WO2020118901A1 - 一种基于深度学习的天线下倾角测量方法 - Google Patents
一种基于深度学习的天线下倾角测量方法 Download PDFInfo
- Publication number
- WO2020118901A1 WO2020118901A1 PCT/CN2019/075902 CN2019075902W WO2020118901A1 WO 2020118901 A1 WO2020118901 A1 WO 2020118901A1 CN 2019075902 W CN2019075902 W CN 2019075902W WO 2020118901 A1 WO2020118901 A1 WO 2020118901A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- antenna
- network
- deep learning
- downtilt angle
- downtilt
- Prior art date
Links
- 238000013135 deep learning Methods 0.000 title claims abstract description 19
- 238000000691 measurement method Methods 0.000 title claims abstract description 9
- 230000006835 compression Effects 0.000 claims abstract description 19
- 238000007906 compression Methods 0.000 claims abstract description 19
- 230000005284 excitation Effects 0.000 claims abstract description 18
- 238000012545 processing Methods 0.000 claims abstract description 11
- 238000012512 characterization method Methods 0.000 claims abstract description 10
- 238000013139 quantization Methods 0.000 claims abstract description 9
- 238000000605 extraction Methods 0.000 claims abstract description 7
- 238000013528 artificial neural network Methods 0.000 claims abstract description 5
- 230000004044 response Effects 0.000 claims abstract description 4
- 238000000034 method Methods 0.000 claims description 18
- 230000004913 activation Effects 0.000 claims description 17
- 230000006870 function Effects 0.000 claims description 16
- 230000008569 process Effects 0.000 claims description 6
- 230000007246 mechanism Effects 0.000 claims description 4
- XWPCYYOZOJKYKQ-UHFFFAOYSA-N 1-(2-chloroethyl)-3-[2-[2-[[2-chloroethyl(nitroso)carbamoyl]amino]ethyldisulfanyl]ethyl]-1-nitrosourea Chemical compound ClCCN(N=O)C(=O)NCCSSCCNC(=O)N(N=O)CCCl XWPCYYOZOJKYKQ-UHFFFAOYSA-N 0.000 claims description 3
- 238000001514 detection method Methods 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 238000010586 diagram Methods 0.000 description 5
- 238000005259 measurement Methods 0.000 description 5
- XEEYBQQBJWHFJM-UHFFFAOYSA-N Iron Chemical compound [Fe] XEEYBQQBJWHFJM-UHFFFAOYSA-N 0.000 description 2
- 238000004891 communication Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000014509 gene expression Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000009194 climbing Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 229910052742 iron Inorganic materials 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01B—MEASURING LENGTH, THICKNESS OR SIMILAR LINEAR DIMENSIONS; MEASURING ANGLES; MEASURING AREAS; MEASURING IRREGULARITIES OF SURFACES OR CONTOURS
- G01B21/00—Measuring arrangements or details thereof, where the measuring technique is not covered by the other groups of this subclass, unspecified or not relevant
- G01B21/22—Measuring arrangements or details thereof, where the measuring technique is not covered by the other groups of this subclass, unspecified or not relevant for measuring angles or tapers; for testing the alignment of axes
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01C—MEASURING DISTANCES, LEVELS OR BEARINGS; SURVEYING; NAVIGATION; GYROSCOPIC INSTRUMENTS; PHOTOGRAMMETRY OR VIDEOGRAMMETRY
- G01C9/00—Measuring inclination, e.g. by clinometers, by levels
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/60—Analysis of geometric attributes
Definitions
- the invention relates to the field of communication measurement, in particular to a method for measuring antenna downtilt angle based on deep learning.
- Antenna downtilt is one of the important parameters that determine the signal coverage of the base station. Not only the downtilt angle of each antenna needs to be accurately designed in the early stage of network planning, after the base station is put into operation, with the development of services, user changes and the surrounding signal environment Changes, it is also necessary to make accurate adjustments to the downtilt angle.
- the inclinometer is generally used to measure the mechanical downtilt angle of the base station antenna.
- the surveyor When using the inclinometer to measure the mechanical downtilt angle of the antenna, the surveyor must climb the iron tower or the pole to measure close to the antenna, which is very dangerous and troublesome, and also makes the accuracy of the measurement suffer. influences.
