CN106570516A - Obstacle recognition method using convolution neural network - Google Patents
Obstacle recognition method using convolution neural network Download PDFInfo
- Publication number
- CN106570516A CN106570516A CN201610804686.9A CN201610804686A CN106570516A CN 106570516 A CN106570516 A CN 106570516A CN 201610804686 A CN201610804686 A CN 201610804686A CN 106570516 A CN106570516 A CN 106570516A
- Authority
- CN
- China
- Prior art keywords
- neural networks
- convolutional neural
- layer
- convolutional
- recognition method
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24133—Distances to prototypes
- G06F18/24137—Distances to cluster centroïds
- G06F18/2414—Smoothing the distance, e.g. radial basis function networks [RBFN]
-
- 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)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses an obstacle recognition method using a convolution neural network (CNN). The method comprises the following steps: firstly, establishing a CNN model; selecting a dispersed data sample; setting a network structure parameter and a training parameter of the CNN; then inputting into the CNN model the dispersed data sample to be processed; finally outputting a classification result of the dispersed data sample performed by the CNN model. The deep learning provided by the present invention is based on a bionic eye system for recognizing the human body as an obstacle, and improves the accuracy of the recognition. Further, the deep learning is used to identify the human body as an obstacle. A method for configuring an interface in an identification process is provided and an identification process and the communication process of the bionic eye system are enhanced.
Description
Technical field
The present invention relates to field of image recognition, particularly a kind of obstacle recognition method of utilization convolutional neural networks CNN.
Background technology
In recent years, deep learning development is swift and violent, particularly in terms of identification has obtained very big development;In deep learning
It is more effective that CNN convolutional neural networks are compared to other deep learning models in terms of image recognition.Human body based on bionic eye
As the target identification technology of barrier;As bionical eye system is exported using video flowing, therefore train object and identification object
Dynamic image is, this only trains identification static images different from conventional most of deep learnings.But dynamic image can draw
Play ghost and similar error;This error can reduce recognition accuracy.
Accordingly, it would be desirable to a kind of recognition methodss of the high barrier of recognition accuracy.
The content of the invention
The purpose of the present invention is to propose to a kind of recognition methodss of the high barrier of recognition accuracy;The method utilizes depth
Habit technology carries out human body as the identification of barrier;Improve and recognized based on the network googlenet in CNN convolutional neural networks
Model.
The purpose of the present invention is achieved through the following technical solutions:
The obstacle recognition method of the utilization convolutional neural networks CNN that the present invention is provided, comprises the following steps:
S1:Set up convolutional neural networks model;
S2:Select dispersion data sample;
S3:The network architecture parameters and training parameter of convolutional neural networks are set;
S4:Dispersion data sample is input in convolutional neural networks model and is processed;
S5:Classification results of the output convolutional neural networks model to dispersion data sample.
Further, the convolutional neural networks model includes convolutional layer, down-sampling layer and Softmax output layers;
The convolutional layer, for the order of magnitude of network parameter is obtained by convolutional calculation;
The down-sampling layer, for obtaining the sub-sample of image;
The Softmax output layers, for obtaining the residual error of output layer and the residual error in intermediate layer;The residual error of the output layer
It is that output valve is worth error amount with category, the residual error in the intermediate layer is the weighted sum of next layer of residual error.
Further, processing procedure of the convolutional neural networks model to dispersion data sample, comprises the following steps that:
S21:The data sample of input is calculated into sampling characteristic pattern by the convolution operator of convolutional layer;
S22:Sampling characteristic pattern is calculated into double sampling characteristic pattern by down-sampling layer;
S23:Double sampling characteristic pattern is obtained being connected entirely by convolution and down-sampling and paves characteristic pattern,
S24:Characteristic pattern is paved in the full connection of output.
Further, the training parameter of the convolutional neural networks model includes learning rate;The learning rate Learning
Rate is set to 0.001-0.005.
