Movatterモバイル変換

[0]ホーム
URL:

CN108154066B - A 3D Object Recognition Method Based on Curvature Feature Recurrent Neural Network - Google Patents

A 3D Object Recognition Method Based on Curvature Feature Recurrent Neural NetworkDownload PDF

Info

Publication number
CN108154066B
CN108154066BCN201611096314.1ACN201611096314ACN108154066BCN 108154066 BCN108154066 BCN 108154066BCN 201611096314 ACN201611096314 ACN 201611096314ACN 108154066 BCN108154066 BCN 108154066B
Authority
CN
China
Prior art keywords
dimensional
curvature
sequence
brnn
target
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201611096314.1A
Other languages
Chinese (zh)
Other versions
CN108154066A (en
Inventor
梁炜
李杨
郑萌
谈金东
彭士伟
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenyang Institute of Automation of CAS
Original Assignee
Shenyang Institute of Automation of CAS
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Shenyang Institute of Automation of CASfiledCriticalShenyang Institute of Automation of CAS
Priority to CN201611096314.1ApriorityCriticalpatent/CN108154066B/en
Publication of CN108154066ApublicationCriticalpatent/CN108154066A/en
Application grantedgrantedCritical
Publication of CN108154066BpublicationCriticalpatent/CN108154066B/en
Activelegal-statusCriticalCurrent
Anticipated expirationlegal-statusCritical

Links

Images

Classifications

Landscapes

Abstract

Translated fromChinese

本发明涉及图像识别技术，为了有效地刻画三维目标在不同视角下的特征，针对三维目标识别过程中存在的图像噪声问题，提出了一种基于曲率特征递归神经网络的三维目标识别方法。首先，本发明通过计算目标三维模型的局部平均高斯曲率和平均均值曲率得出目标三维模型的联合曲率，并通过提取联合曲率局部极大值构成三维模型的曲率草图，利用透射投影变换生成360°二维图像序列作为训练递归神经网络的输入；其次，利用双向递归神经网络(BRNN)作为三维模型多视角序列特征学习方法，在softmax层利用softmax函数求得正确概率最大的识别类别。本发明能够自动提取三维目标与二维图像的共同特征，能够在图像噪声条件下保持较好的鲁棒性和较高的目标识别率。

Figure 201611096314

The invention relates to an image recognition technology. In order to effectively describe the characteristics of a three-dimensional target under different viewing angles, a three-dimensional target recognition method based on a curvature feature recursive neural network is proposed for the image noise problem existing in the three-dimensional target recognition process. First, the present invention obtains the joint curvature of the target three-dimensional model by calculating the local average Gaussian curvature and the average mean curvature of the target three-dimensional model, and forms the curvature sketch of the three-dimensional model by extracting the local maximum value of the joint curvature, and uses the transmission projection transformation to generate 360° The two-dimensional image sequence is used as the input to train the recurrent neural network; secondly, the bidirectional recurrent neural network (BRNN) is used as the multi-view sequence feature learning method of the three-dimensional model, and the softmax function is used in the softmax layer to obtain the recognition category with the highest probability of correctness. The invention can automatically extract the common features of the three-dimensional target and the two-dimensional image, and can maintain good robustness and high target recognition rate under the condition of image noise.

Figure 201611096314

Description

Translated fromChinese
一种基于曲率特征递归神经网络的三维目标识别方法A 3D Object Recognition Method Based on Curvature Feature Recurrent Neural Network

技术领域technical field

本发明涉及图像识别技术领域，具体地说是一种基于曲率特征递归神经网络的三维目标识别方法。The invention relates to the technical field of image recognition, in particular to a three-dimensional target recognition method based on a curvature feature recursive neural network.

背景技术Background technique

三维目标识别是指从任意给定的二维图像场景中自动检测、定位、识别出指定目标模式的过程，是计算机视觉研究的关键问题之一。随着计算机视觉技术的不断发展，三维目标识别越来越广泛地应用于工业检测、增强现实和医学影像等领域。但是，由于受到光照变化、图像噪声和目标遮挡等因素的影响，难以提取三维目标及其在不同视角下二维图像的共同特征，成为三维目标识别亟待解决的问题。3D target recognition refers to the process of automatically detecting, locating, and recognizing the specified target pattern from any given 2D image scene. It is one of the key issues in computer vision research. With the continuous development of computer vision technology, 3D object recognition is more and more widely used in the fields of industrial inspection, augmented reality and medical imaging. However, due to the influence of factors such as illumination changes, image noise and target occlusion, it is difficult to extract the common features of 3D objects and 2D images from different perspectives, which has become an urgent problem to be solved in 3D object recognition.

