CN104134062A - Vein recognition system based on depth neural network - Google Patents
Vein recognition system based on depth neural network Download PDFInfo
- Publication number
- CN104134062A CN104134062A CN201410402818.6A CN201410402818A CN104134062A CN 104134062 A CN104134062 A CN 104134062A CN 201410402818 A CN201410402818 A CN 201410402818A CN 104134062 A CN104134062 A CN 104134062A
- Authority
- CN
- China
- Prior art keywords
- neural network
- layer
- degree
- depth neural
- network model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a vein recognition system based on a depth neural network. The system relates to training of a depth neutral network model, registration of data and recognition and matching. According to the vein recognition system based on the depth neural network, the recognition and matching speed is increased, and meanwhile the high recognition rate is achieved.
Description
Technical field
The present invention relates to vein identification technology field, particularly the vein identification technology field based on degree of depth neural network.
Background technology
At present, the image of finger, palm or hand back vein is mainly obtained in vein identification by infrared C CD camera, and the filtering of uses advanced, image binaryzation, refinement means are extracted feature to digital picture, and the digital picture of vein is stored in computer system.When vein comparison, take in real time vein figure, extract feature, adopt complicated matching algorithm with being stored in main frame medium sized vein eigenwert comparison coupling, thereby individual is carried out to identity authentication, confirm identity.
Because the image of vein is infrared image, thereby there is the feature of the aspects such as signal to noise ratio (S/N ratio) is low, contrast is low.Current existing vein recognition system need to expend the longer processing time in extraction characteristic aspect, and has suitable difficulty.In addition, the current this vein recognition system that extracts feature based on methods such as filtering, binaryzation, refinements, in the process of coupling, the time that need to expend is also longer.
Summary of the invention
In view of the problem existing in vein identification, the present invention proposes a kind of vein recognition system based on degree of depth neural network, has evaded artificial design, has extracted the difficulty of vein recognition image feature, and can mate rapidly identifying.
Vein recognition system based on degree of depth neural network proposed by the invention, comprises S100 training pattern, S200 log-on data, S300 identification coupling three parts.Fig. 1 has provided the overview flow chart of this system.
Degree of depth neural network of the present invention, comprises an input layer, multiple hidden layer and an output layer.Wherein, a pixel of the corresponding input picture of each node of input layer is as input; Between any two nodes in adjacent two layers in model, all exist a band to have the limit of weight; Output layer has multiple nodes, and the output valve of these nodes forms output vector according to the order of specifying.Fig. 2 has provided the structural representation of degree of depth neural network.
In S100 training pattern part, comprise the step of S110-S140.
S110 is by the vein infrared image of a large amount of volunteers' of vein harvester collection human body target site (as finger, palm, the back of the hand etc.).
Corresponding target area carry out size, gray scale normalization in S120 cut-away view picture.
S130 sends the target area image after these a large amount of normalization (hereinafter to be referred as ROI image) into degree of depth neural network model and trains.The first stage of training successively carries out according to the order of input layer, hidden layer, output layer.S131-S136 has provided the process of one deck of training degree of depth neural network.
S131 initialization current layer (i layer) and the internodal weight of lower one deck (i+1 layer).
S132 is according to the set { P of the value of current layer node
i, weighted value set { W between current layer and next node layer
i (i+1)via the scrambler E of current layer and lower one deck
i (i+1)calculate the nodal value set { P of lower one deck
i+1.Wherein, scrambler E
i (i+1)both can be linear transformation, can be also nonlinear transformation.Formula (1) is scrambler conversion signal formula.
P
i+1 =
E i (i+1) (P
i ,W
i(i+1)) (1)。
S133 is by the nodal value set { P of lower one deck
i+1, weighted value set { W between current layer and next node layer
i (i+1)via the demoder D of current layer and lower one deck
(i+1) icalculate the decoding set { PD of current layer
i.Formula (2) is demoder conversion signal formula.
PD
i =
D (i+1)i (P
i+1 ,W
i(i+1)) (2)。
S134 is according to the worth set { P that calculates current layer node
iwith the decoding set { PD of current layer
ierror DEV
i.Formula (3) is error DEV
icomputing formula.
DEV
i = Σ(P
i - PD
i})
2 (3)。
S135 is according to DEV
iadjust the weighted value set { W between current layer and next node layer
i (i+1), and according to the nodal value set { P of the lower one deck of formula (1) renewal
i+1.
S136 repeats the process of S132-S135, until the frequency of training of current layer or DEV
itill reaching requirement.
S140 trains the degree of depth neural network model obtaining to carry out parameter tuning to S130.S141-S145 has provided the detailed process of parameter tuning.
S141 sends ROI image into S130 and trains the degree of depth neural network model obtaining, and successively according to formula (1), via the scrambler of each node layer, obtains the set { P of the value of the output node of output layer
out.
S142 will finally obtain the set { P of value of output node of output layer
outsuccessively according to formula (2), the reverse demoder via each node layer, until input layer obtains decoding set { PD}.
S143 calculates ROI image collection { ROI} and decoding set { the error DEV of PD} of input.Formula (4) is the computing formula of error DEV.
