CN103984959B - A kind of image classification method based on data and task-driven - Google Patents
A kind of image classification method based on data and task-driven Download PDFInfo
- Publication number
- CN103984959B CN103984959B CN201410224860.3A CN201410224860A CN103984959B CN 103984959 B CN103984959 B CN 103984959B CN 201410224860 A CN201410224860 A CN 201410224860A CN 103984959 B CN103984959 B CN 103984959B
- Authority
- CN
- China
- Prior art keywords
- image
- training
- convolutional neural
- neural networks
- data
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 39
- 238000012549 training Methods 0.000 claims abstract description 73
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 56
- 238000012360 testing method Methods 0.000 claims abstract description 32
- 239000000463 material Substances 0.000 claims abstract description 8
- 238000005457 optimization Methods 0.000 claims description 8
- 230000004044 response Effects 0.000 claims description 8
- 238000000605 extraction Methods 0.000 claims description 6
- 238000002939 conjugate gradient method Methods 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 2
- 230000001537 neural effect Effects 0.000 claims description 2
- 230000009466 transformation Effects 0.000 claims description 2
- 230000003044 adaptive effect Effects 0.000 abstract description 2
- 230000006870 function Effects 0.000 description 11
- 238000010586 diagram Methods 0.000 description 5
- 238000004422 calculation algorithm Methods 0.000 description 3
- 238000001514 detection method Methods 0.000 description 3
- 230000007935 neutral effect Effects 0.000 description 3
- 230000009467 reduction Effects 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 239000003086 colorant Substances 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000003542 behavioural effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000007635 classification algorithm Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 238000003709 image segmentation Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Abstract
The invention discloses a kind of image classification method based on data and task-driven, this method includes：Convolutional neural networks structure is designed according to data set scale and picture material；It is trained using given grouped data set pair convolutional neural networks model；Using the convolutional neural networks after training to training set image zooming-out feature representation；The convolutional neural networks that test image is inputted after training, and classified.The inventive method is based on linear Convolution feature learning, can in the way of data-driven implementation model to the adaptive of data set, so as to preferably describe specific data set, error directly to k nearest neighbor by way of task-driven is optimized, and can obtain better performance in k nearest neighbor task；And efficient training can be carried out using GPU in the training stage, need to only use CPU just to realize efficient k nearest neighbor image classification in test phase, be highly suitable for the tasks such as large-scale image classification, retrieval.
Description
Technical field
It is more particularly to a kind of to be based on data and task-driven the present invention relates to Image Classfication Technology field in computer vision
Image classification method.
Background technology
Image classification, which is that computer vision is most basic, one of to study a question, and the problem of it will be solved is exactly a given figure
As whether automatic decision wherein includes certain type objects.Image classification problem is one core topic of vision research, many other to regard
Feel that research will rely on and be related to object point in image classification problem, such as objects in images detection, tracking, image segmentation, video
Class, detection, tracking, behavioural analysis, gesture identification etc..
K nearest neighbor image classification is a kind of image classification method, refers to use k nearest neighbor ballot when to image classification
The most class prediction of occurrence number is the classification of the test sample in mode, i.e. K a nearest image.Except can simply,
Efficiently realize outside being classified to image, k nearest neighbor classification also has a lot of other characteristics.Such as k nearest neighbor image classification can be obtained
Take with the immediate sample of test image, can apply in fields such as image retrieval, face retrieval, video frequency searchings.
Because the selection of grader and image feature representation are two independent processes in conventional art, and k nearest neighbor point
Class is a nonparametric model, and it predicts the spatial distribution for depending critically upon data, namely image feature representation, and this results in figure
As feature representation is not optimal for k nearest neighbor classification, classification performance is impacted.
In recent years, image classification field is quickly grown, and many important breakthroughs are achieved in terms of sorting technique.Currently, word
Bag model is one of image feature representation main flow framework.Word bag model is described by the low-level image feature of the image block to intensive extraction
Statistical nature description is carried out, obtains and the global feature of image is expressed.Word bag model is generally described by low-level image feature, vision word
The step such as generation, low-level image feature coding, feature convergence, classifier training and test is constituted, and before classifier training, we can be with
Think that word bag model uses unsupervised mode and image is expressed, whether the low-level image feature such as traditional SIFT, HOG
Or the expression of word bag model middle level features, does not all use the label information of image, thus such feature representation is near for K
It is not usually optimal for the such printenv model of neighbour's classification.
