CN108154066A - A kind of Three-dimensional target recognition method based on curvature feature recurrent neural network - Google Patents
A kind of Three-dimensional target recognition method based on curvature feature recurrent neural network Download PDFInfo
- Publication number
- CN108154066A CN108154066A CN201611096314.1A CN201611096314A CN108154066A CN 108154066 A CN108154066 A CN 108154066A CN 201611096314 A CN201611096314 A CN 201611096314A CN 108154066 A CN108154066 A CN 108154066A
- Authority
- CN
- China
- Prior art keywords
- curvature
- dimensional
- feature
- neural network
- brnn
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/60—Type of objects
- G06V20/64—Three-dimensional objects
- G06V20/647—Three-dimensional objects by matching two-dimensional images to three-dimensional objects
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Abstract
The present invention relates to image recognition technology, in order to effectively portray feature of the objective under different visual angles, for picture noise problem existing during Three-dimensional target recognition, it is proposed that a kind of Three-dimensional target recognition method based on curvature feature recurrent neural network.First, the present invention obtains the joint curvature of target three-dimensional by the local average Gaussian curvature and average mean curvature that calculate target three-dimensional, and pass through the curvature sketch that extraction joint curvature local maximum forms threedimensional model, convert input of the 360 ° of two-dimensional image sequences of generation as training recurrent neural network by the use of transmission projection;Secondly, by the use of forward-backward recutrnce neural network (BRNN) as threedimensional model various visual angles sequence signature learning method, the identification classification of correct probability maximum is acquired using softmax functions at softmax layers.The present invention can automatically extract the common trait of objective and two dimensional image, and preferable robustness and higher object recognition rate can be kept under the conditions of picture noise.
Description
Technical field
The present invention relates to image identification technical field, it is specifically a kind of based on curvature feature recurrent neural network three
Tie up target identification method.
Background technology
Three-dimensional target recognition refers to detect automatically from any given two dimensional image scene, positions, identifies specified mesh
The process of mark pattern is one of critical issue of computer vision research.It is three-dimensional with the continuous development of computer vision technique
Target identification is applied to the fields such as industrial detection, augmented reality and medical image more and more widely.But due to by illumination
The influence of the factors such as variation, picture noise and target occlusion, it is difficult to extract objective and its two dimensional image under different visual angles
Common trait, become Three-dimensional target recognition urgent problem to be solved.
The key of Three-dimensional target recognition is to find the two dimension expression of three dimensional object model, extracts objective and two dimensional image
Common trait.Existing Three-dimensional target recognition method mainly includes the method based on handmarking's point, based on geometric properties
Method and method based on deep learning etc..Method based on handmarking's point needs the feature in artificial initialization two dimensional image
Point, due to needing man-machine interactively, so such method does not have repeatability;Based on the method for geometric properties by extracting target
Center line skeleton, the information realizations target identification such as contour shape, but such method is identified in the case where image is there are noise
Effect is poor;Method based on deep learning is believed low-level multi-features into semanteme using deep neural network
The high-level feature of breath can well solve the picture noise problem of two dimensional image during Three-dimensional target recognition, but logical
The depth convolutional neural networks often used are beyond expression sequence properties, it is impossible to effectively portray objective under different visual angles
Feature.Therefore, there is an urgent need for propose it is a kind of in different visual angles image to the automatized three-dimensional target identification of picture noise problem robust
Method.
Invention content
The present invention seeks to can more effectively portray feature of the objective under different visual angles, feature extraction is reduced
Journey improves Three-dimensional target recognition accuracy rate, the present invention proposes a kind of based on curvature feature recurrence to the sensitivity of picture noise
The Three-dimensional target recognition method of neural network.
Present invention technical solution used for the above purpose is:It is a kind of based on curvature feature recurrent neural network
Three-dimensional target recognition method, includes the following steps:
Step 1:Calculate the joint curvature of target three-dimensionalExtract joint curvatureLocal maximum form three
The curvature sketch R of dimension moduleSketch;Again to the curvature sketch R of threedimensional modelSketch360 ° of two dimensions of generation are converted using transmission projection
Image Pm, wherein m=1,2 ..., 360;
Step 2:360 ° of two dimensional images are inputted into BRNN, carrying out study using multi-angle feature calculates it under various visual angles
Sequence properties;Identification classification in the softmax layers of correct probability maximum for acquiring using softmax functions sequence properties;Institute
BRNN is stated as forward-backward recutrnce neural network.