三维目标识别的关键是找到三维目标模型的二维表达，提取三维目标和二维图像的共同特征。现有的三维目标识别方法主要包括基于人工标记点的方法、基于几何特征的方法和基于深度学习的方法等。基于人工标记点的方法需要人工初始化二维图像中的特征点，由于需要人工交互，所以此类方法不具有可重复性；基于几何特征的方法通过提取目标的中线骨架、轮廓形状等信息实现目标识别，但是此类方法在图像存在噪声的情况下识别效果较差；基于深度学习的方法利用深度神经网络将低水平的图像特征融合成带有语义信息的高水平特征，能够很好地解决三维目标识别过程中二维图像的图像噪声问题，但是通常使用的深度卷积神经网络无法表达序列属性，不能有效地刻画三维目标在不同视角下的特征。因此，亟需提出一种在不同视角图像中对图像噪声问题鲁棒的自动化三维目标识别方法。The key to 3D object recognition is to find the 2D representation of the 3D object model and extract the common features of the 3D object and the 2D image. Existing 3D target recognition methods mainly include methods based on artificial markers, methods based on geometric features, and methods based on deep learning. The method based on artificial marking points needs to manually initialize the feature points in the two-dimensional image. Due to the need for manual interaction, such methods are not repeatable; the method based on geometric features achieves the target by extracting information such as the midline skeleton, contour shape and other information of the target. However, such methods have poor recognition effect in the presence of noise in the image; deep learning-based methods use deep neural networks to fuse low-level image features into high-level features with semantic information, which can solve the problem of three-dimensional The problem of image noise in two-dimensional images in the process of target recognition, but the commonly used deep convolutional neural network cannot express sequence attributes, and cannot effectively describe the characteristics of three-dimensional targets in different perspectives. Therefore, there is an urgent need to propose an automated 3D object recognition method that is robust to image noise in images from different perspectives.

发明内容SUMMARY OF THE INVENTION

本发明目的是能够更有效地刻画三维目标在不同视角下的特征，降低特征提取过程对图像噪声的敏感程度，提高三维目标识别准确率，本发明提出一种基于曲率特征递归神经网络的三维目标识别方法。The purpose of the invention is to more effectively describe the characteristics of the three-dimensional target under different viewing angles, reduce the sensitivity of the feature extraction process to image noise, and improve the accuracy of the three-dimensional target recognition. recognition methods.

本发明为实现上述目的所采用的技术方案是：一种基于曲率特征递归神经网络的三维目标识别方法，包括以下步骤：The technical scheme adopted by the present invention to achieve the above object is: a three-dimensional target recognition method based on a curvature feature recurrent neural network, comprising the following steps:

步骤1：计算目标三维模型的联合曲率

Figure BDA0001169544040000021
提取联合曲率
Figure BDA0001169544040000022
的局部极大值构成三维模型的曲率草图RSketch；再对三维模型的曲率草图RSketch利用透射投影变换生成360°二维图像Pm，其中m＝1,2,...,360；Step 1: Calculate the joint curvature of the target 3D model
Figure BDA0001169544040000021
Extract joint curvature
Figure BDA0001169544040000022
The local maximum value of the 3D model constitutes the curvature sketch RSketch of the three-dimensional model; then the 360° two-dimensional image Pm is generated by the transmission projection transformation of the curvature sketch RSketch of the three-dimensional model, where m=1,2,...,360;

步骤2：将360°二维图像输入BRNN，利用多角度特征进行学习计算其在多视角下的序列属性；在softmax层利用softmax函数求得序列属性的正确概率最大时的识别类别；所述BRNN为双向递归神经网络。Step 2: Input the 360° two-dimensional image into the BRNN, and use multi-angle features to learn and calculate its sequence attributes under multiple perspectives; use the softmax function in the softmax layer to obtain the recognition category when the correct probability of the sequence attribute is the largest; the BRNN is a bidirectional recurrent neural network.