The content of the invention
In view of this, the main object of the present invention is to provide a kind of image classification method based on data and task-driven,
To realize on large scale image data set more fast and accurately image classification.
In order to achieve the above object, the present invention uses following technical scheme：
A kind of image classification method based on data and task-driven, including：
Data set prepares, and convolutional neural networks structure is designed according to data set scale and picture material；
Model training, is trained using given grouped data set pair convolutional neural networks model；
Feature representation extraction is carried out to training set image using the convolutional neural networks after training；
The convolutional neural networks that test image is inputted after training, are classified using k nearest neighbor mode to test image.
Further, the data set prepares, and convolutional neural networks structure is designed according to data set scale and picture material,
Also include：
Realize that data strengthen at least through the one or more in the following manner：1) image is removed from original image random cropping
The marginal portion of surrounding is to produce the new sample image with nuance；2) added in original image pixels random high
This noise produces new sample image.
Further, the data set prepares, and convolutional neural networks structure is designed according to data set scale and picture material,
Also include：
Image pattern is zoomed into fixed dimension, and pixel is stretched into a vector and is used as the defeated of convolutional neural networks
Enter.
Further, the model training, is trained using given grouped data set pair convolutional neural networks model, tool
Body also includes：
Essential characteristic transformation model is used as using convolutional neural networks；
It is expected that error rate is trained as loss function to convolutional neural networks model based on adjacent PCA；
Optimization method based on gradient carries out network training, and carries out computing using GPU.
Further, the convolutional neural networks using after training carry out extraction feature representation, bag to training set image
Include：
The convolutional neural networks that all training images are inputted after training, take out the response conduct of last layer of full articulamentum
The feature representation of every training image.
Further, the feature representation of training set image is configured to KD- trees and prestored.
Further, the convolutional neural networks that test image is inputted after training, using k nearest neighbor mode to test chart
As being classified, including：
For giving test image, by the image scaling to convolutional neural networks mode input size, convolution is then fed into
Neutral net carries out forward calculation, takes out the response of last layer of full articulamentum as the feature representation of the test image, uses
The expression carries out k nearest neighbor retrieval in the feature representation of training set image, will occur in K nearest training image of feature representation
The most class prediction of number of times is the classification of the test image.
Further, it is described it is expected that error rate is entered as loss function to convolutional neural networks model based on adjacent PCA
Row training, is specifically included：
K nearest neighbor error in classification is estimated using adjacent PCA NCA, N is given to training sample { (xi,yi) | i=
1 ..., N }, wherein xiIt is image pattern, yiIt is its corresponding label, for a sample xiFor, another sample xjWith xiCategory
It is in the other definition of probability of same class
Wherein F () is by the eigentransformation function of the convolutional neural networks；
For adjacent PCA, sample xiThe probability for belonging to classification c and correctly being classified is
It is expected that error rate is
Wherein, yijRepresent sample xiWith sample xjWhether same category, y are belonged toi=yjWhen, yij=1, otherwise yij=0.It is public
Expectation error rate in formula (3) is one of k nearest neighbor classification error rate approximate, with damage of the expectation error rate as the network optimization
Lose function.
Further, the optimization method based on gradient carries out network training and specifically includes one of in the following manner：At random
Gradient decline, conjugate gradient method, quasi-Newton method, L-BFGS.
The above-mentioned image classification method based on data and task-driven that the present invention is provided, compared with prior art with
Lower advantage：
1), using based on linear Convolution feature learning, can in the way of data-driven implementation model to data set
Adaptively, so as to preferably describe specific data set.
2), optimized, convolutional neural networks are carried out in the way of task-driven excellent by the error directly to k nearest neighbor
Change, better performance can be obtained in k nearest neighbor task.
3), efficient training can be carried out using GPU in the training stage, need to only uses CPU just to realize in test phase
Efficient k nearest neighbor image classification, is highly suitable for the tasks such as large-scale image classification, retrieval.