- the GSM-R system With the development of technology, the GSM-R system has emerged. This system is a measurement system tool that can accurately measure the downtilt angle of the antenna without being close to the antenna. It can achieve the inclination of the base station antenna without towering. Measurement work, and can be connected to the test points of each base station to achieve real-time monitoring of the base station antenna inclination.
- an object of the embodiments of the present invention is to provide an antenna downtilt measurement method based on deep learning, so as to measure the antenna downtilt angle safely, effectively, quickly, and accurately.
- An antenna downtilt measurement method based on deep learning includes the following steps: establishing an antenna database and performing quantization processing; inputting antenna images to a deep neural network and entering a feature extraction network to obtain antenna feature images; antenna images entering an SE characterization enhancement network Selectively enhance the inclusion of useful features and suppress useless features; the antenna image enters the target recognition network to identify the antenna candidates and obtain the downtilt angle of the antenna; where the SE characterization enhancement network is equipped with a compression excitation unit to model the dependence between channels , And adaptively adjust the characteristic response value of each channel.
- the compression excitation unit performs compression operation and excitation operation on the antenna picture.
- the compression operation includes the following steps: feature compression along the spatial dimension, turning each two-dimensional feature channel into a real number; compressing global spatial information to form a channel descriptor and using global average pooling to generate each The statistics of the channel.
- the excitation operation is implemented through the threshold mechanism of the sigmoid activation function; the feature dimension is reduced to 1/r of the input; after nonlinear activation, it is then raised back to the original dimension through a fully connected layer; where r is a hyperparameter .
- the excitation operation performs the following processing on the antenna characteristic image: F tr : I ⁇ U, I ⁇ R W′ ⁇ H′ ⁇ C′ , U ⁇ R W ⁇ H ⁇ C ;
- F tr represents the convolution operator of one or more convolution operations
- I represents the original image features
- U represents the acquired image features
- W ⁇ H represents the dimension of the feature map
- C represents the number of channels
- V c [v 1 , v 2 , K v C ] represents the learned convolution kernel set
- v c represents the c-th convolution kernel parameter
- Zc is the c-th element of a statistic generated by the spatial dimension HxW of U;
- the establishment of the antenna database and the quantization process include the following steps: inputting antenna image samples with tags; the tags include the latitude and longitude of the antenna, the direction angle of the antenna, and the downtilt angle of the antenna.
- the establishment of the antenna database and the quantization process include the following steps: classifying the antenna picture samples into 30 levels according to the downtilt angle of 0-16 degrees and the degree span of 0.5 degrees.
- the feature extraction network includes a cascaded nonlinear activation function and a lightweight network.
- the target recognition network is a recognition detection module of faster CNCC.
- the objective function of the target recognition network is
- i represents the index of the candidate box in the network
- pi represents the forward classification prediction value
- pi* represents the predicted value probability of the candidate box.
- the beneficial effect of the embodiment of the present invention is that: a method for measuring the downtilt angle of an antenna based on deep learning adopted in the embodiment of the present invention can be obtained by establishing a sample database and inputting the antenna image to be measured into the deep learning network, and directly outputting it after processing Antenna downtilt angle; at the same time, SE characterization enhances the network's compression excitation unit to enhance the ability to express features through the processing of the channel, the antenna image feature processing is more detailed, and it is easier to accurately obtain the antenna downtilt angle; can avoid climbing measurements It is dangerous and reduces the cost of installing sensors, which can obtain antenna downtilt data more effectively, safely, at low cost, and accurately.
- FIG. 1 is a network structure diagram of an embodiment of the present invention
- FIG. 2 is a structural diagram of a cascaded nonlinear activation function according to an embodiment of the present invention
- FIG. 3 is an example diagram of a lightweight network according to an embodiment of the present invention.
- FIG. 4 is a structural diagram of an SE characterization enhanced network according to an embodiment of the present invention.
- FIG. 5 is a structural diagram of a target recognition network according to an embodiment of the present invention.
- FIG. 6 is a flowchart of steps in an embodiment of the present invention.
