of every a line in vector
The probability value for the tree species shown, probability value is bigger, and the result for representing prediction is more biased to the tree species.
Further, in the training process, one picture of every input is all evaluated with intersection entropy function after obtaining prediction result
Gap between prediction result and true value, and with gradient descent method to one or several full link layer parameters of model into
Row optimization updates,
The formula for intersecting entropy function is as follows:
Wherein, m indicates that the sample number of primary training input, n represent n kind classification, yjiIndicate true label,It represents
The label of prediction.
The formula of gradient descent method is as follows:
In formula, θ: indicate that updated parameter value, θ indicate the parameter that model needs to update, α indicates learning rate, J (θ) table
Show loss function；
Specific training process are as follows: random initializtion Wi, biAnd the value of learning rate α, constantly more using gradient descent method
W in new one or several full articulamentumsi, biValue, to reduce loss value, all training picture samples all inputted this
Network is the update of primary network, after each network updates, goes to calculate accuracy rate with verifying collection, have updated i times when network or
When loss value is lower than j, that is, stop the training of network, select and obtain a classifier in the upper highest model of accuracy rate of verifying collection,
It repeats the above process, different full articulamentums is selected to be updated, then available multiple classifiers are verified with test set
The accuracy rate of each classifier selects the highest classifier of accuracy rate.
Further, activation primitive used in convolutional layer is ReLu activation primitive, and formula is as follows:
A=max (0, z)
In formula, z is representedValue.
Further, available by updating different full articulamentumsA classifier, wherein n indicates pre-
The number of training pattern, kiIndicate the number of plies of each pre-training model.
Further, the calculation formula of accuracy rate is as follows:
Wherein accuracy is accuracy rate, and m is the total number of the image of test, and n is first five result of the model prediction
In any one correct picture number.
The utility model has the advantages that
The present invention compared with the prior art, has the advantage, that
1, using method of the invention, the image based on trees entirety carries out the identification of tree species.It does not need to image data
Collection carries out additional pretreatment, greatly reduces the acquisition difficulty of image data set.And the identification method based on general image
More features can be provided for neural network learning；
2, using method of the invention, the mode based on transfer learning identifies tree species, so that in tree species sample size
Also the good model of generalization ability can be trained in limited situation, and substantially increases the tree species under complex background
Discrimination.
3, method of the invention, by overturning picture, adjustment picture luminance, the modes such as contrast to the quantity of picture into
Row expands, to increase sample size, increases trained number, so that model be made preferably to be optimized, improves mould
The reliability and accuracy rate of type.
4, method of the invention, from network obtain trained on large-scale image data set based on convolutional neural networks
Pre-training model, then pre-training model is trained with the image in tree species image set, is passed through during training
Intersect the accuracy that entropy function carrys out assessment models calculating, and constantly passes through the one or several of gradient descent method more new model
Parameter in full articulamentum, optimizes convolutional neural networks, so that it is guaranteed that convolutional neural networks can accurately identify the kind of trees
Class.
5, method of the invention, training obtains multiple and different classifiers by updating different full articulamentums every time, so
The accuracy rate of each classifier is tested afterwards, by comparing the accuracy rate of each classifier, selects the highest classifier of accuracy rate
Image recognition is carried out, ensure that the reliability of image recognition result.
6, method of the invention, model output result handled with softmax classifier obtain it is final as a result,
The probability that picture to be measured belongs to each tree species is calculated by softmax function, last output result is that probability value is maximum
5 tree species, further improve the accuracy rate of identification, and it is clear to export result.
Detailed description of the invention
Fig. 1 is the wood recognition method flow diagram the present invention is based on transfer learning.
Fig. 2 is 10 kinds of tree experiment sample graphs of the embodiment of the present invention.
Fig. 3 is the embodiment of the present invention to one of enhanced effect picture of experiment sample data.
Fig. 4 is the Technology Roadmap in embodiment nine.
Fig. 5 is result figure of the embodiment of the present invention to wood recognition.
Specific embodiment
The present invention will be further explained with reference to the accompanying drawing.
Embodiment one
A kind of wood recognition method based on transfer learning of the present embodiment, comprising the following steps:
S1 collects Tree image, makes tree species image data set；
S2 carries out data enhancing to raw data set image, to expand the quantity of picture；
S3 obtains more than one pre-training models based on convolutional neural networks trained on large-scale image data set,
S4 is trained pre-training model with tree species image data set, optimizes in pre-training model in the training process
One or several full articulamentums, to train multiple classifiers based on convolutional neural networks；Test the standard of each classifier
True rate selects accuracy rate highest classifier；
S5 carries out wood recognition with the classifier that step S4 is selected, to obtain the result of identification.
Embodiment two
A kind of wood recognition method based on transfer learning of the present embodiment is based on embodiment one, tree species image data set
Production method are as follows: by artificial mode directly natural scene shoot or by crawlers on network batch
Crawl the image data composition image data set of related tree species.Then classify according to trees to image, by tree of the same race
Image is gathered as one, and the label of corresponding set is referred to as with the name of each tree species, data cleansing is then carried out, filters out figure
Picture and the unmatched data of label, and these unmatched data are deleted.
Embodiment three
A kind of wood recognition method based on transfer learning of the present embodiment is based on embodiment two, the specific behaviour of data enhancing
As: general brightness is controlled/spun upside down or adjust by the every picture concentrated to initial data, and contrast is satisfied
Picture is converted with the mode of degree, transformed picture saves as new picture, and is stored in corresponding image collection
In, to expand the quantity of picture.
Example IV
A kind of wood recognition method based on transfer learning of the present embodiment is wanted before being trained based on embodiment three
Obtain more than one pre-training models based on convolutional neural networks trained on large-scale image data set, pre-training mould
Type can be selected from AlexNet, VggNet, the models such as Inception V1~V4, ResNet, in this example using 3 kinds
The pre-training model based on convolutional neural networks come is trained on ImageNet large size image data set, respectively
AlexNet, VggNet and Inception V3.