Brief description of the drawings
Fig. 1 is to be flowed according to image classification method model training of the embodiment of the present invention based on data and task-driven with test
Cheng Tu；
Fig. 2 is trained on handwritten numeral database MNIST using adjacent PCA according to the embodiment of the present invention and obtained
The sub- schematic diagram of convolutional neural networks first layer convolution；
Fig. 3 is to train obtained convolution god on CIFAR-10 databases using adjacent PCA according to the embodiment of the present invention
Through the sub- schematic diagram of network first tier convolution；
Fig. 4 be according to the embodiment of the present invention to MNIST use of numerals method dimensionality reduction proposed by the present invention to 2 dimension result with
And the contrast with other method, the data point of different colours represents different numerals.
Embodiment
For the object, technical solutions and advantages of the present invention are more clearly understood, below in conjunction with specific embodiment, and reference
Accompanying drawing, the present invention is described in more detail.
The present invention thought main points be：1) linear Convolution feature learning is based on, can be realized in the way of data-driven
Model is to the adaptive of data set, so as to preferably describe specific data set；2) present invention is by directly to the error of k nearest neighbor
Optimize, convolutional neural networks are optimized in the way of task-driven, can be obtained in k nearest neighbor task more preferably
Performance；3) efficient training can be carried out using GPU in the training stage, only CPU need to be used just to realize height in test phase
The k nearest neighbor image classification of effect, is highly suitable for the tasks such as large-scale image classification, retrieval.
As shown in figure 1, Fig. 1 top halfs are according to image classification side of the embodiment of the present invention based on data and task-driven
Method model training flow chart.The structure of convolutional neural networks is shown in figure, the sub- size of preceding networks layer convolution is 5 × 5, with
The reduction of trellis diagram size, the sub- size of convolution used below is connected to complete comprising 128 and 64 neurons for 3 × 3, finally
Articulamentum.By the way that training set sample order is sent into convolutional neural networks, and using adjacent PCA loss function from back to front
Error propagation is carried out, the gradient per layer network parameter is calculated, and network is carried out using stochastic gradient descent algorithm with this gradient
Update, to realize that the network model of data-driven and task-driven learns.Fig. 1 the latter half elaborates the test of this method
Journey.The convolutional neural networks that different classes of original image input is trained, the sound that last full articulamentum is obtained by network
The feature representation of image is should be used as, a new non-linear space is constituted, different classes of image can be more within this space
Make a distinction well.
The method of the present invention comprises the following steps：
S1, data set prepare, and convolutional neural networks structure is designed according to data set scale and picture material.
Convolutional neural networks include substantial amounts of model parameter, and parameter means that model is more complicated, easier in training
When occur over-fitting situation, i.e. algorithm on training set performance very well, performance is very poor on test set.Balance training number
It is to prevent over-fitting according to scale and model complexity, obtains the important channel of optimum performance.On the one hand, data volume is bigger,
Model training is less susceptible to over-fitting, and performance is better.But data volume is typically limited, and this is accomplished by using certain method people
Work produces new data, and data enhancing is accomplished by the following way in we：1) image surrounding is gone from original image random cropping
Marginal portion is to produce the new sample image with nuance；2) random Gaussian is added in original image pixels
Produce new sample image.All image patterns are all scaled to fixed dimension, and all pixels stretch into one to
Measure the input as convolutional neural networks.On the other hand, it is necessary to which accordingly control convolution is refreshing in the case of given training set scale
Model complexity through network.Structure of the model complexity of convolutional neural networks generally with model is directly related, the layer of network
Number it is more, every layer of nodes are more, can training parameter it is more, model is more complicated.
S2, based on adjacent PCA expect error rate convolutional neural networks are trained as loss function, optimization side
Method uses stochastic gradient descent, and carries out computing using GPU.