- an embodiment of the present invention discloses an antenna downtilt measurement method based on deep learning, including the following steps: establishing an antenna database and performing quantization processing; inputting antenna images into a deep neural network and entering a feature extraction network In order to obtain the antenna feature image; the antenna picture enters the SE characterization enhancement network to selectively enhance the useful features and suppress the useless features; the antenna picture enters the target recognition network to identify the antenna candidates and obtain the antenna downtilt angle.
- the process of establishing an antenna database and performing quantization includes the following steps: inputting antenna image samples with tags; the tags include antenna latitude and longitude, antenna direction angle, and antenna downtilt angle.
- the establishment of the antenna database and the quantization process include the following steps: classify the antenna image samples into 30 levels according to the downtilt angle of 0-16 degrees and the degree span of 0.5 degrees.
- the feature extraction network includes a cascaded nonlinear activation function and a lightweight network.
- the cascaded nonlinear activation function performs convolution, inversion, cascade, scale transformation and nonlinear unit processing on the input in order to obtain the result.
- the activation function of the cascaded non-linear unit reduces the number of output channels by half. By simply connecting the same output and inverting module, it becomes double, that is, the number of original outputs is reached, which makes the speed increase 2 times without losing accuracy .
- a lightweight module In a lightweight network, a lightweight module generates output activation values for receptive fields of different sizes.
- the antenna features of the deep neural network correspond to a sufficiently large receptive field, which can be achieved by superimposing 3*3 or larger convolution kernels.
- the antenna characteristics correspond to small enough receptive fields to accurately locate small areas of interest.
- the final 1*1 convolution retains the receptive field of the previous layer, which slows the growth of the receptive field of some output features, allowing the lightweight module to accurately capture small-sized targets.
- the traditional convolutional neural network captures the characteristics of the image from the global receptive field to describe the image, but it does not consider the correlation between the convolution filters, that is, the channel correlation.
- the SE characterization enhancement network is provided with a compression excitation unit to model the dependence relationship between channels, and adaptively adjust the characteristic response value of each channel.
- the compression excitation unit models the dependence of each channel to improve the representation ability of the network, and can adjust the features channel by channel, so that the SE representation enhancement network can learn to use global information to selectively enhance the inclusion of useful Information features and suppress useless features.
- the compression excitation unit performs compression operation and excitation operation on the antenna picture.
- the compression operation includes the following steps: feature compression along the spatial dimension, turning each two-dimensional feature channel into a real number; compressing global spatial information to form a channel descriptor and using global average pooling to generate each The statistics of the channel. Since the compression operation uses global information, the SE representation enhancement network can be placed in the low-level and high-level feature expressions, increasing the low-level feature expression and increasing the high-level category correlation.
- the excitation operation is implemented through the threshold mechanism of the sigmoid activation function; the feature dimension is reduced to 1/r of the input; after nonlinear activation, it is then raised back to the original dimension through a fully connected layer; where r is a hyperparameter , In the embodiment, r takes 16.
- r is a hyperparameter
- r takes 16.
- the excitation operation performs the following processing on the antenna characteristic image: F tr : I ⁇ U, I ⁇ R W′ ⁇ H′ ⁇ C′ , U ⁇ R W ⁇ H ⁇ C ;
- F tr represents the convolution operator of one or more convolution operations
- I represents the original image features
- U represents the acquired image features
- W ⁇ H represents the dimension of the feature map
- C represents the number of channels
- V c [v 1 , v 2 , K v C ] represents the learned convolution kernel set
- v c represents the c-th convolution kernel parameter
- Zc is the c-th element of a statistic Z generated by the spatial dimension HxW of U; since each learned convolution kernel is convolved in a local receptive field, it passes through the SE network module Each data unit u c of the output U after conversion needs to be pooled by global average to generate the statistical information Z of the channel.
- ⁇ is the linear activation function operation
- the target recognition network is a recognition detection module of faster CNCC.
- the objective function of the target recognition network is
- i represents the index of the candidate box in the network
- pi represents the forward classification prediction value
- pi* represents the predicted value probability of the candidate box.
- the entire target recognition network obtains the antenna picture candidate frame and detects and recognizes the antenna picture, and directly identifies the antenna downtilt angle.