Embodiment five
A kind of wood recognition method based on transfer learning of the present embodiment is based on example IV, with tree species image data
Collect the method that is trained to pre-training model specifically: first by tree species image data set be divided into training set, verifying collection and
The picture of training set is input in pre-training model by test set, and every picture is by each convolutional layer in pre-training model
With multiple characteristic patterns are obtained after the layer of pond, these characteristic patterns obtain prediction result using each full articulamentum, then with intersect
Gap between entropy function evaluation and foreca result and true value all repeats the above steps to every picture in training set, at this
During a, update is optimized with one or several full link layer parameters of the gradient descent method to model every time, it is all
Training picture sample all inputted the update that the network is primary network, and after each network updates, it is accurate to go to calculate with verifying collection
Rate selects the highest model of accuracy rate and obtains a classifier, with similar method, selects different pre-training models, updates
Different full articulamentums obtains multiple classifiers, then the accuracy rate of each classifier is verified with test set, and it is accurate then to select
The highest classifier of rate.
Embodiment six
A kind of wood recognition method based on transfer learning of the present embodiment is based on embodiment five, instructs to model
When practicing, tree species image data set is first divided into training set, verifying collection and test set, the ratio of general three be 8:1:1 or
Then the modification of dimension of all pictures is fixed size x × y pixel by 7:1.5:1.5, wherein the width of x representative image, y generation
The height of table image, the typically no limitation requirement of the size of x, y, for convenience subsequent calculating, the value of x, y keep one as far as possible
It causes, every picture can pass through the calculating of multiple convolutional layers, pond layer, full articulamentum, convolutional layer after being input in pre-training model
The formula of operation is as follows:
Wherein, LiIndicate i-th layer of convolutional neural networks of characteristic pattern；WiIndicate i-th layer of convolution kernel, WiIt is a tensor；
biIndicate i-th layer of bias vector；Indicate convolution algorithm, convolution is the knot summed after two tensors are multiplied within the scope of certain
Fruit；Act (x) indicates activation primitive；
Activation primitive used in convolutional layer is ReLu activation primitive, and formula is as follows:
A=max (0, z)
In formula, z is representedValue.
The whole process of convolution layer operation is (i-1)-th layer of output image and i-th layer of convolution kernel does convolution operation, then plus
Upper bias matrix bi, then calculated by activation primitive, finally obtain i-th layer of characteristic pattern Li；
What pond layer was done is the range in every n × n of image, take out n × n within the scope of it is maximum one value, and by this
A range is translated according to sequence from left to right, from top to bottom on the image, is finally taken out all values and is formed new picture；
Due to the transfer learning that this method uses, so network need not be from the beginning trained, convolutional layer, pond before pre-training model
The parameter for changing layer is obtained in ImageNet data concentration training, and original image is passed through convolutional neural networks convolutional layer, pond
Change the forward calculation of layer, it can pull out characteristic pattern, these characteristic patterns are often more than the feature of Feature Engineering manual extraction
It is reliable.Context of methods using pull out come characteristic pattern, only with the trained full articulamentum of oneself tree species image data set,
The calculation formula of full articulamentum is as follows:
Z=Xi×Wi+bi
In formula, XiIt is characterized each pixel value of figure, WiFor weight parameter, biFor bias, Z is characterized each of figure
Calculated result of a pixel value in full articulamentum；
The last layer of full articulamentum uses softmax function normalization, exports the vector of a n row, and the value of n is tree
The species number of kind.The formula of Softmax function is as follows:
Wherein, xiRepresent the value of i-th of neuron of output layer；
Vector, that is, model output valve of the n row of output, the current line table of the value representative model prediction of every a line in vector
The probability value for the tree species shown, probability value is bigger, and the result for representing prediction is more biased to the tree species.
Embodiment seven
A kind of wood recognition method based on transfer learning of the present embodiment is based on embodiment six, in the training process, often
It inputs after a picture obtains prediction result and all uses the gap intersected between entropy function evaluation and foreca result and true value, and
Update is optimized with one or several full link layer parameters of the gradient descent method to model,
The formula for intersecting entropy function is as follows:
Wherein, m indicates that the sample number of primary training input, n represent n kind classification, yjiIndicate true label,It represents
The label of prediction.
The formula of gradient descent method is as follows:
In formula, θ: indicate that updated parameter value, θ indicate the parameter that model needs to update, α indicates learning rate, J (θ) table
Show loss function；
Specific training process are as follows: random initializtion Wi, biAnd the value of learning rate α, constantly more using gradient descent method
W in new one or several full articulamentumsi, biValue, to reduce loss value, all training picture samples all inputted this
Network is the update of primary network, after each network updates, goes to calculate accuracy rate with verifying collection, have updated i times when network or
When loss value is lower than j, that is, stop the training of network, the value of i is determined according to data volume, and data volume is more, and the value of i also just needs
If it is bigger, the i in this example is set as 10000.The smaller effect for representing network convergence of Loss value is better, and the value of j is set as in this example
0.001, it selects and collects the upper highest model of accuracy rate in verifying and obtain a classifier, repeat the above process, select different complete
Articulamentum is updated, then available multiple classifiers verify the accuracy rate of each classifier with test set, select standard
The highest classifier of true rate.
The calculation formula of accuracy rate is as follows:
Wherein accuracy is accuracy rate, and m is the total number of the image of test, and n is first five result of the model prediction
In any one correct picture number.
Embodiment eight
A kind of wood recognition method based on transfer learning of the present embodiment is based on embodiment seven, different by updating
Full articulamentum is availableA classifier, wherein n indicates the number of pre-training model, kiIndicate each pre-training
The number of plies of model.