The joint curvature for calculating target three-dimensionalInclude the following steps:
IfIt is the normal vector that a bit (x, y, z) is given on target three-dimensional R;It enablesThen px,py,qx,qyIt is defined as
Calculate the mean-Gaussian curvature around the normal vector of every bit in 3 × 3 neighborhoods on threedimensional model RAveragely
It is worth curvature
Wherein,For average curvature matrix,
Trace () is the mark of matrix,Respectively p, q, px,py,qx,qyAverage value in 3 × 3 neighborhoods;
Define the joint curvature of target three-dimensional RFor:
It is described that 360 ° of two dimensional images are inputted into BRNN, it carries out study using multi-angle feature and calculates it under various visual angles
Sequence properties include the following steps:
Obtain the one-dimensional characteristic sequence T of 360 ° of two dimensional imagesS, s=1,2 ..., 360, then characteristic sequence TSIn BRNN i-th
The output of layer is divided into positive outputWith reversed outputAnd it is exported respectively with the forward direction of the upper sequences of this layer of BRNN
The reversed output of the lower sequences of this layer of BRNNAnd the positive output of last layer BRNNWith reversed outputJust like ShiShimonoseki
System:
Wherein,For the weight matrix of each outlet chamber, b is biasing, and tanh is
Neuron activation functions;
Then characteristic sequence TSIn total output O of BRNNs, the input I of as full articulamentum fcfcFor:
Wherein,Respectively positive output and reversely output are in the connection weight of full articulamentum;
Therefore, characteristic sequence TSIt is in the cumulative output of full articulamentum fcAs sequence properties.
The identification classification in the softmax layers of correct probability maximum for acquiring using softmax functions sequence properties,
Include the following steps:
Correct probability p (the C that recognition result is kth class are calculated using softmax functions at softmax layersk)
Wherein, C is total for identification classification, AkSequence properties for kth class objective are in the cumulative defeated of full articulamentum fc
Go out result;
When then acquiring loss function minimum value using maximum Likelihood, i.e. correct probability p (Ck) it is maximum when
Identify classification k:
Wherein, δ () is Kronecker functionR represents characteristic sequence TSJust
Really identification classification.
The invention has the advantages that and advantage:
1. the joint curvature sketch feature extracting method that the present invention designs, can automatically extract threedimensional model and two dimensional image
Common trait, and combine local average Gaussian curvature used in curvature and local average mean curvature and can effectively solve
Certainly picture noise problem.
2. present invention design multi-angle feature learning forward-backward recutrnce neural network, can consider threedimensional model polygonal simultaneously
Characteristic sequence under degree, can accurately identify objective in two dimensional image at any angle.
Description of the drawings
Fig. 1 is the method for the present invention flow chart;
Fig. 2 is the multi-angle feature learning forward-backward recutrnce neural network framework figure in the method for the present invention.
Specific embodiment
The present invention is described in further detail with reference to the accompanying drawings and embodiments.
The present invention is broadly divided into two parts, is as shown in Figure 1 the method for the present invention flow chart, the specific implementation process is as follows institute
It states.
Step 1:The joint curvature of target three-dimensional is calculated, and passes through extraction joint curvature local maximum and forms three-dimensional
The curvature sketch of model converts input of the 360 ° of two dimensional images of generation as training recurrent neural network by the use of transmission projection;
Step 1.1:IfIt is the normal vector that a bit (x, y, z) is given on threedimensional model.It enablesThen px,py,qx,qyIt is defined asThen threedimensional model
Gaussian curvature GKFor
GK=| C |,
Its curvature matrixThreedimensional model
Mean value curvature MKForTrace () is the mark of matrix.In order to eliminate influence of noise, the present invention calculates
Mean-Gaussian curvature on threedimensional model around the normal vector of every bit in its 3 × 3 neighborhoodWith average mean curvature
WhereinFor average curvature matrix,Respectively p, q, px,py,qx,qyAverage value in 3 × 3 neighborhoods.We define threedimensional model as a result,
Joint curvatureFor
Step 1.2:Extract joint curvatureLocal maximum point form the curvature sketch R of threedimensional model RSketch.It is logical
Cross perspective projection transformation, generation three dimensional curvature sketch RSketch360 ° of two-dimensional projection image Pm, m=1,2 ..., 360, as
The input of BRNN.
Step 2:The present invention is using a kind of depth recurrent neural network (DRNN) as curvature feature recognition methods, DRNN frames
Frame is as shown in Figure 2.Sequence properties of the threedimensional model under various visual angles are portrayed using multi-angle feature learning BRNN, in softmax
Layer acquires the identification classification of correct probability maximum using softmax functions.
Step 2.1:In order to portray the sequentiality of threedimensional model feature under different visual angles, threedimensional model is defined in various visual angles
Under one-dimensional characteristic sequence be TS, s=1,2 ..., 360, then characteristic sequence TSIt is divided into positive output in i-th layer of outputs of BRNNWith reversed outputIt is exported respectively with the forward direction of the upper sequences of this layer of BRNNThe lower sequences of this layer of BRNN it is reversed defeated
Go outAnd the positive output of last layer BRNNWith reversed outputThere is following relationship:
WhereinFor the weight matrix of each outlet chamber, b is bigoted, and tanh is god
Through first activation primitive;Then characteristic sequence TSIn total output O of BRNNs, the input I of as full articulamentum fcfcFor
Wherein,Respectively positive output and reversely output are in the connection weight of full articulamentum.
Step 2.2:Characteristic sequence TSIt is in the cumulative output of full articulamentum fcAs sequence properties.
Softmax layers calculate the correct probability p (C that recognition result is kth class using softmax functionsk)
Wherein C is total for identification classification, AkFor kth class objective sequence properties full articulamentum fc cumulative output
As a result.When then acquiring loss function minimum value using maximum Likelihood, i.e. correct probability p (Ck) it is maximum when identification
Classification k:
Wherein δ () is Kronecker functionR represents characteristic sequence TSCorrect knowledge
Other classification.
Claims (4)
- A kind of 1. Three-dimensional target recognition method based on curvature feature recurrent neural network, which is characterized in that include the following steps:Step 1:Calculate the joint curvature of target three-dimensionalExtract joint curvatureLocal maximum form three-dimensional mould The curvature sketch R of typeSketch;Again to the curvature sketch R of threedimensional modelSketch360 ° of two dimensional images of generation are converted using transmission projection Pm, wherein m=1,2 ..., 360;Step 2:360 ° of two dimensional images are inputted into BRNN, carrying out study using multi-angle feature calculates its sequence under various visual angles Attribute;Identification classification in the softmax layers of correct probability maximum for acquiring using softmax functions sequence properties;It is described BRNN is forward-backward recutrnce neural network.
- 2. according to the Three-dimensional target recognition method based on curvature feature recurrent neural network a kind of described in claim 1, feature It is, the joint curvature for calculating target three-dimensionalInclude the following steps:IfIt is the normal vector that a bit (x, y, z) is given on target three-dimensional R;It enablesThen px,py,qx,qyIt is defined asCalculate the mean-Gaussian curvature around the normal vector of every bit in 3 × 3 neighborhoods on threedimensional model RWith average mean curvatureWherein,For average curvature matrix, trace () It is the mark of matrix,Respectively p, q, px,py,qx,qyAverage value in 3 × 3 neighborhoods;Define the joint curvature of target three-dimensional RFor:
- 3. according to the Three-dimensional target recognition method based on curvature feature recurrent neural network a kind of described in claim 1, feature It is, it is described that 360 ° of two dimensional images are inputted into BRNN, it carries out study using multi-angle feature and calculates its sequence under various visual angles Column Properties include the following steps:Obtain the one-dimensional characteristic sequence T of 360 ° of two dimensional imagesS, s=1,2 ..., 360, then characteristic sequence TSI-th layer of BRNN's Output is divided into positive outputWith reversed outputAnd it is exported respectively with the forward direction of the upper sequences of this layer of BRNNThis layer The reversed output of the lower sequences of BRNNAnd the positive output of last layer BRNNWith reversed outputThere is following relationship:Wherein,For the weight matrix of each outlet chamber, b is biasing, and tanh is nerve First activation primitive;Then characteristic sequence TSIn total output O of BRNNs, the input I of as full articulamentum fcfcFor:Wherein,Respectively positive output and reversely output are in the connection weight of full articulamentum;Therefore, characteristic sequence TSIt is in the cumulative output of full articulamentum fcAs sequence properties.
- 4. according to the Three-dimensional target recognition method based on curvature feature recurrent neural network a kind of described in claim 1, feature It is, the identification classification in the softmax layers of correct probability maximum for acquiring using softmax functions sequence properties, packet Include following steps:Correct probability p (the C that recognition result is kth class are calculated using softmax functions at softmax layersk)Wherein, C is total for identification classification, AkFor kth class objective sequence properties full articulamentum fc cumulative output knot Fruit;When then acquiring loss function minimum value using maximum Likelihood, i.e. correct probability p (Ck) it is maximum when identification class Other k:Wherein, δ () is Kronecker functionR represents characteristic sequence TSCorrect identification Classification.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611096314.1A CN108154066B (en) | 2016-12-02 | 2016-12-02 | Three-dimensional target identification method based on curvature characteristic recurrent neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611096314.1A CN108154066B (en) | 2016-12-02 | 2016-12-02 | Three-dimensional target identification method based on curvature characteristic recurrent neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108154066A true CN108154066A (en) | 2018-06-12 |
CN108154066B CN108154066B (en) | 2021-04-27 |
Family
ID=62470160
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611096314.1A Active CN108154066B (en) | 2016-12-02 | 2016-12-02 | Three-dimensional target identification method based on curvature characteristic recurrent neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108154066B (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109166183A (en) * | 2018-07-16 | 2019-01-08 | 中南大学 | A kind of anatomic landmark point recognition methods and identification equipment |
CN109242955A (en) * | 2018-08-17 | 2019-01-18 | 山东师范大学 | A kind of workpiece manufacturing feature automatic identifying method and device based on single image |
CN109493354A (en) * | 2018-10-10 | 2019-03-19 | 中国科学院上海技术物理研究所 | A kind of target two-dimensional geometry Shape Reconstruction method based on multi-view image |
CN109496316A (en) * | 2018-07-28 | 2019-03-19 | 合刃科技(深圳)有限公司 | Image identification system |
CN110287783A (en) * | 2019-05-18 | 2019-09-27 | 天嗣智能信息科技(上海)有限公司 | A kind of video monitoring image human figure identification method |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101770566A (en) * | 2008-12-30 | 2010-07-07 | 复旦大学 | Quick three-dimensional human ear identification method |
CN104166842A (en) * | 2014-07-25 | 2014-11-26 | 同济大学 | Three-dimensional palm print identification method based on partitioning statistical characteristic and combined expression |
CN104463111A (en) * | 2014-11-21 | 2015-03-25 | 天津工业大学 | Three-dimensional face recognition method fused with multi-scale feature region curvatures |
CN105205478A (en) * | 2015-10-23 | 2015-12-30 | 天津工业大学 | 3-dimensional human face recognition method integrating anthropometry and curvelet transform |
US20160012597A1 (en) * | 2014-07-09 | 2016-01-14 | Nant Holdings Ip, Llc | Feature trackability ranking, systems and methods |
KR101592294B1 (en) * | 2014-09-03 | 2016-02-05 | 배재대학교 산학협력단 | Decimation Method For Complex Three Dimensional Polygonal Mesh Data |
US20160275653A1 (en) * | 2015-03-18 | 2016-09-22 | Kabushiki Kaisha Toshiba | Medical image processing apparatus and method |
CN106097431A (en) * | 2016-05-09 | 2016-11-09 | 王红军 | A kind of object global recognition method based on 3 d grid map |
-
2016
- 2016-12-02 CN CN201611096314.1A patent/CN108154066B/en active Active
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101770566A (en) * | 2008-12-30 | 2010-07-07 | 复旦大学 | Quick three-dimensional human ear identification method |
US20160012597A1 (en) * | 2014-07-09 | 2016-01-14 | Nant Holdings Ip, Llc | Feature trackability ranking, systems and methods |
CN104166842A (en) * | 2014-07-25 | 2014-11-26 | 同济大学 | Three-dimensional palm print identification method based on partitioning statistical characteristic and combined expression |
KR101592294B1 (en) * | 2014-09-03 | 2016-02-05 | 배재대학교 산학협력단 | Decimation Method For Complex Three Dimensional Polygonal Mesh Data |
CN104463111A (en) * | 2014-11-21 | 2015-03-25 | 天津工业大学 | Three-dimensional face recognition method fused with multi-scale feature region curvatures |
US20160275653A1 (en) * | 2015-03-18 | 2016-09-22 | Kabushiki Kaisha Toshiba | Medical image processing apparatus and method |
CN105205478A (en) * | 2015-10-23 | 2015-12-30 | 天津工业大学 | 3-dimensional human face recognition method integrating anthropometry and curvelet transform |
CN106097431A (en) * | 2016-05-09 | 2016-11-09 | 王红军 | A kind of object global recognition method based on 3 d grid map |
Non-Patent Citations (2)
Title |
---|
FENG LIU 等: ""Study on novel Curvature Features for 3D fingerprint recognition"", 《NEUROCOMPUTING》 * |
许俊峰: ""基于模型的任意视点下三维目标识别研究"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109166183A (en) * | 2018-07-16 | 2019-01-08 | 中南大学 | A kind of anatomic landmark point recognition methods and identification equipment |
CN109166183B (en) * | 2018-07-16 | 2023-04-07 | 中南大学 | Anatomical landmark point identification method and identification equipment |
CN109496316A (en) * | 2018-07-28 | 2019-03-19 | 合刃科技(深圳)有限公司 | Image identification system |
WO2020024079A1 (en) * | 2018-07-28 | 2020-02-06 | 合刃科技(深圳)有限公司 | Image recognition system |
CN109496316B (en) * | 2018-07-28 | 2022-04-01 | 合刃科技(深圳)有限公司 | Image recognition system |
CN109242955A (en) * | 2018-08-17 | 2019-01-18 | 山东师范大学 | A kind of workpiece manufacturing feature automatic identifying method and device based on single image |
CN109242955B (en) * | 2018-08-17 | 2023-03-24 | 山东师范大学 | Workpiece manufacturing characteristic automatic identification method and device based on single image |
CN109493354A (en) * | 2018-10-10 | 2019-03-19 | 中国科学院上海技术物理研究所 | A kind of target two-dimensional geometry Shape Reconstruction method based on multi-view image |
CN110287783A (en) * | 2019-05-18 | 2019-09-27 | 天嗣智能信息科技(上海)有限公司 | A kind of video monitoring image human figure identification method |
Also Published As
Publication number | Publication date |
---|---|
CN108154066B (en) | 2021-04-27 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Zhang et al. | Dense attention fluid network for salient object detection in optical remote sensing images | |
CN108154066A (en) | A kind of Three-dimensional target recognition method based on curvature feature recurrent neural network | |
CN107844795B (en) | Convolutional neural networks feature extracting method based on principal component analysis | |
CN109669049B (en) | Particle image velocity measurement method based on convolutional neural network | |
CN108038420B (en) | Human behavior recognition method based on depth video | |
CN107316031A (en) | The image characteristic extracting method recognized again for pedestrian | |
CN108596193B (en) | Method and system for building deep learning network structure aiming at human ear recognition | |
CN110991444B (en) | License plate recognition method and device for complex scene | |
CN110348383B (en) | Road center line and double line extraction method based on convolutional neural network regression | |
CN106919944A (en) | A kind of wide-angle image method for quickly identifying based on ORB algorithms | |
CN110163108A (en) | Robust sonar target detection method based on dual path Fusion Features network | |
CN112767279B (en) | Underwater image enhancement method for generating countermeasure network based on discrete wavelet integration | |
CN115063573A (en) | Multi-scale target detection method based on attention mechanism | |
CN107590831A (en) | A kind of solid matching method based on deep learning | |
CN104298974A (en) | Human body behavior recognition method based on depth video sequence | |
CN105513094A (en) | Stereo vision tracking method and stereo vision tracking system based on 3D Delaunay triangulation | |
CN108665484A (en) | A kind of dangerous source discrimination and system based on deep learning | |
CN110991274A (en) | Pedestrian tumbling detection method based on Gaussian mixture model and neural network | |
CN114926747A (en) | Remote sensing image directional target detection method based on multi-feature aggregation and interaction | |
CN111199558A (en) | Image matching method based on deep learning | |
CN111027586A (en) | Target tracking method based on novel response map fusion | |
CN111612825A (en) | Image sequence motion occlusion detection method based on optical flow and multi-scale context | |
CN107766810B (en) | Cloud and shadow detection method | |
Deng et al. | DRD-Net: Detail-recovery image deraining via context aggregation networks | |
Khan et al. | LRDNet: Lightweight LiDAR aided cascaded feature pools for free road space detection |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CB03 | Change of inventor or designer information |
Inventor after: Liang Wei Inventor after: Li Yang Inventor after: Zheng Meng Inventor after: Peng Shiwei Inventor before: Liang Wei Inventor before: Li Yang Inventor before: Zheng Meng Inventor before: Tan Jindong Inventor before: Peng Shiwei |
|
CB03 | Change of inventor or designer information |