Landscapes
- Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Theoretical Computer Science (AREA)
- Geometry (AREA)
- Radar, Positioning & Navigation (AREA)
- Remote Sensing (AREA)
- Image Analysis (AREA)
- Variable-Direction Aerials And Aerial Arrays (AREA)
Abstract
本发明公开了一种基于深度学习的天线下倾角测量方法，包括以下步骤：建立天线数据库并进行量化处理；天线图片输入到深度神经网络，并进入特征提取网络以获取天线特征图像；天线图片进入SE表征增强网络有选择性地加强包含有用特征并抑制无用特征；天线图片进入目标识别网络对天线进行识别候选并得到天线下倾角角度；其中SE表征增强网络设有压缩激励单元对通道间的依赖关系进行建模，并自适应的调整各通道的特征响应值。通过对天线图片进行深度学习网络的处理得到天线下倾角角度，建立一种方便、安全、有效、准确的天线测量方法。
Description
本发明涉及通信测量领域，特别是一种基于深度学习的天线下倾角测量方法。
在通讯领域里，经常要对天线下倾角进行调整。天线下倾角是决定基站信号覆盖范围的重要参数之一，不但在网络规划的初期需要准确设计每个天线的下倾角，在基站投入运行以后，随着业务的发展，用户的变化以及周围信号环境的变化，还需要对下倾角做出准确调整。
目前对基站天线机械下倾角的测量普遍采用坡度计，使用坡度计测量天线机械下倾角时，测量者必须爬上铁塔或者抱杆贴近天线进行测量，相当危险和麻烦，也使得测量的准确性受到影响。而随着技术发展，出现了GSM-R系统，该系统是一种测量人员可以不用贴近天线就可以准确测量出天线下倾角的测量系工具，能够实现不登塔作业即可完成基站天线倾角的测量工作，并可对各基站测试点进行联网，实现对基站天线倾角的实时监测。但安装传感器，不仅耗时，成本较高，且新旧塔间、基站塔层数及数量等都存在差异性，因而该方法实用性不高，运行周期长，实现较为困难。因此设计出简单操作，性能可靠的角度测量方法就很有必要。
发明内容
为解决上述问题，本发明实施例的目的在于提供一种基于深度学习的天线下 倾角测量方法，以便安全、有效、快速、准确地对天线下倾角进行测量。
本发明实施例解决其问题所采用的技术方案是：
一种基于深度学习的天线下倾角测量方法，包括以下步骤：建立天线数据库并进行量化处理；天线图片输入到深度神经网络，并进入特征提取网络以获取天线特征图像；天线图片进入SE表征增强网络有选择性地加强包含有用特征并抑制无用特征；天线图片进入目标识别网络对天线进行识别候选并得到天线下倾角角度；其中SE表征增强网络设有压缩激励单元对通道间的依赖关系进行建模，并自适应的调整各通道的特征响应值。
进一步，所述压缩激励单元对天线图片进行压缩操作和激励操作。
进一步，所述压缩操作包括以下步骤：顺着空间维度来进行特征压缩，将每个二维的特征通道变成一个实数；压缩全局空间信息形成一个通道描述符并使用全局平均池化来生成各通道的统计量。
进一步，所述激励操作通过sigmoid激活函数的门限机制实现；将特征维度降低到输入的1/r；再经过非线性激活后再通过一个全连接层升回到原来的维度；其中r为超参数。
进一步，所述激励操作对天线特征图像进行以下处理：F
tr：I→U，I∈R
W′×H′×C′，U∈R
W×H×C；
其中，F
tr表示一次或者多次卷积操作的卷积操作符，I表示原始的图像特征，U表示获取的图像特征，W×H表示特征图的维度，C表示通道个数；
其中，V
c＝[v
1，v
2，K v
C]表示学习的卷积核集合，v
c表示第c个卷积核参数，其输出为Uc＝[u1，u2，...u