所述计算目标三维模型的联合曲率

Figure BDA0001169544040000023
包括以下步骤：The joint curvature of the three-dimensional model of the calculation target
Figure BDA0001169544040000023
Include the following steps:

设

Figure BDA0001169544040000024
是目标三维模型R上给定一点(x,y,z)的法向量；令
Figure BDA0001169544040000025
则px,py,qx,qy定义为
Figure BDA0001169544040000026
Assume
Figure BDA0001169544040000024
is the normal vector of a given point (x, y, z) on the target 3D model R; let
Figure BDA0001169544040000025
Then px , py , qx , qy are defined as
Figure BDA0001169544040000026

计算三维模型R上每一点的法向量周围3×3邻域内的平均高斯曲率

Figure BDA0001169544040000027
和平均均值曲率
Figure BDA0001169544040000028
Calculate the average Gaussian curvature in the 3×3 neighborhood around the normal vector of each point on the 3D model R
Figure BDA0001169544040000027
and mean mean curvature
Figure BDA0001169544040000028

Figure BDA0001169544040000029
Figure BDA0001169544040000029

Figure BDA00011695440400000210
Figure BDA00011695440400000210

其中，

Figure BDA00011695440400000211
为平均曲率矩阵，trace(·)是矩阵的迹，
Figure BDA00011695440400000212
分别为p,q,px,py,qx,qy在3×3邻域内的平均值；in,
Figure BDA00011695440400000211
is the mean curvature matrix, trace( ) is the trace of the matrix,
Figure BDA00011695440400000212
are the average values of p, q, px , py , qx , and qy in a 3×3 neighborhood;

定义目标三维模型R的联合曲率

Figure BDA00011695440400000213
为：Define the joint curvature of the target 3D model R
Figure BDA00011695440400000213
for:

Figure BDA00011695440400000214
Figure BDA00011695440400000214

所述将360°二维图像输入BRNN，利用多角度特征进行学习计算出其在多视角下的序列属性，包括以下步骤：The 360° two-dimensional image is input into the BRNN, and the multi-angle features are used to learn and calculate its sequence attributes under the multi-view, including the following steps:

获取360°二维图像的一维特征序列TS,s＝1,2,...,360，则特征序列TS在BRNN第i层的输出分为正向输出

Figure BDA0001169544040000031
和反向输出
Figure BDA0001169544040000032
并且分别与本层BRNN上一序列的正向输出
Figure BDA0001169544040000033
本层BRNN下一序列的反向输出
Figure BDA0001169544040000034
以及上一层BRNN的正向输出
Figure BDA0001169544040000035
和反向输出
Figure BDA0001169544040000036
有如下关系：Obtain the one-dimensional feature sequence TS of the 360° two-dimensional image, s=1,2,...,360, then the output of the feature sequence TS in the i-th layer of BRNN is divided into forward output
Figure BDA0001169544040000031
and reverse output
Figure BDA0001169544040000032
And respectively with the forward output of the previous sequence of BRNN in this layer
Figure BDA0001169544040000033
The reverse output of the next sequence of BRNN in this layer
Figure BDA0001169544040000034
and the forward output of the previous layer of BRNN
Figure BDA0001169544040000035
and reverse output
Figure BDA0001169544040000036
There are the following relationships:

Figure BDA0001169544040000037
Figure BDA0001169544040000037

Figure BDA0001169544040000038
Figure BDA0001169544040000038

其中，

Figure BDA0001169544040000039
为各输出间的权值矩阵，b为偏置，tanh为神经元激活函数；in,
Figure BDA0001169544040000039
is the weight matrix between each output, b is the bias, and tanh is the neuron activation function;

则特征序列TS在BRNN的总输出Os，即为全连接层fc的输入Ifc为：Then the total output Os of the feature sequence TS in the BRNN, that is, the input Ifc of the fully connected layer fc is:

Figure BDA00011695440400000310
Figure BDA00011695440400000310

其中，

Figure BDA00011695440400000311
分别为正向输出和反向输出在全连接层的连接权值；in,
Figure BDA00011695440400000311
are the connection weights of the forward output and the reverse output in the fully connected layer, respectively;

因此，特征序列TS在全连接层fc的累加输出为

Figure BDA00011695440400000312
即为序列属性。Therefore, the cumulative output of the feature sequence TS in the fully connected layer fc is
Figure BDA00011695440400000312
is the sequence attribute.