Convolutional neural networks are widely used in fields such as image classification, detection, segmentations.In such applications,
Convolutional neural networks are normally based on what general object classification criterion was trained, such as logistic is returned, Softmax is returned,
Cross entropy etc..Obtained network is trained using this general standard, can directly predict and obtain result, but to be used for directly carrying out
Handle k nearest neighbor classification problem, then it is not usually optimal.Above-mentioned general classification be end to end, it is direct from image
Predicted, and k nearest neighbor problem is typically to have a feature representation to every image, and k nearest neighbor inspection is carried out using this feature expression
Rope, the classification of test sample is determined by the classification of K nearest samples.If the net that we are trained using general sorting criterion
Network carries out feature representation (can take last layer of response of network as feature) to image, it is difficult to ensure that this feature expression is applicable
In nearest neighbour classification scene.The feature representation of k nearest neighbor problem is preferably directed to for study, we come direct using new training criterion
K nearest neighbor error is optimized, so that the feature representation for ensureing study is optimal under k nearest neighbor problem.
It is irrational directly to use k nearest neighbor error as object function, since it is desired that object function can continuously be led,
So as to be updated using stochastic gradient descent method to network.We are using adjacent PCA (Neighborhood
Component Analysis, NCA) to carry out approximate description to k nearest neighbor error in classification.Given N is to training sample { (xi,yi)|i
=1 ..., N }, wherein xiIt is image pattern, yiIt is its corresponding label.For a sample xiFor, another sample xjWith xi
Belonging to the other definition of probability of same class is
Wherein F () is a highly complex non-linear transform function, by input picture be for conversion into a feature to
Amount, we represent F () using convolutional neural networks here, and take out last layer of response of neutral net as mark sheet
Reach.From formula 1 as can be seen that sample xjWith xiFall into same category of probability be inversely proportional to it is European between both feature representations
Distance.
In k nearest neighbor classification, the prediction classification of a test sample is that occur most in its K nearest sample by counting
Many classifications.For adjacent PCA, sample xiThe probability for belonging to classification c and correctly being classified is
Then it is expected that error rate can be defined as
Wherein yijRepresent sample xiWith sample xjWhether same category, y are belonged toi=yjWhen, yij=1, otherwise yij=0.It is public
Expectation error rate in formula 3 is one of k nearest neighbor classification error rate approximate, excellent as network with the expectation error rate in this patent
The object function of change.The adjacent component lost function that formula 3 is defined can continuously be led, thus can easily be used based on gradient
Optimization method carry out network training, such as stochastic gradient descent, conjugate gradient method, quasi-Newton method, L-BFGS.
S3, the convolutional neural networks for training the input of all training images, take out the response of last layer of full articulamentum
It is used as the feature representation of every training image.After convolutional neural networks are trained, we become this network as feature
Exchange the letters number F (), for the extraction feature from image.K nearest neighbor classification is nonparametric model, and itself is without parameter, it is only necessary to will
The each image input convolutional neural networks of training set, and last layer of full articulamentum is taken out as the mark sheet of the sample
Reach.The feature representation of the training set extracted can be prestored, and be retrieved in test phase for k nearest neighbor.
S4, the convolutional neural networks for training test image input, take out the response conduct of last layer of full articulamentum
The feature representation of this image, and use k nearest neighbor mode in the character pair expression of training set image using this feature expression
Classified, by the classification that the most class prediction of occurrence number in the immediate K image of feature representation is the test image.
The feature representation that the feature representation of test image is extracted in advance with training set image is compared, and is generally required and all training
Sample is all once compared, and this time complexity is proportional to training set size.For very big training set, this shows
Right cost is very high, be it is faster carry out k nearest neighbor lookup, using KD-Tree this data commonly used in nearest _neighbor retrieval
Structure, by the way that K dimension spaces are split, accelerates the speed of neighbour's retrieval.
Fig. 2 is trained on handwritten numeral database MNIST using adjacent PCA according to the embodiment of the present invention and obtained
The sub- schematic diagram of convolutional neural networks first layer convolution.Wherein, the sample in MNIST databases, right figure displaying is shown in left figure
Be convolutional neural networks first layer convolution after training, it can be seen that substantially digital stroke learn.
Fig. 3 is to train obtained convolution god on CIFAR-10 databases using adjacent PCA according to the embodiment of the present invention
Through the sub- schematic diagram of network first tier convolution.Find out from right figure, convolution learnt is some edges and simple bottom vision
Pattern.