c]；
其中，Zc为通过U的空间维度HxW生成的一个统计量的第c个元素；
s＝F
ex(z，W)＝σ(g(z，W))＝σ(W
2δ(W
1z))；
其中，δ是线性激活函数操作，
进一步，所述的建立天线数据库并进行量化处理包括以下步骤：输入带有标签的天线图片样本；所述标签包括天线经纬度、天线方向角和天线下倾角。
优选地，所述的建立天线数据库并进行量化处理包括以下步骤：将天线图片样本按照下倾角的角度0-16度且度数跨度为0.5度分类为30个等级。
优选地，所述特征提取网络包括级联非线性激活函数和轻量级网络。
优选地，所述目标识别网络为faster CNCC的识别检测模块。
进一步，所述目标识别网络的目标函数为
其中，i表示网络中候选框的索引，pi表示前向分类预测值，pi*表示候选框的预测值概率。
本发明实施例的有益效果是：本发明实施例采用的一种基于深度学习的天线下倾角测量方法，可以通过建立样本数据库，并将要测量的天线图片输入到深度学习网络，经过处理直接输出得到天线下倾角角度；同时SE表征增强网络的压缩激励单元通过对通道的处理增强特征的表达能力，对天线图片的特征处理更细致，更容易准确地获得天线下倾角角度；可以避免攀爬测量的危险和减少安装传感器的成本，能更加有效、安全、低成本、准确地获得天线下倾角数据。
下面结合附图和实例对本发明作进一步说明。
图1是本发明实施例的网络结构图；
图2是本发明实施例的级联非线性激活函数的结构图；
图3是本发明实施例的轻量级网络的示例图；
图4是本发明实施例的SE表征增强网络的结构图；
图5是本发明实施例的目标识别网络的结构图；
图6是本发明实施例的步骤流程图。
参照图1和图6，本发明实施例公开了一种基于深度学习的天线下倾角测量方法，包括以下步骤：建立天线数据库并进行量化处理；天线图片输入到深度神经网络，并进入特征提取网络以获取天线特征图像；天线图片进入SE表征增强网络有选择性地加强包含有用特征并抑制无用特征；天线图片进入目标识别网络对天线进行识别候选并得到天线下倾角角度。
在一个实施例中，所述的建立天线数据库并进行量化处理包括以下步骤：输入带有标签的天线图片样本；所述标签包括天线经纬度、天线方向角和天线下倾角。所述的建立天线数据库并进行量化处理包括以下步骤：将天线图片样本按照下倾角的角度0-16度且度数跨度为0.5度分类为30个等级。
参照图2和图3，在一个实施例中，所述特征提取网络包括级联非线性激活函数和轻量级网络。级联非线性激活函数将输入依次进行卷积、取反、级联、尺度变换和非线性单位处理得到结果。级联非线性单元激活函数减少一半输出通道的数量，通过简单的连接相同的输出和取反模块使其变成双倍，即达到原来输出的数量，这使得2倍的速度提升而没有损失精度。
在轻量级网络中，一个轻量级模块产生不同大小的感受野的输出激活值。为了学习捕获大目标的视觉模式，深度神经网络的天线特征对应于足够大的感受野，这可以通过叠加3*3或者更大的卷积核实现。在另外一方面，为了捕获小尺寸的天线，天线特征对应于足够小的感受野来精确定位小的感兴趣区域。最后的1*1卷积保留上一层的感受野，它减慢了一些输出特征的感受野的增长，使得轻量级模块可以精确地捕获小尺寸的目标。
传统的卷积神经网络从全局感受野上去捕获图像的特征来进行图像的描述，但是没有考虑卷积滤波器之间的相关性，即通道相关性。SE表征增强网络设有压缩激励单元对通道间的依赖关系进行建模，并自适应的调整各通道的特征响应值。所述压缩激励单元通过对各通道的依赖性进行建模以提高网络的表示能力，并且可以对特征进行逐通道调整，这样SE表征增强网络就可以学习通过全局信息来有选择性的加强包含有用信息的特征并抑制无用特征。
参照图4，在一个具体实施例中，所述压缩激励单元对天线图片进行压缩操作和激励操作。