所述在softmax层利用softmax函数求得序列属性的正确概率最大时的识别类别，包括以下步骤：Described using the softmax function in the softmax layer to obtain the recognition category when the correct probability of the sequence attribute is the largest, including the following steps:

在softmax层利用softmax函数计算出识别结果为第k类的正确概率p(Ck)In the softmax layer, the softmax function is used to calculate the correct probability p(Ck ) that the recognition result is the kth class

Figure BDA00011695440400000313
Figure BDA00011695440400000313

其中，C为识别类别总数，Ak为第k类三维目标的序列属性在全连接层fc的累加输出结果；Among them, C is the total number of recognition categories, Ak is the cumulative output result of the sequence attribute of the k-th three-dimensional target in the fully connected layer fc;

然后利用最大似然估计方法求得损失函数最小值时，即正确概率p(Ck)最大时的识别类别k：Then the maximum likelihood estimation method is used to obtain the minimum value of the loss function, that is, the identification category k when the correct probability p(Ck ) is the largest:

Figure BDA0001169544040000041
Figure BDA0001169544040000041

其中，δ(·)是克罗内克函数

Figure BDA0001169544040000042
r表示特征序列TS的正确识别类别。where δ( ) is the Kronecker function
Figure BDA0001169544040000042
r represents the correct recognition category of the feature sequence TS.

本发明具有以下有益效果及优点：The present invention has the following beneficial effects and advantages:

1.本发明设计的联合曲率草图特征提取方法，能够自动提取三维模型与二维图像的共同特征，并且联合曲率所使用的局部平均高斯曲率和局部平均均值曲率可以有效的解决图像噪声问题。1. The joint curvature sketch feature extraction method designed by the present invention can automatically extract the common features of the three-dimensional model and the two-dimensional image, and the local average Gaussian curvature and the local average mean curvature used by the joint curvature can effectively solve the problem of image noise.

2.本发明设计多角度特征学习双向递归神经网络，能够同时考虑三维模型在多角度下的特征序列，能够在任意角度的二维图像中准确识别三维目标。2. The present invention designs a multi-angle feature learning bidirectional recurrent neural network, which can simultaneously consider the feature sequences of the three-dimensional model under multiple angles, and can accurately identify the three-dimensional target in the two-dimensional image at any angle.

附图说明Description of drawings

图1为本发明方法流程图；Fig. 1 is the flow chart of the method of the present invention;

图2为本发明方法中的多角度特征学习双向递归神经网络框架图。FIG. 2 is a frame diagram of a bidirectional recurrent neural network for multi-angle feature learning in the method of the present invention.

具体实施方式Detailed ways

下面结合附图及实施例对本发明做进一步的详细说明。The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

本发明主要分为两部分，如图1所示为本发明方法流程图，具体实现过程如下所述。The present invention is mainly divided into two parts, as shown in FIG. 1 is a flow chart of the method of the present invention, and the specific implementation process is as follows.

步骤1：计算目标三维模型的联合曲率，并通过提取联合曲率局部极大值构成三维模型的曲率草图，利用透射投影变换生成360°二维图像作为训练递归神经网络的输入；Step 1: Calculate the joint curvature of the target three-dimensional model, and form the curvature sketch of the three-dimensional model by extracting the local maximum value of the joint curvature, and use the transmission projection transformation to generate a 360° two-dimensional image as the input for training the recurrent neural network;

步骤1.1：设

Figure BDA0001169544040000043
是三维模型上给定一点(x,y,z)的法向量。令
Figure BDA0001169544040000051
则px,py,qx,qy定义为
Figure BDA0001169544040000052
则三维模型的高斯曲率GK为Step 1.1: Set up
Figure BDA0001169544040000043
is the normal vector of a given point (x, y, z) on the 3D model. make
Figure BDA0001169544040000051
Then px , py , qx , qy are defined as
Figure BDA0001169544040000052
Then the Gaussian curvature GK of the three-dimensional model is

GK＝|C|，GK = |C|,

其中曲率矩阵

Figure BDA0001169544040000053
三维模型的均值曲率MK为
Figure BDA0001169544040000054
trace(·)是矩阵的迹。为了消除噪声影响，本发明计算三维模型上每一点的法向量周围其3×3邻域内的平均高斯曲率
Figure BDA0001169544040000055
和平均均值曲率
Figure BDA0001169544040000056
where the curvature matrix
Figure BDA0001169544040000053
The mean curvature MK of the 3D model is
Figure BDA0001169544040000054
trace( ) is the trace of the matrix. In order to eliminate the influence of noise, the present invention calculates the average Gaussian curvature in the 3×3 neighborhood around the normal vector of each point on the three-dimensional model
Figure BDA0001169544040000055
and mean mean curvature
Figure BDA0001169544040000056