Fig. 4 be the method dimensionality reduction that is proposed according to the embodiment of the present invention to MNIST use of numerals this patent to 2 dimensions results with
And the contrast with other method, the data point of different colours represents different numerals.As can be seen that different digital mark sheets
Realize and preferably cluster up to spatial distribution, suitable for next being classified using k nearest neighbor mode classification.
In a word, the present invention proposes a kind of new image classification method based on data and task-driven, is rolled up using deep layer
Product neutral net carries out model training as feature representation model, and using stochastic gradient descent algorithm on GPU.Convolutional Neural
After the completion of network training, for extracting feature representation from image and carrying out image classification using k nearest neighbor way of search.Test table
The bright invention compared with image classification algorithms of the main flow based on k nearest neighbor have feature representation have strong identification, model training by
Data-driven, task-driven, test process are extremely efficient, the k nearest neighbor image classification being suitable under large scale data and retrieval.
Particular embodiments described above, has been carried out further in detail to the purpose of the present invention, technical scheme and beneficial effect
Describe in detail it is bright, should be understood that the foregoing is only the present invention specific embodiment, be not intended to limit the invention, it is all
Within the spirit and principles in the present invention, any modification, equivalent substitution and improvements done etc., should be included in the guarantor of the present invention
Within the scope of shield.
Claims (7)
1. a kind of image classification method based on data and task-driven, it is characterised in that including：
Data set prepares, and convolutional neural networks structure is designed according to data set scale and picture material；
Model training, is trained using given grouped data set pair convolutional neural networks model；
Feature representation extraction is carried out to training set image using the convolutional neural networks after training；
The convolutional neural networks that test image is inputted after training, are classified using k nearest neighbor mode to test image；
Wherein, the model training, is trained using given grouped data set pair convolutional neural networks model, specifically also wrapped
Include：
Essential characteristic transformation model is used as using convolutional neural networks；
It is expected that error rate is trained as loss function to convolutional neural networks model based on adjacent PCA；
Optimization method based on gradient carries out network training, and carries out computing using GPU；
It is expected that error rate is trained as loss function to convolutional neural networks model based on adjacent PCA, specifically include：
K nearest neighbor error in classification is estimated using adjacent PCA NCA, N is given to training sample { (xi, yi) | i=
1 ..., N }, wherein, xiIt is image pattern, yiIt is its corresponding label, for a sample xiFor, another sample xjWith xi
Belonging to the other definition of probability of same class is
Wherein, F () is the eigentransformation function of the convolutional neural networks；
For adjacent PCA, sample xiThe probability for belonging to classification c and correctly being classified is
It is expected that error rate is
Wherein, yijRepresent sample xiWith sample xjWhether same category, y are belonged toi=yjWhen, yij=1, otherwise yij=0；The phase
Prestige error rate is one of k nearest neighbor classification error rate approximate, with loss function of the expectation error rate as the network optimization.
2. the image classification method according to claim 1 based on data and task-driven, it is characterised in that data set is accurate
It is standby, convolutional neural networks structure is designed according to data set scale and picture material, in addition to：
Realize that data strengthen at least through the one or more in the following manner：1) image surrounding is gone from original image random cropping
Marginal portion to produce the new sample image with nuance；2) random Gaussian is added in original image pixels to make an uproar
Sound produces new sample image.
3. the image classification method according to claim 1 based on data and task-driven, it is characterised in that data set is accurate
It is standby, convolutional neural networks structure is designed according to data set scale and picture material, in addition to：
Image pattern is zoomed into fixed dimension, and pixel is stretched into a vector as the input of convolutional neural networks.
4. the image classification method according to claim 1 based on data and task-driven, it is characterised in that use training
Convolutional neural networks afterwards carry out extraction feature representation to training set image, including：
The convolutional neural networks that all training images are inputted after training, the response for taking out last layer of full articulamentum is used as every
The feature representation of training image.
5. the image classification method according to claim 4 based on data and task-driven, it is characterised in that by training set
The feature representation of image is configured to KD- trees and prestored.