Further, the training parameter of the convolutional neural networks model includes batchsize parameters;The batchsize
Parameter is 6-24;
Further, in the convolutional neural networks model, Softmax output layers are provided with 3-10 output channel.
Further, the softmax output layers are two graders.
As a result of above-mentioned technical proposal, the present invention has the advantage that:
The deep learning that the present invention is provided is to carry out human body as the identification of barrier based on bionical eye system;Improve identification
Accuracy rate.Human body is carried out as the identification of barrier by the use of depth learning technology simultaneously;Configure there is provided in identification process
The method of interface, strengthens the communication process of identification process and bionical eye system.
Other advantages of the present invention, target and feature will be illustrated to a certain extent in the following description, and
And to a certain extent, based on being will be apparent to investigating hereafter to those skilled in the art, Huo Zheke
To be instructed from the practice of the present invention.The present invention target and other advantages can be realized by description below and
Obtain.
Description of the drawings
The description of the drawings of the present invention is as follows.
Fig. 1 is neuronal structure model.
Fig. 2 is convolutional neural networks topology.
Fig. 3 is Googlenet network structures.
Fig. 4 is to train the error of 400 times to be distributed.
Fig. 5 is to change the curve of error change that learning rate trains 400 times.
Fig. 6 is 400 curve of error distributions of training.
Fig. 7 is 400 curve of error of training.
Fig. 8-1 is the design sketch that two sorter networks contrast that recognition result is behaved.
Fig. 8-2 is that two sorter networks contrast recognition result is other design sketchs.
Fig. 8-3 is that two sorter networks contrast design sketch of the recognition result for car.
Fig. 9 is the obstacle recognition method flow chart of improved convolutional neural networks CNN.
Specific embodiment
The invention will be further described with reference to the accompanying drawings and examples.
Embodiment 1
As shown in figure 9, a kind of obstacle recognition method of utilization convolutional neural networks CNN of the present embodiment offer, including
Following steps:
S1:Set up convolutional neural networks model;
S2:Select dispersion data sample;
S3:The network architecture parameters and training parameter of convolutional neural networks are set;
S4:Dispersion data sample is input in convolutional neural networks model and is processed;
S5:Classification results of the output convolutional neural networks model to dispersion data sample.
The convolutional neural networks model includes convolutional layer, down-sampling layer and Softmax output layers;
The convolutional layer, is distinguished come simulation feature by convolution for CNN, and is shared by the weights of convolution and pond
To drop
The order of magnitude of low network parameter, completes the tasks such as classification finally by traditional neural network;
The down-sampling layer, for the principle using image local correlation, carries out sub-sample to image, it is possible to reduce number
Retain useful information simultaneously according to treating capacity;
The Softmax output layers, the residual error for the output layer of CNN are different from the residual computations mode in intermediate layer, defeated
The residual error for going out layer is that output valve is worth error amount with category, and the residual error of middle each layer is from the weighting of next layer of residual error
With.
Processing procedure of the convolutional neural networks model to dispersion data sample, comprises the following steps that:
S21:The data sample of input is calculated into sampling characteristic pattern by the convolution operator of convolutional layer;
S22:Sampling characteristic pattern is calculated into double sampling characteristic pattern by down-sampling layer;
S23:Double sampling characteristic pattern is obtained being connected entirely by convolution and down-sampling and paves characteristic pattern,
S24:Characteristic pattern is paved in the full connection of output.
The training parameter of the convolutional neural networks model includes learning rate;The learning rate Learning Rate are arranged
For 0.001-0.005.The training parameter of the convolutional neural networks model includes batchsize parameters;The batchsize ginsengs
Number is 6-24;In the convolutional neural networks model, Softmax output layers are provided with 3-10 output channel.State softmax defeated
Go out layer for two graders.
The training parameter of the convolutional neural networks model includes learning rate;The learning rate Learning Rate are arranged
For 0.001-0.005.The training parameter of the convolutional neural networks model includes batchsize parameters;The batchsize ginsengs
Number is 6-24;In the convolutional neural networks model, Softmax output layers are provided with 3-10 output channel.The softmax
Output layer is two graders.