Method in this example by transfer learning has trained following 8 classifiers altogether:
Wherein, AlexNet-fc8 " indicates the fc8 layer of only training AlexNet model, and " AlexNet-fc8, fc7 " are indicated
Fc8 and fc7 layers of training AlexNet model." AlexNet-fc8, fc7, fc6 " indicate the fc8 of training AlexNet model,
Three layers behind fc7 and fc6.And so on, VggNet-16 model indicates meaning with AlexNet model."Inception-
V3 " indicates the last layer of only training Inception-V3 model." Inception-V3-a1 " is indicated in Inception-V3
Increase a full articulamentum below, then two layers below of training.
Embodiment nine
A kind of wood recognition method based on transfer learning of the present embodiment, referring to Fig.1, comprising the following steps:
Make tree species image data set.Production method are as follows: a part is directly shot in natural scene by artificial mode
It obtaining, another part is obtained on network by web crawlers technology, finally again by manually these Data Integrations get up, into
Row data cleansing, to obtain the tree species data set for being suitble to training；Experimental data sample in this example include: american beech,
Button ball, black walnut, Dong Fanghongshan, ginkgo, Acer palmatum ' Atropurpureum' tree, southern yulan, tulip white poplar, white oak and kahikatea totally 10
1593 images of a tree species, wherein american beech 104 is opened, button ball 130 is opened, black walnut 154 is opened, Dong Fanghongshan 102
It opens, ginkgo 270 opens, 183, Acer palmatum ' Atropurpureum' tree, yulan 175 opens, tulip white poplar 133 opens, white oak 202 opens, kahikatea 140 in south
, as shown in Fig. 2, wherein every kind of tree only has chosen a picture as representative；
Data enhancing is carried out to raw data set image, using following methods progress data enhancing: first, for figure
As color carries out data enhancing, comprising: brightness of image, saturation degree, contrast etc.；Second, water is carried out to original image
Flat/flip vertical；Third zooms in and out picture using modes such as arest neighbors difference, bilinearity difference, bicubic differences, so
The new picture of five Area generations of random cropping picture afterwards.The data of above-mentioned several ways are realized by Python programming language
Enhancing.Experiment effect figure is as shown in figure 3, only show here by an enhanced effect of picture data；
Download based on ImageNet image data set train come pre-training model, the resource on network is very rich
Richness, such as:https://github.com/ tensorflow/models/tree/master/research/slim, this reality
Apply example selection Inception_v3 pre-training model.
Model downloading finishes, so that it may write program based on this model.In the embodiment of the present invention neural network build and
Operation is all to call TensorFlow deep learning library by Python programming language at windows10 to implement to complete.
The specific implementation steps are as follows:
1. carrying out the design of program UI using WebAPP Development Framework MUI and 5+Runtime.
2. database uses the MySQL of open source community version, it is mainly used for storing the image data that user uploads.
3. server-side uses PHP program language, it is mainly used for executing business operation.Specific logic is that user is by taking pictures
Or photograph album chooses catalogue of the mode uploading pictures of picture to server, the mould trained using Python based on front
Type writes identification script, identifies the picture under above-mentioned catalogue, finally calls Python script to obtain recognition result by PHP.
Technology path is as shown in Figure 4:
The initialization learning rate of model is set, suitable optimization algorithm is selected, certain iterative steps are finally set.It can
To open having trained for model.One time model training is completed, and calculates accuracy rate of the model on test set.It then proceedes to
Model initialization parameter is adjusted, second of training is opened.So circulation, finally comparison obtains the best model of an effect.This
Learning rate is set as 0.001 to being provided that when obtaining highest accuracy rate in inventive embodiments, uses Adam optimizer iteration
100000 times；
The identification of tree species is carried out using trained model, eventually exports the recognition result of top-5.As shown in figure 5,
The left side is a ginkgo tree Image of shooting, and the right is the wood recognition obtained based on the method for the present invention as a result, top-5 is identified
As a result the highest scoring of middle ginkgo, identification are correct.
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 (10)
1. a kind of wood recognition method based on transfer learning, which comprises the following steps:
S1 collects Tree image, makes tree species image data set；
S2 carries out data enhancing to raw data set image, to expand the quantity of picture；
S3 obtains more than one pre-training models based on convolutional neural networks trained on large-scale image data set,
S4 is trained pre-training model with tree species image data set, optimizes one in pre-training model in the training process
Or several full articulamentums, to train multiple classifiers based on convolutional neural networks；The accuracy rate of each classifier is tested,
Select accuracy rate highest classifier；
S5 carries out wood recognition with the classifier that step S4 is selected, to obtain the result of identification.
2. the wood recognition method according to claim 1 based on transfer learning, which is characterized in that in the step S1,
The production method of tree species image data set are as follows: directly shot in natural scene by artificial mode or existed by crawlers
The image data composition image data set for crawling related tree species of batch on network.Then classify according to trees to image,
Gather the image of tree of the same race as one, the label of corresponding set is referred to as with the name of each tree species, it is clear then to carry out data
It washes, filters out image and the unmatched data of label, and these unmatched data are deleted.
3. the wood recognition method according to claim 1 based on transfer learning, which is characterized in that in the step S2,
Data enhance concrete operations are as follows: by the every picture concentrated to initial data to control/spin upside down or adjust general
Brightness, contrast, the mode of saturation degree convert picture, and transformed picture saves as new picture, and is stored in phase
In the image collection answered, to expand the quantity of picture.
4. the wood recognition method according to claim 1 based on transfer learning, which is characterized in that in the step S3,
It obtains 3 kinds and trains the pre-training model based on convolutional neural networks come on ImageNet large size image data set.
5. the wood recognition method according to claim 1 based on transfer learning, which is characterized in that the step S4 is specific
Are as follows: tree species image data set is divided into training set, verifying collection and test set first, the picture of training set is input to pre-training
In model, every picture obtains multiple characteristic patterns after each convolutional layer and pond layer in pre-training model, these features
Figure obtains prediction result using each full articulamentum, then with the difference intersected between entropy function evaluation and foreca result and true value
Away from all repeating the above steps to every picture in training set, in this process, every time with gradient descent method to model
One or several full link layer parameters optimize update, and it is primary network that all training picture samples, which all inputted the network,
Update, after each network updates, go to calculate accuracy rate with verifying collection, select the highest model of accuracy rate and obtain a classification
Device selects different pre-training models with similar method, updates different full articulamentums, obtains multiple classifiers, then with surveying
Examination collection verifies the accuracy rate of each classifier, then selects the highest classifier of accuracy rate.
6. the wood recognition method according to claim 5 based on transfer learning, which is characterized in that instructed to model
It is first fixed size by the modification of dimension of all pictures, every picture can pass through multiple after being input in pre-training model when practicing
The formula of the calculating of convolutional layer, pond layer, full articulamentum, convolution layer operation is as follows:
Wherein, LiIndicate i-th layer of convolutional neural networks of characteristic pattern；WiIndicate i-th layer of convolution kernel, WiIt is a tensor；biIt indicates
I-th layer of bias vector；Indicate convolution algorithm, convolution is the result summed after two tensors are multiplied within the scope of certain；act
(x) activation primitive is indicated；
The whole process of convolution layer operation is that (i-1)-th layer of output image and i-th layer of convolution kernel do convolution operation, along with inclined
Set matrix bi, then calculated by activation primitive, finally obtain i-th layer of characteristic pattern Li；
What pond layer was done is the range in every n × n of image, takes out a maximum value within the scope of n × n, and by this model
It encloses and is translated on the image according to sequence from left to right, from top to bottom, finally take out all values and form new picture；
The calculation formula of full articulamentum is as follows:
Z=Xi×Wi+bi
In formula, XiIt is characterized each pixel value of figure, WiFor weight parameter, biFor bias, Z is characterized each picture of figure
Calculated result of the element value in full articulamentum；
The last layer of full articulamentum uses softmax function normalization, exports the vector of a n row, and the value of n is the kind of tree species
Class number.The formula of Softmax function is as follows:
Wherein, xiRepresent the value of i-th of neuron of output layer；
Vector, that is, model output valve of the n row of output, the current line of the value representative model prediction of every a line indicates in vector
The probability value of tree species, probability value is bigger, and the result for representing prediction is more biased to the tree species.
7. the wood recognition method according to claim 6 based on transfer learning, which is characterized in that in the training process,
One picture of every input all uses the gap intersected between entropy function evaluation and foreca result and true value after obtaining prediction result, and
Update is optimized with one or several full link layer parameters of the gradient descent method to model,
The formula for intersecting entropy function is as follows:
Wherein, m indicates that the sample number of primary training input, n represent n kind classification, yjiIndicate true label,Represent prediction
Label.
The formula of gradient descent method is as follows:
In formula, θ: indicate that updated parameter value, θ indicate the parameter that model needs to update, α indicates that learning rate, J (θ) indicate damage
Lose function；
Specific training process are as follows: random initializtion Wi, biAnd the value of learning rate α, one is constantly updated using gradient descent method
W in a or several full articulamentumsi, biValue, to reduce loss value, all training picture samples all inputted the network
It for the update of primary network, after each network updates, goes to calculate accuracy rate with verifying collection, when network has updated i times or loss
When value is lower than j, that is, stop the training of network, selects and obtain a classifier in the upper highest model of accuracy rate of verifying collection, repeat
The above process selects different full articulamentums to be updated, then available multiple classifiers verify each point with test set
The accuracy rate of class device selects the highest classifier of accuracy rate.
8. the wood recognition method according to claim 6 based on transfer learning, which is characterized in that used in convolutional layer
Activation primitive is ReLu activation primitive, and formula is as follows:
A=max (0, z)
In formula, z is representedValue.
9. the wood recognition method according to claim 7 based on transfer learning, which is characterized in that different by updating
Full articulamentum is availableA classifier, wherein n indicates the number of pre-training model, kiIndicate each pre-training
The number of plies of model.