6. the image classification method according to claim 1 based on data and task-driven, it is characterised in that by test chart
As the convolutional neural networks after input training, test image is classified using k nearest neighbor mode, including：
For giving test image, by the image scaling to convolutional neural networks mode input size, convolutional Neural is then fed into
Network carries out forward calculation, takes out the response of last layer of full articulamentum as the feature representation of the test image, uses the table
K nearest neighbor retrieval is carried out up in the feature representation in training set image, by occurrence number in K nearest training image of feature representation
Most class predictions is the classification of the test image.
7. the image classification method according to claim 1 based on data and task-driven, it is characterised in that based on gradient
Optimization method carry out network training specifically include one of in the following manner：Stochastic gradient descent, conjugate gradient method, quasi-Newton method,
L-BFGS。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410224860.3A CN103984959B (en) | 2014-05-26 | 2014-05-26 | A kind of image classification method based on data and task-driven |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410224860.3A CN103984959B (en) | 2014-05-26 | 2014-05-26 | A kind of image classification method based on data and task-driven |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103984959A CN103984959A (en) | 2014-08-13 |
| CN103984959B true CN103984959B (en) | 2017-07-21 |
Family
ID=51276921
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410224860.3A Active CN103984959B (en) | 2014-05-26 | 2014-05-26 | A kind of image classification method based on data and task-driven |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103984959B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106326931A (en) * | 2016-08-25 | 2017-01-11 | 南京信息工程大学 | Mammary gland molybdenum target image automatic classification method based on deep learning |
Families Citing this family (42)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106663186B (en) * | 2014-07-28 | 2018-08-21 | 北京市商汤科技开发有限公司 | method and system for face recognition |
| CN104182735A (en) * | 2014-08-18 | 2014-12-03 | 厦门美图之家科技有限公司 | Training optimization pornographic picture or video detection method based on convolutional neural network |
| CN104217225B (en) * | 2014-09-02 | 2018-04-24 | 中国科学院自动化研究所 | A kind of sensation target detection and mask method |
| CN104361363B (en) * | 2014-11-25 | 2018-01-16 | 中国科学院自动化研究所 | Depth deconvolution feature learning network, generation method and image classification method |
| WO2016090520A1 (en) * | 2014-12-10 | 2016-06-16 | Xiaogang Wang | A method and a system for image classification |
| CN104517122A (en) * | 2014-12-12 | 2015-04-15 | 浙江大学 | Image target recognition method based on optimized convolution architecture |
| CN104572965A (en) * | 2014-12-31 | 2015-04-29 | 南京理工大学 | Search-by-image system based on convolutional neural network |
| CN104700118A (en) * | 2015-03-18 | 2015-06-10 | 中国科学院自动化研究所 | Pulmonary nodule benignity and malignancy predicting method based on convolutional neural networks |
| CN104778448B (en) * | 2015-03-24 | 2017-12-15 | 孙建德 | A kind of face identification method based on structure adaptive convolutional neural networks |
| CN106056529B (en) * | 2015-04-03 | 2020-06-02 | 阿里巴巴集团控股有限公司 | Method and equipment for training convolutional neural network for picture recognition |
| CN104794501B (en) * | 2015-05-14 | 2021-01-05 | 清华大学 | Pattern recognition method and device |
| CN104991959B (en) * | 2015-07-21 | 2019-11-05 | 北京京东尚科信息技术有限公司 | A kind of method and system of the same or similar image of information retrieval based on contents |
| CN105022836B (en) * | 2015-08-14 | 2018-07-03 | 中国科学技术大学 | Compact depth CNN aspect indexing methods based on SIFT insertions |
| CN106485192B (en) * | 2015-09-02 | 2019-12-06 | 富士通株式会社 | Training method and device of neural network for image recognition |
| CN105447866A (en) * | 2015-11-22 | 2016-03-30 | 南方医科大学 | X-ray chest radiograph bone marrow suppression processing method based on convolution neural network |