Embodiment 2
The recognition methodss that the present embodiment is provided are carrying out obstacle recognition using convolutional neural networks CNN;In image
There is in terms of identification extraordinary effect, below artificial neural network and convolutional neural networks CNN are illustrated:
Artificial neural network is made up of countless neuronal cells, and each neuronal cell is as shown in figure 1, Fig. 1 is god
Jing meta structure models, A=X1W1+X2W2+ ...+XnWn, output Z=F (A);
Wherein, a neuron can have n input Xi, and each input has a Wi, F (x) swashing for neurocyte
Function living, if excitation value exceedes threshold value, exports 1, otherwise exports 0.
And the neuron of layer upon layer constitutes artificial neural network, per layer has N number of neuron, by full connection or portion
The method for dividing connection, the output of preceding layer are the input of later layer, and the model framework is artificial neural network, and works as the number of plies
More to a certain extent, then become deep neural network.
CNN convolutional neural networks are the one kind in deep neural network.One simple convolutional neural networks such as Fig. 2 institute
Show, Fig. 2 is convolutional neural networks topology, in figure, 1 represents that the first convolutional calculation, 2 represent that first time secondary acquisition, 3 represent second
Secondary convolutional calculation, 4 represent that second secondary acquisition, 5 represent full connection tiling characteristic pattern;11 represent input 32*32 image, 12
Represent that 6 5*5 convolution operators obtain 6 28*28 characteristic patterns, 13 expressions and obtain sample graph 14-14,14 16 10*10 spies of expression
Levy figure, 15 16 5*5 characteristic patterns of expression, 16 expression outputs;Logical Gaussian connections;Concrete principle is as follows:
Input 32*32 images, (strengthen former feature, and reduce by the convolution operator that a convolutional layer is 6 5*5
Noise), the characteristic pattern of 6 28*28 is obtained, 6 14*14 characteristic patterns are obtained by a down-sampling then.Rolled up by one again
Product and down-sampling (typically collocation occurs two-by-two), obtain the characteristic pattern of 16 5*5, then pave characteristic pattern by full connection, become
For 120 to 84 features, N number of output is obtained by full connection again then, the present embodiment N takes 10.
Convolutional neural networks it is advantageous that weights are shared, and reduce training parameter, become the structure of the network of complexity
It is simpler.And Googlenet is one type network structure;It is highly effective in terms of image recognition, the accuracy rate of identification
From 70% transient rise to 84%.
As shown in Fig. 2 Fig. 2 is Googlenet network structures, the Googlenet network structures that the present embodiment is provided, including
Convolutional layer, down-sampling layer, Softmax output layers.
Output at least only needs to three outputs, at most needs ten to export, is moving object due to what is processed based on bionical eye system
Body, moving person are substantially people, and the object of other motions is simultaneously few, it is impossible to realize subdivision, therefore using three outputs;Select sample
Originally it is trained;
As shown in figure 3, Fig. 3 is distributed for the training error of sample non-dispersive formula, Fig. 3 is to train the error of 400 times to be distributed;
Until to 400 times, sample loss still has very big fluctuation, illustrates that training effect is poor after training 200 times, therefore
The method is infeasible, that is, when training, sample is chosen and needs more to disperse, i.e., human body too can not collect as the sample of barrier
In, can otherwise cause training not good,
As shown in figure 4, Fig. 4 is to change the curve of error change that learning rate trains 400 times, illustrate learning rate also to training
Also very crucial, in the sample training of non-dispersive formula, learning rate is too high to cause training to be absorbed in locally optimal solution.And for point
Scattered style sheet, learning rate are higher, and convergence rate is faster, but its error robustness is poorer, therefore by above-mentioned analysis, have modified original
This training parameter, is that Learning Rate are changed to 0.001 by learning rate, although convergence rate is reduced, but as long as by increasing
Plus frequency of training can just reach due effect.