10. the wood recognition method according to claim 7 or 9 based on transfer learning, which is characterized in that the meter of accuracy rate
It is as follows to calculate formula:
Wherein accuracy is accuracy rate, and m is the total number of the image of test, and n is to appoint in first five result of the model prediction
One correct picture number of meaning.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811234969.XA CN109508650A (en) | 2018-10-23 | 2018-10-23 | A kind of wood recognition method based on transfer learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811234969.XA CN109508650A (en) | 2018-10-23 | 2018-10-23 | A kind of wood recognition method based on transfer learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109508650A true CN109508650A (en) | 2019-03-22 |
Family
ID=65746076
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811234969.XA Pending CN109508650A (en) | 2018-10-23 | 2018-10-23 | A kind of wood recognition method based on transfer learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109508650A (en) |
Cited By (49)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110033045A (en) * | 2019-04-17 | 2019-07-19 | 内蒙古工业大学 | A kind of method and apparatus of trained identification image atomization |
| CN110147772A (en) * | 2019-05-23 | 2019-08-20 | 河海大学常州校区 | A kind of underwater dam surface crack recognition methods based on transfer learning |
| CN110263735A (en) * | 2019-06-25 | 2019-09-20 | 北京林业大学 | A method of tree species classification being carried out to artificial forest high-spectral data using Three dimensional convolution neural network |
| CN110263863A (en) * | 2019-06-24 | 2019-09-20 | 南京农业大学 | Fine granularity mushroom phenotype recognition methods based on transfer learning Yu bilinearity InceptionResNetV2 |
| CN110363249A (en) * | 2019-07-22 | 2019-10-22 | 西南林业大学 | A kind of category of annatto recognition methods neural network based |
| CN110390341A (en) * | 2019-07-26 | 2019-10-29 | 长安大学 | Green based on convolutional neural networks, which is open to traffic, delivers cargo radioactive source image-recognizing method |
| CN110458043A (en) * | 2019-07-20 | 2019-11-15 | 中国船舶重工集团公司第七二四研究所 | A kind of SAR target identification method based on transfer learning and the output of full articulamentum |
| CN110458109A (en) * | 2019-08-13 | 2019-11-15 | 西南林业大学 | A kind of tealeaves disease recognition system and working method based on image recognition technology |
| CN110503051A (en) * | 2019-08-27 | 2019-11-26 | 西南林业大学 | A kind of precious timber identifying system and method based on image recognition technology |
| CN110503154A (en) * | 2019-08-27 | 2019-11-26 | 携程计算机技术（上海）有限公司 | Method, system, electronic equipment and the storage medium of image classification |
| CN110619059A (en) * | 2019-08-13 | 2019-12-27 | 浙江工业大学 | Building marking method based on transfer learning |
| CN110781921A (en) * | 2019-09-25 | 2020-02-11 | 浙江农林大学 | Depth residual error network and transfer learning-based muscarinic image identification method and device |
| CN110795623A (en) * | 2019-10-10 | 2020-02-14 | 深圳供电局有限公司 | Image enhancement training method and system and computer readable storage medium |
| CN110929722A (en) * | 2019-11-04 | 2020-03-27 | 浙江农林大学 | Tree detection method based on whole tree image |
| CN110991511A (en) * | 2019-11-26 | 2020-04-10 | 中原工学院 | Sunflower crop seed sorting method based on deep convolutional neural network |
| CN111242063A (en) * | 2020-01-17 | 2020-06-05 | 江苏大学 | Small sample classification model construction method based on transfer learning and iris classification application |
| CN111273767A (en) * | 2020-01-13 | 2020-06-12 | 重庆邮电大学 | Hearing-aid brain computer interface system based on deep migration learning |
| CN111382787A (en) * | 2020-03-06 | 2020-07-07 | 芯薇(上海)智能科技有限公司 | Target detection method based on deep learning |
| CN111462068A (en) * | 2020-03-30 | 2020-07-28 | 华南理工大学 | Bolt and nut detection method based on transfer learning |
| CN111476283A (en) * | 2020-03-31 | 2020-07-31 | 上海海事大学 | Glaucoma fundus image identification method based on transfer learning |
| CN111611924A (en) * | 2020-05-21 | 2020-09-01 | 东北林业大学 | Mushroom identification method based on deep migration learning model |
| CN111611972A (en) * | 2020-06-01 | 2020-09-01 | 南京信息工程大学 | Crop leaf type identification method based on multi-view multi-task ensemble learning |
| CN111652326A (en) * | 2020-06-30 | 2020-09-11 | 华南农业大学 | Improved fruit maturity identification method and identification system based on MobileNet v2 network |
| CN111666984A (en) * | 2020-05-20 | 2020-09-15 | 海南电网有限责任公司电力科学研究院 | Intelligent overvoltage identification method based on transfer learning |
| CN111695640A (en) * | 2020-06-18 | 2020-09-22 | 南京信息职业技术学院 | Foundation cloud picture recognition model training method and foundation cloud picture recognition method |
| CN111723738A (en) * | 2020-06-19 | 2020-09-29 | 安徽工业大学 | Coal petrography group microscopic image classification method and system based on transfer learning |
| CN111738303A (en) * | 2020-05-28 | 2020-10-02 | 华南理工大学 | Long-tail distribution image identification method based on hierarchical learning |
| CN111739108A (en) * | 2020-06-12 | 2020-10-02 | 鞍钢集团矿业有限公司 | Iron ore powder grade rapid estimation method based on convolutional neural network |
| CN111784639A (en) * | 2020-06-05 | 2020-10-16 | 浙江大学 | Oral panoramic film dental caries depth identification method based on deep learning |
| CN111783802A (en) * | 2020-07-21 | 2020-10-16 | 南昌航空大学 | Metal fracture image identification method based on deep learning |
| CN112036472A (en) * | 2020-08-28 | 2020-12-04 | 长安大学 | Visual image classification method and system for power system |
| CN112036495A (en) * | 2020-09-02 | 2020-12-04 | 常熟理工学院 | Garment image classification method based on web crawler and transfer learning |
| CN112270402A (en) * | 2020-10-20 | 2021-01-26 | 山东派蒙机电技术有限公司 | Training method and system for tire wear identification model |
| CN112287839A (en) * | 2020-10-29 | 2021-01-29 | 广西科技大学 | SSD infrared image pedestrian detection method based on transfer learning |
| CN112364899A (en) * | 2020-10-27 | 2021-02-12 | 西安科技大学 | Abrasive grain ferrographic image intelligent identification method based on virtual image and transfer learning |
| CN112435237A (en) * | 2020-11-24 | 2021-03-02 | 山西三友和智慧信息技术股份有限公司 | Skin lesion segmentation method based on data enhancement and depth network |
| CN112541555A (en) * | 2020-12-22 | 2021-03-23 | 中国医学科学院北京协和医院 | Deep learning-based classifier model training method |
| CN112926679A (en) * | 2021-03-26 | 2021-06-08 | 江苏大学 | Method for enhancing background change resistance of classification model in intellectual property image recognition |
| CN113066053A (en) * | 2021-03-11 | 2021-07-02 | 紫东信息科技(苏州)有限公司 | Model migration-based duodenum self-training classification method and system |
| CN113298791A (en) * | 2021-05-31 | 2021-08-24 | 中电福富信息科技有限公司 | Image detection method of mixed cartoon based on deep learning |
| CN113536298A (en) * | 2021-06-11 | 2021-10-22 | 浙江工业大学 | Defense method for bias poisoning attack of deep learning model |
| CN113610121A (en) * | 2021-07-22 | 2021-11-05 | 哈尔滨工程大学 | Cross-domain task deep learning identification method |
| CN113706511A (en) * | 2021-08-31 | 2021-11-26 | 佛山市南海区广工大数控装备协同创新研究院 | Composite material damage detection method based on deep learning |
| WO2022134834A1 (en) * | 2020-12-23 | 2022-06-30 | 深圳壹账通智能科技有限公司 | Potential event predicting method, apparatus and device, and storage medium |
| CN115100522A (en) * | 2022-06-27 | 2022-09-23 | 南京邮电大学 | Intelligent garden aerial photography tree species self-adaptive identification method based on tree growth model |
| CN115147737A (en) * | 2022-06-07 | 2022-10-04 | 滁州学院 | City tree species classification method based on ConvNeXt transfer learning |
| CN115719430A (en) * | 2022-10-28 | 2023-02-28 | 河北舒隽科技有限公司 | Method for identifying male and female of Taixing chick |
| CN116310511A (en) * | 2023-02-13 | 2023-06-23 | 南京林业大学 | Solid wood floor color classification method based on self-learning |
| CN116912879A (en) * | 2023-06-02 | 2023-10-20 | 广州大学 | Ear brushing identity recognition method based on transfer learning |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106096605A (en) * | 2016-06-02 | 2016-11-09 | 史方 | A kind of image obscuring area detection method based on degree of depth study and device |
| CN107507197A (en) * | 2017-08-18 | 2017-12-22 | 东北大学 | A kind of pulmonary parenchyma extracting method based on clustering algorithm and convolutional neural networks |
| CN108038499A (en) * | 2017-12-06 | 2018-05-15 | 浙江大学 | A kind of seeds sorting technique and system based on deep learning |
| US9984455B1 (en) * | 2017-06-05 | 2018-05-29 | Hana Resources, Inc. | Organism growth prediction system using drone-captured images |
| CN108256426A (en) * | 2017-12-15 | 2018-07-06 | 安徽四创电子股份有限公司 | A kind of facial expression recognizing method based on convolutional neural networks |
| CN108363961A (en) * | 2018-01-24 | 2018-08-03 | 东南大学 | Bridge pad disease recognition method based on transfer learning between convolutional neural networks |
-
2018
- 2018-10-23 CN CN201811234969.XA patent/CN109508650A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106096605A (en) * | 2016-06-02 | 2016-11-09 | 史方 | A kind of image obscuring area detection method based on degree of depth study and device |
| US9984455B1 (en) * | 2017-06-05 | 2018-05-29 | Hana Resources, Inc. | Organism growth prediction system using drone-captured images |
| CN107507197A (en) * | 2017-08-18 | 2017-12-22 | 东北大学 | A kind of pulmonary parenchyma extracting method based on clustering algorithm and convolutional neural networks |
| CN108038499A (en) * | 2017-12-06 | 2018-05-15 | 浙江大学 | A kind of seeds sorting technique and system based on deep learning |
| CN108256426A (en) * | 2017-12-15 | 2018-07-06 | 安徽四创电子股份有限公司 | A kind of facial expression recognizing method based on convolutional neural networks |
| CN108363961A (en) * | 2018-01-24 | 2018-08-03 | 东南大学 | Bridge pad disease recognition method based on transfer learning between convolutional neural networks |
Non-Patent Citations (2)
| Title |
|---|
| 赵鹏超，戚大伟: "基于卷积神经网络和树叶纹理的树种识别研究", 《森林工程》 * |
| 陈慧岩: "《智能车辆理论与应用》", 31 July 2018 * |
Cited By (70)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110033045A (en) * | 2019-04-17 | 2019-07-19 | 内蒙古工业大学 | A kind of method and apparatus of trained identification image atomization |
| CN110147772A (en) * | 2019-05-23 | 2019-08-20 | 河海大学常州校区 | A kind of underwater dam surface crack recognition methods based on transfer learning |
| CN110147772B (en) * | 2019-05-23 | 2022-09-20 | 河海大学常州校区 | Migration learning-based underwater dam body surface crack identification method |
| CN110263863A (en) * | 2019-06-24 | 2019-09-20 | 南京农业大学 | Fine granularity mushroom phenotype recognition methods based on transfer learning Yu bilinearity InceptionResNetV2 |
| CN110263863B (en) * | 2019-06-24 | 2021-09-10 | 南京农业大学 | Fine-grained fungus phenotype identification method based on transfer learning and bilinear InceptionResNet V2 |
| CN110263735A (en) * | 2019-06-25 | 2019-09-20 | 北京林业大学 | A method of tree species classification being carried out to artificial forest high-spectral data using Three dimensional convolution neural network |
| CN110458043A (en) * | 2019-07-20 | 2019-11-15 | 中国船舶重工集团公司第七二四研究所 | A kind of SAR target identification method based on transfer learning and the output of full articulamentum |
| CN110363249A (en) * | 2019-07-22 | 2019-10-22 | 西南林业大学 | A kind of category of annatto recognition methods neural network based |
| CN110390341A (en) * | 2019-07-26 | 2019-10-29 | 长安大学 | Green based on convolutional neural networks, which is open to traffic, delivers cargo radioactive source image-recognizing method |
| CN110390341B (en) * | 2019-07-26 | 2022-12-06 | 长安大学 | Convolutional neural network-based image recognition method for radioactive source of green traffic vehicle carried goods |
| CN110619059B (en) * | 2019-08-13 | 2021-07-27 | 浙江工业大学 | Building marking method based on transfer learning |
| CN110619059A (en) * | 2019-08-13 | 2019-12-27 | 浙江工业大学 | Building marking method based on transfer learning |
| CN110458109A (en) * | 2019-08-13 | 2019-11-15 | 西南林业大学 | A kind of tealeaves disease recognition system and working method based on image recognition technology |
| CN110503154A (en) * | 2019-08-27 | 2019-11-26 | 携程计算机技术（上海）有限公司 | Method, system, electronic equipment and the storage medium of image classification |
| CN110503051A (en) * | 2019-08-27 | 2019-11-26 | 西南林业大学 | A kind of precious timber identifying system and method based on image recognition technology |
| CN110503051B (en) * | 2019-08-27 | 2023-04-07 | 西南林业大学 | Precious wood identification system and method based on image identification technology |
| CN110781921A (en) * | 2019-09-25 | 2020-02-11 | 浙江农林大学 | Depth residual error network and transfer learning-based muscarinic image identification method and device |
| CN110795623A (en) * | 2019-10-10 | 2020-02-14 | 深圳供电局有限公司 | Image enhancement training method and system and computer readable storage medium |
| CN110795623B (en) * | 2019-10-10 | 2022-10-14 | 深圳供电局有限公司 | Image enhancement training method and system and computer readable storage medium |
| CN110929722A (en) * | 2019-11-04 | 2020-03-27 | 浙江农林大学 | Tree detection method based on whole tree image |
| CN110991511A (en) * | 2019-11-26 | 2020-04-10 | 中原工学院 | Sunflower crop seed sorting method based on deep convolutional neural network |
| CN111273767A (en) * | 2020-01-13 | 2020-06-12 | 重庆邮电大学 | Hearing-aid brain computer interface system based on deep migration learning |
| CN111242063B (en) * | 2020-01-17 | 2023-08-25 | 江苏大学 | Small sample classification model construction method based on transfer learning and iris classification application |
| CN111242063A (en) * | 2020-01-17 | 2020-06-05 | 江苏大学 | Small sample classification model construction method based on transfer learning and iris classification application |
| CN111382787A (en) * | 2020-03-06 | 2020-07-07 | 芯薇(上海)智能科技有限公司 | Target detection method based on deep learning |
| CN111462068B (en) * | 2020-03-30 | 2023-03-21 | 华南理工大学 | Bolt and nut detection method based on transfer learning |
| CN111462068A (en) * | 2020-03-30 | 2020-07-28 | 华南理工大学 | Bolt and nut detection method based on transfer learning |
| CN111476283A (en) * | 2020-03-31 | 2020-07-31 | 上海海事大学 | Glaucoma fundus image identification method based on transfer learning |
| CN111666984B (en) * | 2020-05-20 | 2023-08-25 | 海南电网有限责任公司电力科学研究院 | Overvoltage intelligent identification method based on transfer learning |
| CN111666984A (en) * | 2020-05-20 | 2020-09-15 | 海南电网有限责任公司电力科学研究院 | Intelligent overvoltage identification method based on transfer learning |
| CN111611924A (en) * | 2020-05-21 | 2020-09-01 | 东北林业大学 | Mushroom identification method based on deep migration learning model |
| CN111611924B (en) * | 2020-05-21 | 2022-03-25 | 东北林业大学 | Mushroom identification method based on deep migration learning model |
| CN111738303B (en) * | 2020-05-28 | 2023-05-23 | 华南理工大学 | Long-tail distribution image recognition method based on hierarchical learning |
| CN111738303A (en) * | 2020-05-28 | 2020-10-02 | 华南理工大学 | Long-tail distribution image identification method based on hierarchical learning |
| CN111611972B (en) * | 2020-06-01 | 2024-01-05 | 南京信息工程大学 | Crop leaf type identification method based on multi-view multi-task integrated learning |
| CN111611972A (en) * | 2020-06-01 | 2020-09-01 | 南京信息工程大学 | Crop leaf type identification method based on multi-view multi-task ensemble learning |
| CN111784639A (en) * | 2020-06-05 | 2020-10-16 | 浙江大学 | Oral panoramic film dental caries depth identification method based on deep learning |
| CN111739108B (en) * | 2020-06-12 | 2023-09-15 | 鞍钢集团矿业有限公司 | Rapid iron ore powder grade estimation method based on convolutional neural network |
| CN111739108A (en) * | 2020-06-12 | 2020-10-02 | 鞍钢集团矿业有限公司 | Iron ore powder grade rapid estimation method based on convolutional neural network |
| CN111695640B (en) * | 2020-06-18 | 2024-04-09 | 南京信息职业技术学院 | Foundation cloud picture identification model training method and foundation cloud picture identification method |
| CN111695640A (en) * | 2020-06-18 | 2020-09-22 | 南京信息职业技术学院 | Foundation cloud picture recognition model training method and foundation cloud picture recognition method |
| CN111723738A (en) * | 2020-06-19 | 2020-09-29 | 安徽工业大学 | Coal petrography group microscopic image classification method and system based on transfer learning |
| CN111652326A (en) * | 2020-06-30 | 2020-09-11 | 华南农业大学 | Improved fruit maturity identification method and identification system based on MobileNet v2 network |
| CN111652326B (en) * | 2020-06-30 | 2023-04-28 | 华南农业大学 | Fruit maturity identification method and system based on MobileNet v2 network improvement |
| CN111783802A (en) * | 2020-07-21 | 2020-10-16 | 南昌航空大学 | Metal fracture image identification method based on deep learning |
| CN112036472A (en) * | 2020-08-28 | 2020-12-04 | 长安大学 | Visual image classification method and system for power system |
| CN112036495A (en) * | 2020-09-02 | 2020-12-04 | 常熟理工学院 | Garment image classification method based on web crawler and transfer learning |
| CN112270402A (en) * | 2020-10-20 | 2021-01-26 | 山东派蒙机电技术有限公司 | Training method and system for tire wear identification model |
| CN112364899A (en) * | 2020-10-27 | 2021-02-12 | 西安科技大学 | Abrasive grain ferrographic image intelligent identification method based on virtual image and transfer learning |