Figure BDA0001169544040000057
Figure BDA0001169544040000057

Figure BDA0001169544040000058
Figure BDA0001169544040000058

其中

Figure BDA0001169544040000059
为平均曲率矩阵，
Figure BDA00011695440400000510
分别为p,q,px,py,qx,qy在3×3邻域内的平均值。由此，我们定义三维模型的联合曲率
Figure BDA00011695440400000511
为in
Figure BDA0001169544040000059
is the mean curvature matrix,
Figure BDA00011695440400000510
are the average values of p, q, px , py , qx , and qy in a 3×3 neighborhood, respectively. From this, we define the joint curvature of the 3D model
Figure BDA00011695440400000511
for

Figure BDA00011695440400000512
Figure BDA00011695440400000512

步骤1.2：提取联合曲率

Figure BDA00011695440400000513
的局部最大值点构成三维模型R的曲率草图RSketch。通过透视投影变换，生成三维曲率草图RSketch的360°二维投影图像Pm,m=1,2,...,360，作为BRNN的输入。Step 1.2: Extract the joint curvature
Figure BDA00011695440400000513
The local maximum points of , constitute the curvature sketch RSketch of the 3D model R . Through perspective projection transformation, a 360° two-dimensional projection image Pm , m=1,2,...,360 of the three-dimensional curvature sketch RSketch is generated as the input of the BRNN.

步骤2：本发明采用一种深度递归神经网络(DRNN)作为曲率特征识别方法，DRNN框架如图2所示。利用多角度特征学习BRNN刻画三维模型在多视角下的序列属性，在softmax层利用softmax函数求得正确概率最大的识别类别。Step 2: The present invention adopts a deep recurrent neural network (DRNN) as the curvature feature identification method, and the DRNN framework is shown in FIG. 2 . The multi-angle feature learning BRNN is used to describe the sequence attributes of the 3D model under multi-view, and the softmax function is used in the softmax layer to obtain the recognition category with the highest probability of correctness.

步骤2.1：为了刻画三维模型在不同视角下特征的序列性，定义三维模型在多视角下的一维特征序列为TS,s＝1,2,...,360，则特征序列TS在BRNN第i层的输出分为正向输出

Figure BDA0001169544040000061
和反向输出
Figure BDA0001169544040000062
分别与本层BRNN上一序列的正向输出
Figure BDA0001169544040000063
本层BRNN下一序列的反向输出
Figure BDA0001169544040000064
以及上一层BRNN的正向输出
Figure BDA0001169544040000065
和反向输出
Figure BDA0001169544040000066
有如下关系：Step 2.1: In order to describe the sequence of the features of the 3D model under different perspectives, define the one-dimensional feature sequence of the 3D model in multiple perspectives as TS , s=1,2,...,360, then the feature sequence TS is in The output of the i-th layer of BRNN is divided into forward output
Figure BDA0001169544040000061
and reverse output
Figure BDA0001169544040000062
Respectively with the forward output of the previous sequence of the BRNN in this layer
Figure BDA0001169544040000063
The reverse output of the next sequence of BRNN in this layer
Figure BDA0001169544040000064
and the forward output of the previous layer of BRNN
Figure BDA0001169544040000065
and reverse output
Figure BDA0001169544040000066
There are the following relationships:

Figure BDA0001169544040000067
Figure BDA0001169544040000067

Figure BDA0001169544040000068
Figure BDA0001169544040000068

其中

Figure BDA0001169544040000069
为各输出间的权值矩阵，b为偏执，tanh为神经元激活函数；则特征序列TS在BRNN的总输出Os，即为全连接层fc的输入Ifc为in
Figure BDA0001169544040000069
is the weight matrix between each output, b is paranoia, and tanh is the neuron activation function; then the total output Os of the feature sequence TS in the BRNN is the input Ifc of the fully connected layer fc is

Figure BDA00011695440400000610
Figure BDA00011695440400000610

其中，

Figure BDA00011695440400000611
分别为正向输出和反向输出在全连接层的连接权值。in,
Figure BDA00011695440400000611
are the connection weights of the forward output and reverse output in the fully connected layer, respectively.