After it have selected distributing sample distribution and relatively low learning rate, can obtain being illustrated in fig. 6 shown below curve of error.
Fig. 5 is 400 curve of error distributions of training,
Be compared to two width figures, it can be seen that carry out error and constantly stably decline, though have fluctuation, but still receive scope it
It is interior.Note its vertical coordinate scope in relatively Fig. 3, the fluctuation in Fig. 3 up to more than ten and minimum also have 2, and Fig. 5 its error
Fluctuate and not less than 1, illustrate that the method is effective and feasible.
3 figure of the above is all that picture is analyzed as sample using on IMAGENET, the dispersion of above analysis verification sample and
Suitable learning rate has very big relation to training effect.
Googlenet is despite a convolutional neural networks, shared by weights, makes network structure become simpler,
But it is still a deep learning framework, and its network architecture parameters and training parameter are all a very big data volumes;Simplify net
Network structure, carries out computing, training using GPU and recognizes, improve training and recognition speed.
The present embodiment improves network structure by reducing batchsize, that is, the number of picture is input into while reducing the number of plies
Mesh, it is specific as follows:
Select identical training sample and test sample to be tested, only change batchsize parameters.
Test result is as follows:
Test 1:Batchsize is 6, and frequency of training is 800, test sample;
Number is 271, recognizes that positive exact figures are:251, accuracy rate is 92.6%;
Test 2:Batchsize is 12, and frequency of training is 400, test specimens;
This number is 271, recognizes that positive exact figures are:241, accuracy rate is 88.9%;
Test 3:Batchsize is 18, and frequency of training is 400, test specimens;
This number is 271, recognizes that positive exact figures are:250, accuracy rate is 92.5%;
Test 4:Batchsize is 18, and frequency of training is 800, test specimens;
This number is 271, recognizes that positive exact figures are:253, accuracy rate is 93.5%;
Test 5:Batchsize is 24, and frequency of training is 400, test specimens;
This number is 271, recognizes that positive exact figures are 240, accurate 88.5%;
From above it is each test in can be seen that come batchsize for accuracy impact less, it can also be seen that to train time
The accuracy rate of number 400 and 800 is also more or less the same,
Fig. 6 is 400 curve of error of training, from curve of error, trains already close to saturation, therefore training time later
Number can be controlled in 400-800.Comprehensively it is analyzed above, at the beginning of batchsize, is set to 12.The method greatly simplifies googlenet nets
Network, reduces data volume so that the depth network model can realize GPU computings, improve rapidly training speed.Will be heavy multiple
Miscellaneous calculating gives GPU process, and program remainder by CPU process, so as to realize efficient performance, compared to it is traditional only
CPU process, substantially reduces cycle of training, construction cycle.The computing that CPU needs more than ten hour is depended merely on, in GPU ten rather
Clock is it is achieved that or even faster.But this needs the performance depending on equipment, the size of video memory, therefore the size of batchsize
Also may require that appropriate adjustment.
Due to the use of CPU, the shortening of cycle of training and recognition speed quickening, bionical eye system can be for small part
Data can be trained in real time, i.e., after bionical eye system is locked to the object for moving, will can be somebody's turn to do after determining the label of object
Image after moving object carries out training identification in real time, improves accuracy rate.
The human body based on bionical eye system that the present embodiment is provided is as obstacle recognition method, specific as follows:
The image for selecting the bionical eye system of substantial amounts of reality to be identified, carries out labeling and then is trained again,
Actual to obtain a large amount of samples for obtaining people during training set, other barriers take second place.I.e. human body as barrier and other
Data sample quantity needs are similar, and softmax is revised as two classification, i.e. people and inhuman two class, as a result as follows:
(1) people's sample:1000, inhuman sample:100, frequency of training:400, test sample number:418, recognize positive exact figures:
392, accuracy rate 93.7%
(2) people's sample:300, inhuman sample:300, frequency of training:600, test sample number:418, recognize positive exact figures:
411, accuracy rate is 98.3%
(3) people's sample:1000, inhuman sample number:300, frequency of training:600, test sample number:418, recognize positive exact figures:
403, accuracy rate:96.3%
By above test analysis:Can obtain that quantity between each exemplar is similar can to improve accuracy of identification.Overcome motion
Ghost problem;I.e. deep learning neural network model is affected less by motion ghost, and correctly can be recognized.