| CN112287839A (en) * | 2020-10-29 | 2021-01-29 | 广西科技大学 | SSD infrared image pedestrian detection method based on transfer learning |
| CN112435237A (en) * | 2020-11-24 | 2021-03-02 | 山西三友和智慧信息技术股份有限公司 | Skin lesion segmentation method based on data enhancement and depth network |
| CN112541555A (en) * | 2020-12-22 | 2021-03-23 | 中国医学科学院北京协和医院 | Deep learning-based classifier model training method |
| WO2022134834A1 (en) * | 2020-12-23 | 2022-06-30 | 深圳壹账通智能科技有限公司 | Potential event predicting method, apparatus and device, and storage medium |
| CN113066053B (en) * | 2021-03-11 | 2023-10-10 | 紫东信息科技(苏州)有限公司 | Model migration-based duodenum self-training classification method and system |
| CN113066053A (en) * | 2021-03-11 | 2021-07-02 | 紫东信息科技(苏州)有限公司 | Model migration-based duodenum self-training classification method and system |
| CN112926679A (en) * | 2021-03-26 | 2021-06-08 | 江苏大学 | Method for enhancing background change resistance of classification model in intellectual property image recognition |
| CN112926679B (en) * | 2021-03-26 | 2024-06-11 | 江苏大学 | Method for enhancing background change resistance of classification model in intellectual property graph recognition |
| CN113298791A (en) * | 2021-05-31 | 2021-08-24 | 中电福富信息科技有限公司 | Image detection method of mixed cartoon based on deep learning |
| CN113536298B (en) * | 2021-06-11 | 2024-04-30 | 浙江工业大学 | Deep learning model bias poisoning attack-oriented defense method |
| CN113536298A (en) * | 2021-06-11 | 2021-10-22 | 浙江工业大学 | Defense method for bias poisoning attack of deep learning model |
| CN113610121B (en) * | 2021-07-22 | 2023-09-29 | 哈尔滨工程大学 | Cross-domain task deep learning identification method |
| CN113610121A (en) * | 2021-07-22 | 2021-11-05 | 哈尔滨工程大学 | Cross-domain task deep learning identification method |
| CN113706511A (en) * | 2021-08-31 | 2021-11-26 | 佛山市南海区广工大数控装备协同创新研究院 | Composite material damage detection method based on deep learning |
| CN115147737A (en) * | 2022-06-07 | 2022-10-04 | 滁州学院 | City tree species classification method based on ConvNeXt transfer learning |
| CN115100522A (en) * | 2022-06-27 | 2022-09-23 | 南京邮电大学 | Intelligent garden aerial photography tree species self-adaptive identification method based on tree growth model |
| CN115100522B (en) * | 2022-06-27 | 2024-07-12 | 南京邮电大学 | Intelligent garden aerial tree species self-adaptive identification method based on tree growth model |
| CN115719430A (en) * | 2022-10-28 | 2023-02-28 | 河北舒隽科技有限公司 | Method for identifying male and female of Taixing chick |
| CN116310511A (en) * | 2023-02-13 | 2023-06-23 | 南京林业大学 | Solid wood floor color classification method based on self-learning |
| CN116310511B (en) * | 2023-02-13 | 2024-07-30 | 南京林业大学 | Solid wood floor color classification method based on self-learning |
| CN116912879A (en) * | 2023-06-02 | 2023-10-20 | 广州大学 | Ear brushing identity recognition method based on transfer learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109508650A (en) | A kind of wood recognition method based on transfer learning | |
| CN106874956B (en) | The construction method of image classification convolutional neural networks structure | |
| CN108171266A (en) | A kind of learning method of multiple target depth convolution production confrontation network model | |
| CN108961245A (en) | Picture quality classification method based on binary channels depth parallel-convolution network | |
| CN110929610B (en) | Plant disease identification method and system based on CNN model and transfer learning | |
| CN107016406A (en) | The pest and disease damage image generating method of network is resisted based on production | |
| CN107945118A (en) | A kind of facial image restorative procedure based on production confrontation network | |
| CN109478232A (en) | The identification of weeds in natural environment | |
| CN107316066A (en) | Image classification method and system based on multi-path convolutional neural networks | |
| CN110211173A (en) | A kind of paleontological fossil positioning and recognition methods based on deep learning | |
| CN108961675A (en) | Fall detection method based on convolutional neural networks | |
| CN109325495A (en) | A kind of crop image segmentation system and method based on deep neural network modeling | |
| CN108230278A (en) | A kind of image based on generation confrontation network goes raindrop method | |
| CN110543895B (en) | Image classification method based on VGGNet and ResNet | |
| Zamuda et al. | Vectorized procedural models for animated trees reconstruction using differential evolution | |
| CN109886153A (en) | A kind of real-time face detection method based on depth convolutional neural networks | |
| CN114359727A (en) | Tea disease identification method and system based on lightweight optimization Yolo v4 | |
| CN109615010A (en) | Chinese traditional medicinal materials recognition method and system based on double scale convolutional neural networks | |
| CN114170137B (en) | Pepper disease identification method, identification system and computer readable storage medium | |
| CN104298974A (en) | Human body behavior recognition method based on depth video sequence | |
| CN110287882A (en) | A kind of big chrysanthemum kind image-recognizing method based on deep learning | |
| CN110399888A (en) | A kind of go judgment system based on MLP neural network and computer vision | |
| CN109190666B (en) | Flower image classification method based on improved deep neural network | |
| CN109978074A (en) | Image aesthetic feeling and emotion joint classification method and system based on depth multi-task learning | |
| CN116129260A (en) | Forage grass image recognition method based on deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190322 |
|
| WD01 | Invention patent application deemed withdrawn after publication |