| CN105469400B (en) * | 2015-11-23 | 2019-02-26 | 广州视源电子科技股份有限公司 | The quick method and system for identifying, marking of electronic component polar orientation |
| CN105590099B (en) * | 2015-12-22 | 2019-02-01 | 中国石油大学(华东) | A kind of more people's Activity recognition methods based on improvement convolutional neural networks |
| CN105354565A (en) * | 2015-12-23 | 2016-02-24 | 北京市商汤科技开发有限公司 | Full convolution network based facial feature positioning and distinguishing method and system |
| CN105718890A (en) * | 2016-01-22 | 2016-06-29 | 北京大学 | Method for detecting specific videos based on convolution neural network |
| CN105718960B (en) * | 2016-01-27 | 2019-01-04 | 北京工业大学 | Based on convolutional neural networks and the matched image order models of spatial pyramid |
| US10282588B2 (en) * | 2016-06-09 | 2019-05-07 | Siemens Healthcare Gmbh | Image-based tumor phenotyping with machine learning from synthetic data |
| CN105894046B (en) * | 2016-06-16 | 2019-07-02 | 北京市商汤科技开发有限公司 | Method and system, the computer equipment of convolutional neural networks training and image procossing |
| CN106203442B (en) * | 2016-06-28 | 2019-04-05 | 北京小白世纪网络科技有限公司 | A kind of copy image feature extracting method based on deep learning |
| CN106373112B (en) * | 2016-08-31 | 2020-08-04 | 北京比特大陆科技有限公司 | Image processing method and device and electronic equipment |
| CN107871497A (en) * | 2016-09-23 | 2018-04-03 | 北京眼神科技有限公司 | Audio recognition method and device |
| CN106650694A (en) * | 2016-12-30 | 2017-05-10 | 江苏四点灵机器人有限公司 | Human face recognition method taking convolutional neural network as feature extractor |
| CN107145898B (en) * | 2017-04-14 | 2020-04-10 | 北京航星机器制造有限公司 | Radiographic image classification method based on neural network |
| CN107256384A (en) * | 2017-05-22 | 2017-10-17 | 汕头大学 | A kind of card recognition and method of counting based on image and signal transacting |
| CN107564070B (en) * | 2017-09-05 | 2020-06-12 | 国网浙江省电力公司湖州供电公司 | Large-scene binocular ranging and correcting method in overhead power transmission channel image monitoring |
| CN107609586A (en) * | 2017-09-08 | 2018-01-19 | 深圳市唯特视科技有限公司 | A kind of visual characteristic learning method based on self-supervision |
| CN110348428B (en) * | 2017-11-01 | 2023-03-24 | 腾讯科技（深圳）有限公司 | Fundus image classification method and device and computer-readable storage medium |
| CN108010049A (en) * | 2017-11-09 | 2018-05-08 | 华南理工大学 | Split the method in human hand region in stop-motion animation using full convolutional neural networks |
| CN107958219A (en) * | 2017-12-06 | 2018-04-24 | 电子科技大学 | Image scene classification method based on multi-model and Analysis On Multi-scale Features |
| CN108288282B (en) * | 2017-12-26 | 2022-04-08 | 浙江工业大学 | Adaptive feature selection target tracking method based on convolutional neural network |
| CN108235770B (en) * | 2017-12-29 | 2021-10-19 | 达闼机器人有限公司 | Image identification method and cloud system |
| CN108304546B (en) * | 2018-01-31 | 2020-09-01 | 杭州电子科技大学 | Medical image retrieval method based on content similarity and Softmax classifier |
| WO2019222734A1 (en) * | 2018-05-18 | 2019-11-21 | Google Llc | Learning data augmentation policies |
| CN109584142A (en) * | 2018-12-05 | 2019-04-05 | 网易传媒科技（北京）有限公司 | Image Intensified System and method, training method, medium and electronic equipment |
| CN109740611A (en) * | 2019-01-25 | 2019-05-10 | 中电健康云科技有限公司 | Tongue image analysis method and device |
| CN110110780B (en) * | 2019-04-30 | 2023-04-07 | 南开大学 | Image classification method based on antagonistic neural network and massive noise data |
| CN110689961B (en) * | 2019-09-03 | 2022-12-09 | 重庆大学 | Gastric cancer disease risk detection device based on big data analysis technology |
| US11157811B2 (en) | 2019-10-28 | 2021-10-26 | International Business Machines Corporation | Stub image generation for neural network training |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103177264A (en) * | 2013-03-14 | 2013-06-26 | 中国科学院自动化研究所 | Image classification method based on visual dictionary global topological representations |
| CN103366180A (en) * | 2013-06-14 | 2013-10-23 | 山东大学 | Cell image segmentation method based on automatic feature learning |