Due to bionic eye system identification is dynamic object, i.e. video flowing, and the picture that the object is represented in video flowing has
A lot, therefore recognition methodss are different from the single picture of usual identification, the moving object is distinguished generally by identification plurality of pictures
Body is so as to improve after precision, i.e. lock motion object, rapid to collect N pictures (N>=10), by N pictures each label
Softmax output valves are added, and the object is then identified as the label of maximum representative.
Human body is identified by the present embodiment as barrier, using distributing sample, relatively low learning rate, from
And avoid the situation for causing that local optimum is absorbed in due to the unstable of loss function, and each exemplar quantity is close,
Ensure that accuracy of identification;And in terms of network structure, be input into simultaneously picture how much i.e. batchsize by reducing, reduce network
Structured data amount, so as to reach the purpose that can run GPU, accelerates training and the speed for recognizing.Finally in human body as obstacle
Preferable accuracy is obtained in the identification of thing, if obtaining more samples, enters row label, trained, then constantly can also carry
High precision precision, until saturation, due to the presence of GPU, this process will become very rapid.If Fig. 8 is with current network
A design sketch obtaining of identification.
Fig. 8-1,8-2,8-3 are to be identified the effect that obtains, and its tool by two sorter networks i.e. people and inhuman two class
Body numerical value.
Wherein, the parameter of Fig. 8-1 is:Predicted class1is#0;[0.99799794,0.00200206];
0.99799794 accuracy rate behaved for identification target, 0.00200206 is the inhuman accuracy rate of identification target.
The parameter of Fig. 8-2 is:Predicted class1is#1;[0.98484403,0.01515592];
The parameter of Fig. 8-3 is:Predicted class1is#0;[1.231197e-04,9.998768e-01];
Due to being two classification, i.e. only two classes, people and other, wherein 0 represents that for people, 1 represents other, from softmax
Output valve can be seen that two Classification and Identification accuracy have reached more than 90%, although there may be training sample and test sample excessively
Dull or similar problem, but the precision has been described that this model can accomplish good effect.
The recognition methodss that the present embodiment is provided are by combining the bionical eye system human body in practical application as barrier
Identifying system, is improved to googlenet models, and the googlenet models after lot of experiment validation, improvement make experiment
Effect has reached good effect, discrimination is improve a lot.Illustrate the truly feasible reliability of the innovatory algorithm.
Finally illustrate, above example is only unrestricted to illustrate technical scheme, although with reference to compared with
Good embodiment has been described in detail to the present invention, it will be understood by those within the art that, can be to the skill of the present invention
Art scheme is modified or equivalent, and without deviating from the objective and scope of the technical program, which all should be covered in the present invention
Protection domain in the middle of.
Claims (7)
1. a kind of obstacle recognition method of utilization convolutional neural networks CNN, it is characterised in that:Comprise the following steps:
S1:Set up convolutional neural networks model;
S2:Select dispersion data sample;
S3:The network architecture parameters and training parameter of convolutional neural networks are set;
S4:Dispersion data sample is input in convolutional neural networks model and is processed;
S5:Classification results of the output convolutional neural networks model to dispersion data sample.
2. as claimed in claim 1 using the obstacle recognition method of convolutional neural networks CNN, it is characterised in that:The volume
Product neural network model includes convolutional layer, down-sampling layer and Softmax output layers;
The convolutional layer, for the order of magnitude of network parameter is obtained by convolutional calculation;
The down-sampling layer, for obtaining the sub-sample of image;
The Softmax output layers, for obtaining the residual error of output layer and the residual error in intermediate layer;The residual error of the output layer is defeated
Going out value is worth error amount with category, and the residual error in the intermediate layer is the weighted sum of next layer of residual error.
3. as claimed in claim 1 using the obstacle recognition method of convolutional neural networks CNN, it is characterised in that:The volume
Processing procedure of the product neural network model to dispersion data sample, comprises the following steps that:
S21:The data sample of input is calculated into sampling characteristic pattern by the convolution operator of convolutional layer;
S22:Sampling characteristic pattern is calculated into double sampling characteristic pattern by down-sampling layer;
S23:Double sampling characteristic pattern is obtained being connected entirely by convolution and down-sampling and paves characteristic pattern,
S24:Characteristic pattern is paved in the full connection of output.
4. as claimed in claim 1 using the obstacle recognition method of convolutional neural networks CNN, it is characterised in that:The volume
The training parameter of product neural network model includes learning rate;The learning rate Learning Rate are set to 0.001-0.005.
5. as claimed in claim 1 using the obstacle recognition method of convolutional neural networks CNN, it is characterised in that:The volume
The training parameter of product neural network model includes batchsize parameters;The batchsize parameters are 6-24.
6. as claimed in claim 1 using the obstacle recognition method of convolutional neural networks CNN, it is characterised in that:The volume
In product neural network model, Softmax output layers are provided with 3-10 output channel.
7. as claimed in claim 1 using the obstacle recognition method of convolutional neural networks CNN, it is characterised in that:It is described
Softmax output layers are two graders.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610804686.9A CN106570516A (en) | 2016-09-06 | 2016-09-06 | Obstacle recognition method using convolution neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610804686.9A CN106570516A (en) | 2016-09-06 | 2016-09-06 | Obstacle recognition method using convolution neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106570516A true CN106570516A (en) | 2017-04-19 |
Family
ID=58531657
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610804686.9A Pending CN106570516A (en) | 2016-09-06 | 2016-09-06 | Obstacle recognition method using convolution neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106570516A (en) |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107944386A (en) * | 2017-11-22 | 2018-04-20 | 天津大学 | Visual scene recognition methods based on convolutional neural networks |
CN108038466A (en) * | 2017-12-26 | 2018-05-15 | 河海大学 | Multichannel human eye closure recognition methods based on convolutional neural networks |
CN108805882A (en) * | 2018-05-29 | 2018-11-13 | 杭州视氪科技有限公司 | A kind of water surface and puddle detection method |
CN108921119A (en) * | 2018-07-12 | 2018-11-30 | 电子科技大学 | A kind of barrier real-time detection and classification method |
CN109463809A (en) * | 2018-11-27 | 2019-03-15 | 浙江理工大学 | A kind of production method of personalized fit brassiere |
CN110646574A (en) * | 2019-10-08 | 2020-01-03 | 张家港江苏科技大学产业技术研究院 | Unmanned ship-based water quality conductivity autonomous detection system and method |
CN110889319A (en) * | 2019-09-04 | 2020-03-17 | 昆山研祥智能科技有限公司 | Road condition detection method and system |
CN111191559A (en) * | 2019-12-25 | 2020-05-22 | 国网浙江省电力有限公司泰顺县供电公司 | Overhead line early warning system obstacle identification method based on time convolution neural network |
CN111967434A (en) * | 2020-08-31 | 2020-11-20 | 湖北科技学院 | Machine vision anti-pinch system based on deep learning |
CN112463643A (en) * | 2020-12-16 | 2021-03-09 | 郑州航空工业管理学院 | Software quality prediction method |
CN112656431A (en) * | 2020-12-15 | 2021-04-16 | 中国科学院深圳先进技术研究院 | Electroencephalogram-based attention recognition method and device, terminal equipment and storage medium |
CN112669292A (en) * | 2020-12-31 | 2021-04-16 | 上海工程技术大学 | Method for detecting and classifying defects on painted surface of aircraft skin |
CN113449569A (en) * | 2020-03-27 | 2021-09-28 | 威海北洋电气集团股份有限公司 | Mechanical signal health state classification method and system based on distributed deep learning |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104809443A (en) * | 2015-05-05 | 2015-07-29 | 上海交通大学 | Convolutional neural network-based license plate detection method and system |
CN105590102A (en) * | 2015-12-30 | 2016-05-18 | 中通服公众信息产业股份有限公司 | Front car face identification method based on deep learning |
CN105912611A (en) * | 2016-04-05 | 2016-08-31 | 中国科学技术大学 | CNN based quick image search method |
CN105913087A (en) * | 2016-04-11 | 2016-08-31 | 天津大学 | Object identification method based on optimal pooled convolutional neural network |
WO2016168235A1 (en) * | 2015-04-17 | 2016-10-20 | Nec Laboratories America, Inc. | Fine-grained image classification by exploring bipartite-graph labels |
-
2016
- 2016-09-06 CN CN201610804686.9A patent/CN106570516A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2016168235A1 (en) * | 2015-04-17 | 2016-10-20 | Nec Laboratories America, Inc. | Fine-grained image classification by exploring bipartite-graph labels |
CN104809443A (en) * | 2015-05-05 | 2015-07-29 | 上海交通大学 | Convolutional neural network-based license plate detection method and system |
CN105590102A (en) * | 2015-12-30 | 2016-05-18 | 中通服公众信息产业股份有限公司 | Front car face identification method based on deep learning |
CN105912611A (en) * | 2016-04-05 | 2016-08-31 | 中国科学技术大学 | CNN based quick image search method |
CN105913087A (en) * | 2016-04-11 | 2016-08-31 | 天津大学 | Object identification method based on optimal pooled convolutional neural network |
Non-Patent Citations (2)
Title |
---|
卢宏涛等: "深度卷积神经网络在计算机视觉中的应用研究综述", 《数据采集与处理》 * |
石祥滨等: ""基于深度学习混合模型迁移学习的图像分类"", 《系统仿真学报》 * |
Cited By (22)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107944386A (en) * | 2017-11-22 | 2018-04-20 | 天津大学 | Visual scene recognition methods based on convolutional neural networks |
CN107944386B (en) * | 2017-11-22 | 2019-11-22 | 天津大学 | Visual scene recognition methods based on convolutional neural networks |
CN108038466A (en) * | 2017-12-26 | 2018-05-15 | 河海大学 | Multichannel human eye closure recognition methods based on convolutional neural networks |
CN108038466B (en) * | 2017-12-26 | 2021-11-16 | 河海大学 | Multi-channel human eye closure recognition method based on convolutional neural network |
CN108805882B (en) * | 2018-05-29 | 2021-09-03 | 杭州视氪科技有限公司 | Water surface and water pit detection method |
CN108805882A (en) * | 2018-05-29 | 2018-11-13 | 杭州视氪科技有限公司 | A kind of water surface and puddle detection method |
CN108921119B (en) * | 2018-07-12 | 2021-10-26 | 电子科技大学 | Real-time obstacle detection and classification method |
CN108921119A (en) * | 2018-07-12 | 2018-11-30 | 电子科技大学 | A kind of barrier real-time detection and classification method |
CN109463809A (en) * | 2018-11-27 | 2019-03-15 | 浙江理工大学 | A kind of production method of personalized fit brassiere |
CN110889319A (en) * | 2019-09-04 | 2020-03-17 | 昆山研祥智能科技有限公司 | Road condition detection method and system |
CN110646574A (en) * | 2019-10-08 | 2020-01-03 | 张家港江苏科技大学产业技术研究院 | Unmanned ship-based water quality conductivity autonomous detection system and method |
CN110646574B (en) * | 2019-10-08 | 2022-02-08 | 张家港江苏科技大学产业技术研究院 | Unmanned ship-based water quality conductivity autonomous detection system and method |
CN111191559A (en) * | 2019-12-25 | 2020-05-22 | 国网浙江省电力有限公司泰顺县供电公司 | Overhead line early warning system obstacle identification method based on time convolution neural network |
CN111191559B (en) * | 2019-12-25 | 2023-07-11 | 国网浙江省电力有限公司泰顺县供电公司 | Overhead line early warning system obstacle recognition method based on time convolution neural network |
CN113449569A (en) * | 2020-03-27 | 2021-09-28 | 威海北洋电气集团股份有限公司 | Mechanical signal health state classification method and system based on distributed deep learning |
CN113449569B (en) * | 2020-03-27 | 2023-04-25 | 威海北洋电气集团股份有限公司 | Mechanical signal health state classification method and system based on distributed deep learning |
CN111967434A (en) * | 2020-08-31 | 2020-11-20 | 湖北科技学院 | Machine vision anti-pinch system based on deep learning |
CN111967434B (en) * | 2020-08-31 | 2023-04-07 | 湖北科技学院 | Machine vision anti-pinch system based on deep learning |
CN112656431A (en) * | 2020-12-15 | 2021-04-16 | 中国科学院深圳先进技术研究院 | Electroencephalogram-based attention recognition method and device, terminal equipment and storage medium |
CN112463643A (en) * | 2020-12-16 | 2021-03-09 | 郑州航空工业管理学院 | Software quality prediction method |
CN112669292A (en) * | 2020-12-31 | 2021-04-16 | 上海工程技术大学 | Method for detecting and classifying defects on painted surface of aircraft skin |
CN112669292B (en) * | 2020-12-31 | 2022-09-30 | 上海工程技术大学 | Method for detecting and classifying defects on painted surface of aircraft skin |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106570516A (en) | Obstacle recognition method using convolution neural network | |
CN111259930B (en) | General target detection method of self-adaptive attention guidance mechanism | |
CN103890781B (en) | Retina encoder for machine vision | |
CN104732240B (en) | A kind of Hyperspectral imaging band selection method using neural network sensitivity analysis | |
Mendoza et al. | Interval type-2 fuzzy logic and modular neural networks for face recognition applications | |
CN106781489B (en) | A kind of road network trend prediction method based on recurrent neural network | |
CN107273845A (en) | A kind of facial expression recognizing method based on confidence region and multiple features Weighted Fusion | |
CN106951825A (en) | A kind of quality of human face image assessment system and implementation method | |
CN108509976A (en) | The identification device and method of animal | |
CN106980858A (en) | The language text detection of a kind of language text detection with alignment system and the application system and localization method | |
CN108399248A (en) | A kind of time series data prediction technique, device and equipment | |
CN106023195A (en) | BP neural network image segmentation method and device based on adaptive genetic algorithm | |
CN106446942A (en) | Crop disease identification method based on incremental learning | |
CN107341506A (en) | A kind of Image emotional semantic classification method based on the expression of many-sided deep learning | |
CN108615071A (en) | The method and device of model measurement | |
CN108921285A (en) | Single-element classification method in sequence based on bidirectional valve controlled Recognition with Recurrent Neural Network | |
CN105913450A (en) | Tire rubber carbon black dispersity evaluation method and system based on neural network image processing | |
CN110135502A (en) | A kind of image fine granularity recognition methods based on intensified learning strategy | |
CN109657582A (en) | Recognition methods, device, computer equipment and the storage medium of face mood | |
CN109903339B (en) | Video group figure positioning detection method based on multi-dimensional fusion features | |
CN104103033A (en) | Image real-time processing method | |
CN112037179B (en) | Method, system and equipment for generating brain disease diagnosis model | |
CN108447057A (en) | SAR image change detection based on conspicuousness and depth convolutional network | |
CN106909938A (en) | Viewing angle independence Activity recognition method based on deep learning network | |
CN108416458A (en) | A kind of tunnel rich water rock mass Synthetic Geological Prediction Ahead of Construction method based on BP neural network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170419 |