DEV = Σ(ROI - PD)
2 。 (4)。
S144, according to the DEV calculating, finely tunes the parameter of degree of depth neural network model.
S145 repeats the process of S141-S144, until tuning number of times or DEV reach requirement.
In the S200 log-on data stage, comprise S210-S230 step.
S210 gathers the vein infrared image at accredited personnel's human body target position by vein harvester.
Corresponding target area carry out size, gray scale normalization in S220 cut-away view picture.
The input layer of the degree of depth neural network model that S230 trains the input of ROI image, via degree of depth neural network model, the proper vector of then storage depth neural network output.
At S300 identification matching stage, comprise S310-S350 step.
S310 gathers the vein infrared image at personnel's to be identified human body target position by vein harvester.
Corresponding target area carry out size, gray scale normalization in S320 cut-away view picture.
The input layer of the degree of depth neural network model that S330 trains the input of ROI image, via degree of depth neural network model, obtains the proper vector of this vein image sample to be identified.
The proper vector of the proper vector that S340 obtains S330 and registered vein database is calculated distance.
If S350 distance value is less than the distance threshold training in training data, judge that the sample of coupling to be identified and selected registered vein sample come from same people's same collection position; Otherwise, judge that the two does not come from same people's same collection position.
The invention has the beneficial effects as follows, on the one hand, the operation of removed the front a large amount of image processing of tional identification matching process from, extracting feature, only need send into vein image degree of depth neural network, gets final product generating feature vector; On the other hand, in identification matching process, there is higher discrimination.
Brief description of the drawings
Fig. 1 has provided the overview flow chart of the vein recognition system based on degree of depth neural network.
Fig. 2 has provided the structural representation of degree of depth neural network.
Fig. 3 has provided the process flow diagram of training a layer depth neural network.
Fig. 4 has provided the process flow diagram of degree of depth neural network parameter tuning.
Claims (9)
1. the vein recognition system based on degree of depth neural network, is characterized in that: comprise training degree of depth neural network model, log-on data, identification coupling three parts.
2. degree of depth neural network model according to claim 1, is characterized in that: comprise an input layer, multiple hidden layer and an output layer.
3. degree of depth neural network model according to claim 1, is characterized in that: any two nodes that come from adjacent layer exist a band to have the limit of weight.
4. degree of depth neural network model according to claim 1, is characterized in that: between every two adjacent layers, have respectively a scrambler and a demoder.
5. encoder according to claim 4, is characterized in that, scrambler can be both linear transformation, can be also nonlinear transformation.
6. training degree of depth neural network model according to claim 1, is characterized in that: the vein image of training set is intercepted to target area; The region intercepting is carried out to the normalization of gray scale, area size's size successively; By a node input degree of depth neural network model of the input layer of the corresponding degree of depth neural network model of the pixel of the image after every width normalization; First training successively encodes according to the scrambler of equivalent layer according to the order of input layer, hidden layer, output layer, then the coding result of equivalent layer is decoded by the demoder of equivalent layer, adjust the weight of equivalent layer by calculating the error of the input of equivalent layer and the result of decoding; After the weight training of degree of depth neural network completes, carry out parameter tuning.
7. parameter tuning according to claim 6, is characterized in that, the degree of depth neural network model before parameter tuning is sent in the target area that the vein image of training set is intercepted, and via the scrambler of each layer of model, obtains the output vector of output layer; Then, by reverse the output vector of the output layer demoder via each layer of degree of depth neural network, obtain the decoded result of input layer; Finally, according to the error between the input of input layer and input layer decoded result, the parameter of model is carried out to tuning.
8. log-on data according to claim 1, is characterized in that, intercepts target area for the vein image of registration; The region intercepting is carried out to the normalization of gray scale, area size's size successively; Image after normalization is sent into the degree of depth neural network model training, the proper vector of storage depth neural network output.
9. identification coupling according to claim 1, it is characterized in that, the vein image of coupling to be identified is intercepted to target area, and it is carried out respectively to the normalization of gray scale, area size's size, then the image after normalization is sent into the degree of depth neural network model training, the proper vector of storing in the proper vector of degree of depth neural network output and log-on data process claimed in claim 8 is calculated to distance, finally judge identification matching result according to distance value.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410402818.6A CN104134062A (en) | 2014-08-18 | 2014-08-18 | Vein recognition system based on depth neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410402818.6A CN104134062A (en) | 2014-08-18 | 2014-08-18 | Vein recognition system based on depth neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN104134062A true CN104134062A (en) | 2014-11-05 |
Family
ID=51806736
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201410402818.6A Pending CN104134062A (en) | 2014-08-18 | 2014-08-18 | Vein recognition system based on depth neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN104134062A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105160361A (en) * | 2015-09-30 | 2015-12-16 | 东软集团股份有限公司 | Image identification method and apparatus |
CN105787497A (en) * | 2014-12-23 | 2016-07-20 | 阿里巴巴集团控股有限公司 | Account-stealing case analysis method and device |
CN106127217A (en) * | 2015-05-07 | 2016-11-16 | 西门子保健有限责任公司 | The method and system that neutral net detects is goed deep into for anatomical object for approximation |
CN108596246A (en) * | 2018-04-23 | 2018-09-28 | 浙江科技学院 | The method for building up of soil heavy metal content detection model based on deep neural network |
CN109074650A (en) * | 2016-05-25 | 2018-12-21 | 科磊股份有限公司 | It applies for semiconductor from input picture and generates through analog image |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2004102993A (en) * | 2003-08-11 | 2004-04-02 | Hitachi Ltd | Personal identification device and method |
CN101006923A (en) * | 2007-01-23 | 2007-08-01 | 天津理工大学 | Identity recognition instrument based on characteristics of subcutaneous vein of dorsum of hand and recognition method |
MY135197A (en) * | 2003-08-26 | 2008-02-29 | Hitachi Ltd | Personal identification device and method |
-
2014
- 2014-08-18 CN CN201410402818.6A patent/CN104134062A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2004102993A (en) * | 2003-08-11 | 2004-04-02 | Hitachi Ltd | Personal identification device and method |
MY135197A (en) * | 2003-08-26 | 2008-02-29 | Hitachi Ltd | Personal identification device and method |
CN101006923A (en) * | 2007-01-23 | 2007-08-01 | 天津理工大学 | Identity recognition instrument based on characteristics of subcutaneous vein of dorsum of hand and recognition method |
Non-Patent Citations (2)
Title |
---|
卓维: "深度神经网络的快速学习算法", 《嘉应学院学报》 * |
肖潇: "基于手指静脉的身份识别技术研究", 《中国优秀硕士学位论文全文数据库》 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105787497A (en) * | 2014-12-23 | 2016-07-20 | 阿里巴巴集团控股有限公司 | Account-stealing case analysis method and device |
CN106127217A (en) * | 2015-05-07 | 2016-11-16 | 西门子保健有限责任公司 | The method and system that neutral net detects is goed deep into for anatomical object for approximation |
CN106127217B (en) * | 2015-05-07 | 2019-11-05 | 西门子保健有限责任公司 | The method and system for going deep into neural network for approximation to detect for anatomical object |
CN105160361A (en) * | 2015-09-30 | 2015-12-16 | 东软集团股份有限公司 | Image identification method and apparatus |
CN109074650A (en) * | 2016-05-25 | 2018-12-21 | 科磊股份有限公司 | It applies for semiconductor from input picture and generates through analog image |
CN109074650B (en) * | 2016-05-25 | 2023-09-15 | 科磊股份有限公司 | Generating simulated images from input images for semiconductor applications |
CN108596246A (en) * | 2018-04-23 | 2018-09-28 | 浙江科技学院 | The method for building up of soil heavy metal content detection model based on deep neural network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110458844B (en) | Semantic segmentation method for low-illumination scene | |
Song et al. | Region-based quality estimation network for large-scale person re-identification | |
Zhang et al. | Non-rigid object tracking via deep multi-scale spatial-temporal discriminative saliency maps | |
CN108875732B (en) | Model training and instance segmentation method, device and system and storage medium | |
Yang et al. | Super normal vector for activity recognition using depth sequences | |
Chen et al. | Gait recognition based on improved dynamic Bayesian networks | |
CN107492121B (en) | Two-dimensional human body bone point positioning method of monocular depth video | |
US20170300744A1 (en) | Method and apparatus for determining identity identifier of face in face image, and terminal | |
CN104134062A (en) | Vein recognition system based on depth neural network | |
CN106096568A (en) | A kind of pedestrian's recognition methods again based on CNN and convolution LSTM network | |
Liu et al. | A cross-modal adaptive gated fusion generative adversarial network for RGB-D salient object detection | |
Ghazal et al. | Human posture classification using skeleton information | |
CN112580523A (en) | Behavior recognition method, behavior recognition device, behavior recognition equipment and storage medium | |
CN108875586B (en) | Functional limb rehabilitation training detection method based on depth image and skeleton data multi-feature fusion | |
CN110826447A (en) | Restaurant kitchen staff behavior identification method based on attention mechanism | |
Jalal et al. | Daily human activity recognition using depth silhouettes and transformation for smart home | |
CN102682452A (en) | Human movement tracking method based on combination of production and discriminant | |
CN105718873A (en) | People stream analysis method based on binocular vision | |
CN110570443B (en) | Image linear target extraction method based on structural constraint condition generation model | |
CN109035172A (en) | A kind of non-local mean Ultrasonic Image Denoising method based on deep learning | |
CN106815855A (en) | Based on the human body motion tracking method that production and discriminate combine | |
Uddin et al. | Human Activity Recognition via 3-D joint angle features and Hidden Markov models | |
CN104794446A (en) | Human body action recognition method and system based on synthetic descriptors | |
CN112733710A (en) | Method for training a neural network for irrigation water pressure control of an irrigation device | |
Zhou et al. | A study on attention-based LSTM for abnormal behavior recognition with variable pooling |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20141105 |
|
RJ01 | Rejection of invention patent application after publication |