| CN103544506A (en) * | 2013-10-12 | 2014-01-29 | Tcl集团股份有限公司 | Method and device for classifying images on basis of convolutional neural network |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8582807B2 (en) * | 2010-03-15 | 2013-11-12 | Nec Laboratories America, Inc. | Systems and methods for determining personal characteristics |
-
2014
- 2014-05-26 CN CN201410224860.3A patent/CN103984959B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103177264A (en) * | 2013-03-14 | 2013-06-26 | 中国科学院自动化研究所 | Image classification method based on visual dictionary global topological representations |
| CN103366180A (en) * | 2013-06-14 | 2013-10-23 | 山东大学 | Cell image segmentation method based on automatic feature learning |
| CN103544506A (en) * | 2013-10-12 | 2014-01-29 | Tcl集团股份有限公司 | Method and device for classifying images on basis of convolutional neural network |
Non-Patent Citations (2)
| Title |
|---|
| ImageNet Classification with Deep Convolutional Neural Networks;Alex Krizhevsky;《Advances in Neural Information Processing Systems》;20121231;第25卷;第1106-1114页 * |
| 基于CNN和随机弹性形变的相似手写汉字识别;高学等;《华南理工大学学报(自然科学版)》;20140131;第42卷(第1期);第72-76、83页 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106326931A (en) * | 2016-08-25 | 2017-01-11 | 南京信息工程大学 | Mammary gland molybdenum target image automatic classification method based on deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103984959A (en) | 2014-08-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103984959B (en) | A kind of image classification method based on data and task-driven | |
| US10846566B2 (en) | Method and system for multi-scale cell image segmentation using multiple parallel convolutional neural networks | |
| US20190228268A1 (en) | Method and system for cell image segmentation using multi-stage convolutional neural networks | |
| CN108229338B (en) | Video behavior identification method based on deep convolution characteristics | |
| Kümmerer et al. | DeepGaze II: Reading fixations from deep features trained on object recognition | |
| CN106157319B (en) | The conspicuousness detection method in region and Pixel-level fusion based on convolutional neural networks | |
| CN109948714B (en) | Chinese scene text line identification method based on residual convolution and recurrent neural network | |
| CN112966691B (en) | Multi-scale text detection method and device based on semantic segmentation and electronic equipment | |
| CN107943856A (en) | A kind of file classification method and system based on expansion marker samples | |
| CN110334589B (en) | High-time-sequence 3D neural network action identification method based on hole convolution | |
| CN107506761A (en) | Brain image dividing method and system based on notable inquiry learning convolutional neural networks | |
| CN108875076B (en) | Rapid trademark image retrieval method based on Attention mechanism and convolutional neural network | |
| CN104217225A (en) | A visual target detection and labeling method | |
| CN106611427A (en) | A video saliency detection method based on candidate area merging | |
| CN106778687A (en) | Method for viewing points detecting based on local evaluation and global optimization | |
| CN106960415A (en) | A kind of method for recovering image based on pixel-recursive super-resolution model | |
| CN103996018A (en) | Human-face identification method based on 4DLBP | |
| CN105184772A (en) | Adaptive color image segmentation method based on super pixels | |
| CN105279519A (en) | Remote sensing image water body extraction method and system based on cooperative training semi-supervised learning | |
| CN103745233B (en) | The hyperspectral image classification method migrated based on spatial information | |
| Freytag et al. | Labeling examples that matter: Relevance-based active learning with gaussian processes | |
| Sharma et al. | Deep eigen space based ASL recognition system | |
| CN104484886A (en) | Segmentation method and device for MR image | |
| CN110084136A (en) | Context based on super-pixel CRF model optimizes indoor scene semanteme marking method | |
| CN114387270B (en) | Image processing method, image processing device, computer equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant |