CN104268524A - Convolutional neural network image recognition method based on dynamic adjustment of training targets - Google Patents
Convolutional neural network image recognition method based on dynamic adjustment of training targets Download PDFInfo
- Publication number
- CN104268524A CN104268524A CN201410492769.XA CN201410492769A CN104268524A CN 104268524 A CN104268524 A CN 104268524A CN 201410492769 A CN201410492769 A CN 201410492769A CN 104268524 A CN104268524 A CN 104268524A
- Authority
- CN
- China
- Prior art keywords
- layer
- convolutional neural
- output
- training
- neural networks
- 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
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
Abstract
The invention discloses a convolutional neural network image recognition method based on dynamic adjustment of training targets. The method comprises the three steps of training model building, data registration, and recognition and matching. In the training model building process, training samples are input into a convolutional neural network model, the mean value of output vectors of output layers of the same class of samples is worked out to serve as the target vector of the current round of training, and then the error between the output vector of each sample and the target vector is reversely transmitted to adjust the parameters of a convolutional neural network till the training requirement is met; in the data registration process, sample images needing to be registered are input into the convolutional neural network model, and output results of the output layers serve as feature vectors to be stored; in the recognition and matching process, sample images needing to be recognized and matched are input into the convolutional neural network, output results of the output layers serve as feature vectors, and then matching is carried out on the feature vectors and registered feature vectors so as to provide the judgment result of recognition and matching.
Description
Technical field
The present invention relates to image identification technical field, especially based on the image identification technical field of convolutional neural networks.
Background technology
At present, the output layer of known convolutional neural networks is classification coding.This just requires, after model training terminates, the image to be identified used a model needs and the sample of training pattern has identical classification, and the classification number that model can identify also is fixing.
Such convolutional neural networks is applicable in other field of image recognition of fixed class, and, in other field of image recognition of many fixed class, achieve good effect, as hand-written Letter identification, Handwritten Digit Recognition etc.
But, also have many field of image recognition, the particularly field of image recognition of biological characteristic, as recognition of face, fingerprint recognition, hand vein recognition etc., after model training, the image pattern that identifies and the sample of model training is needed to belong to a different category, and, need the classification number that identifies to change along with the data variation registered in use procedure.This just makes known convolutional neural networks can not be used for these field of image recognition.
Summary of the invention
The field of image recognition of non-limiting classification can not be used for overcome existing convolutional neural networks model, the present invention proposes a kind of image-recognizing method of the convolutional neural networks based on dynamic conditioning training objective.After the method does not limit training pattern, the sample of discriminator and training sample have identical classification, and the classification number also not limiting classification immobilizes.Meanwhile, the method has higher discrimination.
The image-recognizing method of a kind of convolutional neural networks based on dynamic conditioning training objective proposed by the invention comprises S100 training pattern, S200 log-on data and S300 and identifies coupling three parts.Fig. 1 gives overview flow chart of the present invention.
The convolutional neural networks of dynamic conditioning training objective of the present invention, its structure is: comprise an input layer I, after input layer, is alternately distributed convolutional layer C
1, down-sampled layer S
1..., convolutional layer C
k, down-sampled layer S
k, convolutional layer C
k+1, an in the end convolutional layer C
k+1be several hidden layers H afterwards
1, H
2... H
n, an in the end hidden layer H
nafterwards, be output layer O.Fig. 2 gives the structural representation of convolutional neural networks.
Each node of described input layer corresponds to a pixel of input picture.Described input picture, can be the original image gathered, also can be through the image after filtering or normalization.
Every one deck of described convolutional layer comprises multiple characteristic pattern, the characteristic pattern of same layer measure-alike, and the pixel of each characteristic pattern, corresponding to the pixel set of some characteristic pattern respective window positions of specifying of front one deck.
Every one deck of described down-sampled layer comprises the characteristic pattern of multiple same size; The characteristic pattern of often opening of down-sampled layer corresponds to a characteristic pattern of front one deck convolutional layer; The pixel of the characteristic pattern of down-sampled layer corresponds to the sample area of front one deck individual features figure; All sample area do not have lap.
Each node of each node of described hidden layer and each node of front one deck and later layer is connected each other by the limit of Weight.
It is a real number that each node of described output layer exports.
Described training pattern S100 training pattern comprises the step of S110-S130.
S110 initialization convolutional neural networks, this process comprises the step of S111-S116.
S111 arranges training time end condition parameter.
S112 arranges number and the convolution window size of the characteristic pattern of each convolutional layer.
S113 arranges the number of the characteristic pattern of each down-sampled layer and down-sampled ratio.
S114 arranges the corresponding relation of each convolutional layer and front one deck characteristic pattern.
S115 arranges the node number of output layer, makes this node number identical with the binary-coded figure place of the classification of training sample.
The weight parameter on the limit in S116 initialization convolutional neural networks is random number.
S120 performs training process, repeats the step of S121-S128, till meeting training end condition.
Training image is inputted the input layer of convolutional neural networks by S121.
If S122 current layer is convolutional layer, then according to convolutional calculation formula, convolutional calculation is carried out to the view data in selected by front one deck, thus obtains the result of the characteristic pattern of convolutional layer.Its computation process is: chosen by front one deck view data and corresponding convolution kernel to calculate the convolution of respective regions according to formula (1); Then the convolution results calculated is drawn the result of calculation of Sigm function according to formula (2); Finally, according to formula (3), to respectively choosing the convolution results of image to sue for peace, obtain the corresponding convolution results of convolutional layer.Fig. 3 gives the schematic diagram of convolution process.
（1）。
（2）。
S = Σ S
i （3）。
If S123 current layer is down-sampled layer, then carry out down-sampled calculating according to the corresponding characteristic pattern of front one deck, thus obtain the characteristic pattern result of down-sampled layer.Formula (4) gives down-sampled computing formula.Wherein, C is the area data of last layer characteristic pattern, N
1with N
2be respectively the down-sampled multiple of last layer characteristic pattern in two dimensions of image.Fig. 4 gives the schematic diagram of down-sampled process.
（4）。
S124 repeated execution of steps S122 and step S123, till the result of calculation of last convolutional layer completes.
S125, according to the weight on the output data of the last node layer of hidden layer, the limit between hidden layer and last node layer, calculates the output data of hidden layer node.Repeat, till the output of last hidden layer has been calculated.Formula (5) gives the computing method of hidden layer node.Fig. 5 gives the schematic diagram of hidden layer structure.
y = Sigm (Σw
ix
i + b) （5）。
S126 is according to the Output rusults of last hidden layer, and the weight on limit between hidden layer and output layer, calculates the output vector of output layer.The computing formula of output layer is identical with formula (5).
S127 generates after output vector until all samples, calculates the average of each similar sample respectively, trains such other style object vector originally as epicycle.Formula (6) gives the computing method of object vector.
（6）。
S128 is by the object vector error of calculation of each training sample and its classification epicycle.
（7）。
The error that step S128 calculates is inputted convolutional neural networks by output layer by S129, successively adjusts the weight parameter in convolutional neural networks between each layer.
Described S200 log-on data comprises the step of S210-S220.
The convolutional neural networks that S210 will need the image input of registration to be trained by S100, through performing step S122-S126.
The Output rusults of the output layer that step S210 obtains by S220 is stored as the proper vector of this registered images.
Described S300 identifies that coupling comprises the step of S310-S330.
The convolutional neural networks that the input of the image of coupling to be identified is trained by S100 by S310, through performing step S122-S126.
The Output rusults of last hidden layer that step S310 obtains by S320 is as the proper vector of this matching image to be identified, and the proper vector of the log-on data stored with S200 calculates distance.
S330, according to the result of calculation of step S320, provides the court verdict identifying coupling.
The invention has the beneficial effects as follows, by the target in dynamically adjusting training process, after training is terminated, the proper vector that convolutional neural networks exports can make the distance of similar sample as far as possible little; This just makes, and in identification coupling, the distance of the sample of coupling to be identified and the proper vector of registered storage, as the foundation identifying coupling, thus can reach the object of classifying to non-limiting classification sample.Facts have proved, method proposed by the invention has higher discrimination.
Accompanying drawing explanation
Fig. 1 gives a kind of overview flow chart of image-recognizing method of the convolutional neural networks based on dynamic conditioning training objective.
Fig. 2 gives the structural representation of convolutional neural networks.
Fig. 3 gives the schematic diagram of convolution process.
Fig. 4 gives the schematic diagram of down-sampled process.
Fig. 5 gives the schematic diagram of hidden layer structure.
Claims (10)
1. based on an image-recognizing method for the convolutional neural networks of dynamic conditioning training objective, it is characterized in that: comprise training volume model, log-on data, identification coupling three parts.
2. convolutional neural networks according to claim 1, it is characterized in that: comprise an input layer, after input layer, be alternately distributed convolutional layer, down-sampled layer ..., down-sampled layer, convolutional layer, be in the end several hidden layers after a convolutional layer, in the end after a hidden layer, it is output layer.
3. input layer according to claim 2, is characterized in that: input layer corresponds to input picture, and each node corresponds to a pixel of input picture; Input picture can be original image, also can be through the image after filtering or normalization.
4. convolutional layer according to claim 2, it is characterized in that: each convolutional layer comprises the measure-alike of the characteristic pattern of multiple characteristic pattern same layer, and the pixel of each characteristic pattern, corresponding to the pixel set of some characteristic pattern respective window positions that front one deck is specified.
5. down-sampled layer according to claim 2, is characterized in that: each down-sampled layer comprised the characteristic pattern of a same size; The characteristic pattern of often opening of down-sampled layer corresponds to a characteristic pattern of front one deck convolutional layer; The pixel of the characteristic pattern of down-sampled layer corresponds to the sample area of front one deck individual features figure; All sample area do not have lap.
6. hidden layer according to claim 2, is characterized in that: each node of each node of this layer and each node of front one deck and later layer is connected each other by the limit of Weight.
7. output layer according to claim 2, is characterized in that: the output valve of each node of this layer is real number.
8. training pattern according to claim 1, is characterized in that: the parameter of initialization convolutional neural networks is random number; Training sample is inputted convolutional neural networks, the output vector of the output layer of similar training sample is averaged, as such sample epicycle object vector; The error of the epicycle object vector of each sample output vector and its generic is entered convolutional neural networks from output layer back transfer, the parameter of adjustment convolutional neural networks; Then, again training sample is inputted convolutional neural networks, repeat said process, till training reaches requirement.
9. log-on data according to claim 1, is characterized in that: the view data of registration is sent into convolutional neural networks, using the proper vector of the output vector of output layer as this sample, is stored in database.
10. identification coupling according to claim 1, it is characterized in that: view data to be identified is sent into convolutional neural networks, using the proper vector of the output vector of output layer as this sample, registered proper vector in this proper vector and database according to claim 9 is calculated distance, thus makes the court verdict identifying coupling.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410492769.XA CN104268524A (en) | 2014-09-24 | 2014-09-24 | Convolutional neural network image recognition method based on dynamic adjustment of training targets |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410492769.XA CN104268524A (en) | 2014-09-24 | 2014-09-24 | Convolutional neural network image recognition method based on dynamic adjustment of training targets |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN104268524A true CN104268524A (en) | 2015-01-07 |
Family
ID=52160044
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410492769.XA Pending CN104268524A (en) | 2014-09-24 | 2014-09-24 | Convolutional neural network image recognition method based on dynamic adjustment of training targets |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104268524A (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105096279A (en) * | 2015-09-23 | 2015-11-25 | 成都融创智谷科技有限公司 | Digital image processing method based on convolutional neural network |
| CN105138973A (en) * | 2015-08-11 | 2015-12-09 | 北京天诚盛业科技有限公司 | Face authentication method and device |
| CN105389596A (en) * | 2015-12-21 | 2016-03-09 | 长沙网动网络科技有限公司 | Method for enabling convolutional neural network to be suitable for recognition of pictures of various sizes |
| CN105469100A (en) * | 2015-11-30 | 2016-04-06 | 广东工业大学 | Deep learning-based skin biopsy image pathological characteristic recognition method |
| WO2016112797A1 (en) * | 2015-01-15 | 2016-07-21 | 阿里巴巴集团控股有限公司 | Method and device for determining image display information |
| WO2016183766A1 (en) * | 2015-05-18 | 2016-11-24 | Xiaogang Wang | Method and apparatus for generating predictive models |
| CN107609645A (en) * | 2017-09-21 | 2018-01-19 | 百度在线网络技术（北京）有限公司 | Method and apparatus for training convolutional neural networks |
| US10395141B2 (en) | 2017-03-20 | 2019-08-27 | Sap Se | Weight initialization for machine learning models |
| CN110288082A (en) * | 2019-06-05 | 2019-09-27 | 北京字节跳动网络技术有限公司 | Convolutional neural networks model training method, device and computer readable storage medium |
| WO2019200735A1 (en) * | 2018-04-17 | 2019-10-24 | 平安科技（深圳）有限公司 | Livestock feature vector acquisition method, apparatus, computer device and storage medium |
| CN110598504A (en) * | 2018-06-12 | 2019-12-20 | 北京市商汤科技开发有限公司 | Image recognition method and device, electronic equipment and storage medium |
| CN113486918A (en) * | 2021-05-19 | 2021-10-08 | 浙江大华技术股份有限公司 | Image identification method and device based on dynamic adjustment of feature vector distribution trend |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104268521A (en) * | 2014-09-23 | 2015-01-07 | 朱毅 | Image recognition method based on convolutional neural network in non-finite category |
-
2014
- 2014-09-24 CN CN201410492769.XA patent/CN104268524A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104268521A (en) * | 2014-09-23 | 2015-01-07 | 朱毅 | Image recognition method based on convolutional neural network in non-finite category |
Non-Patent Citations (3)
| Title |
|---|
| 肖柏旭: "基于卷积网络的人脸检测的研究与实现", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 许可: "卷积神经网络在图像识别上的应用的研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 陈先昌: "基于卷积神经网络的深度学习算法与应用研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2016112797A1 (en) * | 2015-01-15 | 2016-07-21 | 阿里巴巴集团控股有限公司 | Method and device for determining image display information |
| WO2016183766A1 (en) * | 2015-05-18 | 2016-11-24 | Xiaogang Wang | Method and apparatus for generating predictive models |
| CN105138973A (en) * | 2015-08-11 | 2015-12-09 | 北京天诚盛业科技有限公司 | Face authentication method and device |
| CN105138973B (en) * | 2015-08-11 | 2018-11-09 | 北京天诚盛业科技有限公司 | The method and apparatus of face authentication |
| CN105096279A (en) * | 2015-09-23 | 2015-11-25 | 成都融创智谷科技有限公司 | Digital image processing method based on convolutional neural network |
| CN105469100B (en) * | 2015-11-30 | 2018-10-12 | 广东工业大学 | Skin biopsy image pathological characteristics recognition methods based on deep learning |
| CN105469100A (en) * | 2015-11-30 | 2016-04-06 | 广东工业大学 | Deep learning-based skin biopsy image pathological characteristic recognition method |
| CN105389596B (en) * | 2015-12-21 | 2018-05-29 | 长沙网动网络科技有限公司 | The method that convolutional neural networks are suitable for identification sizes picture |
| CN105389596A (en) * | 2015-12-21 | 2016-03-09 | 长沙网动网络科技有限公司 | Method for enabling convolutional neural network to be suitable for recognition of pictures of various sizes |
| US10395141B2 (en) | 2017-03-20 | 2019-08-27 | Sap Se | Weight initialization for machine learning models |
| CN107609645A (en) * | 2017-09-21 | 2018-01-19 | 百度在线网络技术（北京）有限公司 | Method and apparatus for training convolutional neural networks |
| CN107609645B (en) * | 2017-09-21 | 2024-04-02 | 百度在线网络技术（北京）有限公司 | Method and apparatus for training convolutional neural network |
| WO2019200735A1 (en) * | 2018-04-17 | 2019-10-24 | 平安科技（深圳）有限公司 | Livestock feature vector acquisition method, apparatus, computer device and storage medium |
| CN110598504A (en) * | 2018-06-12 | 2019-12-20 | 北京市商汤科技开发有限公司 | Image recognition method and device, electronic equipment and storage medium |
| CN110598504B (en) * | 2018-06-12 | 2023-07-21 | 北京市商汤科技开发有限公司 | Image recognition method and device, electronic equipment and storage medium |
| CN110288082A (en) * | 2019-06-05 | 2019-09-27 | 北京字节跳动网络技术有限公司 | Convolutional neural networks model training method, device and computer readable storage medium |
| CN113486918A (en) * | 2021-05-19 | 2021-10-08 | 浙江大华技术股份有限公司 | Image identification method and device based on dynamic adjustment of feature vector distribution trend |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104268524A (en) | Convolutional neural network image recognition method based on dynamic adjustment of training targets | |
| CN104268521A (en) | Image recognition method based on convolutional neural network in non-finite category | |
| CN107529650B (en) | Closed loop detection method and device and computer equipment | |
| CN104615983B (en) | Activity recognition method based on recurrent neural network and human skeleton motion sequence | |
| CN108596053B (en) | Vehicle detection method and system based on SSD and vehicle posture classification | |
| CN108596258B (en) | Image classification method based on convolutional neural network random pooling | |
| CN106295637B (en) | A kind of vehicle identification method based on deep learning and intensified learning | |
| CN107203753B (en) | Action recognition method based on fuzzy neural network and graph model reasoning | |
| CN106778604B (en) | Pedestrian re-identification method based on matching convolutional neural network | |
| CN108108764B (en) | Visual SLAM loop detection method based on random forest | |
| CN104537647B (en) | A kind of object detection method and device | |
| CN108734208B (en) | Multi-source heterogeneous data fusion system based on multi-mode deep migration learning mechanism | |
| CN111160176B (en) | Fusion feature-based ground radar target classification method for one-dimensional convolutional neural network | |
| CN107274378B (en) | Image fuzzy type identification and parameter setting method based on fusion memory CNN | |
| CN113222011B (en) | Small sample remote sensing image classification method based on prototype correction | |
| CN107705322A (en) | Motion estimate tracking and system | |
| CN107563446A (en) | A kind of micro OS object detection method | |
| CN113705769A (en) | Neural network training method and device | |
| CN104008379A (en) | Object recognition method based on surf | |
| CN107832789B (en) | Feature weighting K nearest neighbor fault diagnosis method based on average influence value data transformation | |
| CN114693983B (en) | Training method and cross-domain target detection method based on image-instance alignment network | |
| CN103955950A (en) | Image tracking method utilizing key point feature matching | |
| CN109740672B (en) | Multi-stream feature distance fusion system and fusion method | |
| CN107564013B (en) | Scene segmentation correction method and system fusing local information | |
| CN111079930B (en) | Data set quality parameter determining method and device and electronic equipment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C41 | Transfer of patent application or patent right or utility model | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20160226 Address after: Siming Lake Road Xiamen city Fujian province 361004 of No. 138 room 2508 Applicant after: XIAMEN TONGYUAN MICRO INTELLIGENT TECHNOLOGY CO., LTD. Address before: Siming Lake Road Xiamen city Fujian province 361004 of No. 138 room 2508 Applicant before: Zhu Yi |
|
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20150107 |