步骤2.2：特征序列TS在全连接层fc的累加输出为

Figure BDA00011695440400000612
即为序列属性。在softmax层利用softmax函数计算识别结果为第k类的正确概率p(Ck)Step 2.2: The cumulative output of the feature sequence TS in the fully connected layer fc is
Figure BDA00011695440400000612
is the sequence attribute. In the softmax layer, the softmax function is used to calculate the correct probability p(Ck ) that the recognition result is the kth class

Figure BDA00011695440400000613
Figure BDA00011695440400000613

其中C为识别类别总数，Ak为第k类三维目标的序列属性在全连接层fc的累加输出结果。然后利用最大似然估计方法求得损失函数最小值时，即正确概率p(Ck)最大时的识别类别k：Among them, C is the total number of recognition categories, and Ak is the cumulative output result of the sequence attributes of the k-th three-dimensional objects in the fully connected layer fc. Then the maximum likelihood estimation method is used to obtain the minimum value of the loss function, that is, the identification category k when the correct probability p(Ck ) is the largest:

Figure BDA00011695440400000614
Figure BDA00011695440400000614

其中δ(·)是克罗内克函数

Figure BDA0001169544040000071
r表示特征序列TS的正确识别类别。where δ( ) is the Kronecker function
Figure BDA0001169544040000071
r represents the correct recognition category of the feature sequence TS.

Claims (3)

1. A three-dimensional target identification method based on curvature characteristic recurrent neural network is characterized by comprising the following steps:
step 1: calculating joint curvature of a three-dimensional model of an object
Figure FDA0002972861850000011
Extracting combined curvatures
Figure FDA0002972861850000012
The local maximum values form a curvature sketch R of the three-dimensional modelSketch(ii) a Then, a curvature sketch R of the three-dimensional model is conductedSketchGeneration of a 360 DEG two-dimensional image P using transmission projective transformationmWherein m is 1, 2.., 360;
step 2: inputting a 360-degree two-dimensional image into the BRNN, and utilizing multi-angle characteristics to learn and calculate sequence attributes of the image under multiple visual angles; obtaining the identification category when the correct probability of the sequence attribute is maximum by utilizing a softmax function in a softmax layer; the BRNN is a bidirectional recurrent neural network;
the joint curvature of the three-dimensional model of the calculation target
Figure FDA0002972861850000013
The method comprises the following steps:
is provided with
Figure FDA0002972861850000014
Is a normal vector of a given point (x, y, z) on the target three-dimensional model R; order to
Figure FDA0002972861850000015
Then p isx,py,qx,qyIs defined as
Figure FDA0002972861850000016
Calculating the mean Gaussian curvature in a 3 × 3 neighborhood around the normal vector of each point on the three-dimensional model R
Figure FDA0002972861850000017
And mean curvature
Figure FDA0002972861850000018
Figure FDA0002972861850000019
Figure FDA00029728618500000110
Wherein,
Figure FDA00029728618500000111
being the mean curvature matrix, trace (-) is the trace of the matrix,
Figure FDA00029728618500000112
are respectively p, q, px,py,qx,qyAverage in the 3 × 3 neighborhood;
defining a joint curvature of a three-dimensional model R of an object
Figure FDA00029728618500000113
Comprises the following steps:
Figure FDA00029728618500000114
2. the method for identifying three-dimensional objects based on curvature feature recurrent neural network as claimed in claim 1, wherein said inputting 360 ° two-dimensional image into BRNN, using multi-angle feature to learn and calculate its sequence attribute under multi-view, comprises the following steps:
one-dimensional characteristic sequence T for acquiring 360-degree two-dimensional imageSS 1,2, 360, then the signature sequence TSOutput at the i-th layer of BRNN is divided into forward output
Figure FDA0002972861850000021
And reverse output
Figure FDA0002972861850000022
And respectively output with a sequence on the BRNN of the local layer in the forward direction
Figure FDA0002972861850000023
Reverse output of BRNN next sequence at this layer
Figure FDA0002972861850000024
And the forward output of the upper layer BRNN
Figure FDA0002972861850000025
And reverse output
Figure FDA0002972861850000026
The following relationships exist:
Figure FDA0002972861850000027
Figure FDA0002972861850000028
wherein,
Figure FDA0002972861850000029
b is a bias, and tanh is a neuron activation function;
then the characteristic sequence TSTotal output O at BRNNsI.e. input I of full connection level fcfcComprises the following steps:
Figure FDA00029728618500000210
wherein,
Figure FDA00029728618500000211
respectively is the connection weight of the forward output and the reverse output on the full connection layer;
thus, the signature sequence TSThe cumulative output at full connection level fc is
Figure FDA00029728618500000212
I.e. the sequence property.
3. The three-dimensional object recognition method based on the curvature feature recurrent neural network as claimed in claim 1, wherein the recognition class when the correct probability of the sequence attribute is maximum is found by the softmax layer by using the softmax function, comprising the following steps:
calculating the correct probability p (C) of the recognition result being the kth class by utilizing a softmax function at a softmax layerk)
Figure FDA00029728618500000213
Wherein C is the total number of identification categories, AkAccumulating and outputting a result of the sequence attribute of the kth three-dimensional target at the full connection layer fc;
then, the maximum likelihood estimation method is used to obtain the minimum value of the loss function, i.e. the correct probability p (C)k) Maximum recognition category k:
Figure FDA0002972861850000031
wherein δ (·) is a kronecker function
Figure FDA0002972861850000032
r represents a characteristic sequence TSCorrect identification category of; t isSTo acquire a one-dimensional sequence of features for a 360 ° two-dimensional image, s 1, 2.
CN201611096314.1A2016-12-022016-12-02 A 3D Object Recognition Method Based on Curvature Feature Recurrent Neural NetworkActiveCN108154066B (en)

Priority Applications (1)

Application NumberPriority DateFiling DateTitle
CN201611096314.1ACN108154066B (en)2016-12-022016-12-02 A 3D Object Recognition Method Based on Curvature Feature Recurrent Neural Network

Applications Claiming Priority (1)

Application NumberPriority DateFiling DateTitle
CN201611096314.1ACN108154066B (en)2016-12-022016-12-02 A 3D Object Recognition Method Based on Curvature Feature Recurrent Neural Network

Publications (2)

Publication NumberPublication Date
CN108154066A CN108154066A (en)2018-06-12
CN108154066Btrue CN108154066B (en)2021-04-27

Family

ID=62470160

Family Applications (1)

Application NumberTitlePriority DateFiling Date
CN201611096314.1AActiveCN108154066B (en)2016-12-022016-12-02 A 3D Object Recognition Method Based on Curvature Feature Recurrent Neural Network

Country Status (1)

CountryLink
CN (1)CN108154066B (en)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication numberPriority datePublication dateAssigneeTitle
CN109166183B (en)*2018-07-162023-04-07中南大学Anatomical landmark point identification method and identification equipment
CN109496316B (en)*2018-07-282022-04-01合刃科技（深圳）有限公司Image recognition system
CN109242955B (en)*2018-08-172023-03-24山东师范大学Workpiece manufacturing characteristic automatic identification method and device based on single image
CN109493354B (en)*2018-10-102021-08-06中国科学院上海技术物理研究所 A Reconstruction Method of Target 2D Geometric Shape Based on Multi-view Image
CN110287783A (en)*2019-05-182019-09-27天嗣智能信息科技(上海)有限公司 A method for humanoid recognition in video surveillance images
CN117315397B (en)*2023-10-112024-06-07电子科技大学Classification method for noise data containing labels based on class curvature

Citations (6)

* Cited by examiner, † Cited by third party
Publication numberPriority datePublication dateAssigneeTitle
CN101770566A (en)*2008-12-302010-07-07复旦大学Quick three-dimensional human ear identification method
CN104166842A (en)*2014-07-252014-11-26同济大学Three-dimensional palm print identification method based on partitioning statistical characteristic and combined expression
CN104463111A (en)*2014-11-212015-03-25天津工业大学Three-dimensional face recognition method fused with multi-scale feature region curvatures
CN105205478A (en)*2015-10-232015-12-30天津工业大学3-dimensional human face recognition method integrating anthropometry and curvelet transform
KR101592294B1 (en)*2014-09-032016-02-05배재대학교 산학협력단Decimation Method For Complex Three Dimensional Polygonal Mesh Data
CN106097431A (en)*2016-05-092016-11-09王红军A kind of object global recognition method based on 3 d grid map

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication numberPriority datePublication dateAssigneeTitle
US9984473B2 (en)*2014-07-092018-05-29Nant Holdings Ip, LlcFeature trackability ranking, systems and methods
US10019784B2 (en)*2015-03-182018-07-10Toshiba Medical Systems CorporationMedical image processing apparatus and method

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication numberPriority datePublication dateAssigneeTitle
CN101770566A (en)*2008-12-302010-07-07复旦大学Quick three-dimensional human ear identification method
CN104166842A (en)*2014-07-252014-11-26同济大学Three-dimensional palm print identification method based on partitioning statistical characteristic and combined expression
KR101592294B1 (en)*2014-09-032016-02-05배재대학교 산학협력단Decimation Method For Complex Three Dimensional Polygonal Mesh Data
CN104463111A (en)*2014-11-212015-03-25天津工业大学Three-dimensional face recognition method fused with multi-scale feature region curvatures
CN105205478A (en)*2015-10-232015-12-30天津工业大学3-dimensional human face recognition method integrating anthropometry and curvelet transform
CN106097431A (en)*2016-05-092016-11-09王红军A kind of object global recognition method based on 3 d grid map

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
"Study on novel Curvature Features for 3D fingerprint recognition";Feng Liu 等;《Neurocomputing》;20151130;第168卷;第599-608页*
"基于模型的任意视点下三维目标识别研究";许俊峰;《中国优秀硕士学位论文全文数据库 信息科技辑》;20160115(第01期);全文*

Also Published As

Publication numberPublication date
CN108154066A (en)2018-06-12

Similar Documents

PublicationPublication DateTitle
CN108154066B (en) A 3D Object Recognition Method Based on Curvature Feature Recurrent Neural Network
US12270727B2 (en)Underwater detection method and system for contact leakage of tunnel joints of dam culvert
CN109816049B (en) A deep learning-based assembly monitoring method, device and readable storage medium
CN107103613B (en)A kind of three-dimension gesture Attitude estimation method
CN108537191B (en) A three-dimensional face recognition method based on structured light camera
CN110543878A (en) A Neural Network-Based Recognition Method of Pointer Meter Readings
CN103247045B (en)A kind of method obtaining artificial scene principal direction and image border from multi views
CN108764065A (en)A kind of method of pedestrian's weight identification feature fusion assisted learning
CN110766746B (en)3D driver posture estimation method based on combined 2D-3D neural network
CN110084304A (en)A kind of object detection method based on generated data collection
CN107424161B (en) A Coarse-to-fine Image Layout Estimation Method for Indoor Scenes
CN110246181A (en)Attitude estimation model training method, Attitude estimation method and system based on anchor point
CN116958420A (en) A high-precision modeling method for the three-dimensional face of a digital human teacher
CN109035172A (en)A kind of non-local mean Ultrasonic Image Denoising method based on deep learning
CN105513094A (en)Stereo vision tracking method and stereo vision tracking system based on 3D Delaunay triangulation
CN115937520A (en)Point cloud moving target segmentation method based on semantic information guidance
CN104077775A (en)Shape matching method and device combining skeleton feature points and shape context
CN110490915B (en) A point cloud registration method based on convolution restricted Boltzmann machine
CN115049842B (en)Method for detecting damage of aircraft skin image and positioning 2D-3D
CN113780240B (en)Object pose estimation method based on neural network and rotation characteristic enhancement
CN104182968A (en)Method for segmenting fuzzy moving targets by wide-baseline multi-array optical detection system
CN116416376A (en)Three-dimensional hair reconstruction method, system, electronic equipment and storage medium
CN116310095A (en)Multi-view three-dimensional reconstruction method based on deep learning
CN113808273A (en) A Disordered Incremental Sparse Point Cloud Reconstruction Method for Numerical Simulation of Ship Traveling Waves
He et al.A generative feature-to-image robotic vision framework for 6D pose measurement of metal parts

Legal Events

DateCodeTitleDescription
PB01Publication
PB01Publication
SE01Entry into force of request for substantive examination
SE01Entry into force of request for substantive examination
GR01Patent grant
GR01Patent grant
CB03Change of inventor or designer information
CB03Change 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
[8]ページ先頭
©2009-2025 Movatter.jp