进一步，所述压缩操作包括以下步骤：顺着空间维度来进行特征压缩，将每个二维的特征通道变成一个实数；压缩全局空间信息形成一个通道描述符并使用全局平均池化来生成各通道的统计量。由于压缩操作使用了全局信息，因此SE表征增强网络可以放在低层和高层特征表达中，增加低层的特征表达同时增加高层的类别相关性。
进一步，所述激励操作通过sigmoid激活函数的门限机制实现；将特征维度降低到输入的1/r；再经过非线性激活后再通过一个全连接层升回到原来的维度；其中r为超参数，实施例中r取16。门限机制中使用瓶颈形式的两个全连接 层；这样相比只用一个全连接层，好处在于：具有更多的非线性，可以更好地拟合通道间复杂的相关性；极大地减少了参数量和计算量。
进一步，所述激励操作对天线特征图像进行以下处理：F
tr：I→U，I∈R
W′×H′×C′，U∈R
W×H×C；
其中，F
tr表示一次或者多次卷积操作的卷积操作符，I表示原始的图像特征，U表示获取的图像特征，W×H表示特征图的维度，C表示通道个数；
其中，V
c＝[v
1，v
2，K v
C]表示学习的卷积核集合，v
c表示第c个卷积核参数，其输出为Uc＝[u1，u2，...u
c]；
其中，Zc为通过U的空间维度HxW生成的一个统计量Z的第c个元素；由于每个学习好的卷积核都是以一个局部感受野的方式进行卷积，因此在经过SE网络模块转换之后的输出U的各个数据单元u
c需通过全局平均池化以产生通道的统计信息Z。
s＝F
ex(z，W)＝σ(g(z，W))＝σ(W
2δ(W
1z))；
其中，δ是线性激活函数操作，
其为了能够在多个通道在经过多次激活函数情况下学习多个通道之间非互斥关系。
参照图5，在一个实施例中，优选地，所述目标识别网络为faster CNCC的识别检测模块。
进一步，所述目标识别网络的目标函数为
其中，i表示网络中候选框的索引，pi表示前向分类预测值，pi*表示候选框的预测值概率。
整个目标识别网络获取天线图片候选框并对天线图片进行检测识别，直接识别出天线下倾角角度。
以上所述，只是本发明的较佳实施例而已，本发明并不局限于上述实施方式，只要其以相同的手段达到本发明的技术效果，都应属于本发明的保护范围。
Claims (10)
- 一种基于深度学习的天线下倾角测量方法，其特征在于，包括以下步骤：建立天线数据库并进行量化处理；天线图片输入到深度神经网络，并进入特征提取网络以获取天线特征图像；天线图片进入SE表征增强网络有选择性地加强包含有用特征并抑制无用特征；天线图片进入目标识别网络对天线进行识别候选并得到天线下倾角角度；其中SE表征增强网络设有压缩激励单元对通道间的依赖关系进行建模，并自适应的调整各通道的特征响应值。
- 根据权利要求1所述的一种基于深度学习的天线下倾角测量方法，其特征在于，所述压缩激励单元对天线图片进行压缩操作和激励操作。
- 根据权利要求2所述的一种基于深度学习的天线下倾角测量方法，其特征在于，所述压缩操作包括以下步骤：顺着空间维度来进行特征压缩，将每个二维的特征通道变成一个实数；压缩全局空间信息形成一个通道描述符并使用全局平均池化来生成各通道的统计量。
- 根据权利要求2或3任一项所述的一种基于深度学习的天线下倾角测量方法，其特征在于，所述激励操作通过sigmoid激活函数的门限机制实现；将特征维度降低到输入的1/r；再经过非线性激活后再通过一个全连接层升回到原来的维度；其中r为超参数。
- 根据权利要求4所述的一种基于深度学习的天线下倾角测量方法，其特征在于，所述激励操作对天线特征图像进行以下处理：F tr：I→U，I∈R W′×H′×C′，U∈R W×H×C；其中，F tr表示一次或者多次卷积操作的卷积操作符，I表示原始的图像特征，U表示获取的图像特征，W×H表示特征图的维度，C表示通道个数；
其中，V c＝[v 1，v 2，K v C]表示学习的卷积核集合，v c表示第c个卷积核参数，其输出为Uc＝[u1，u2，...u c]；
其中，Zc为通过U的空间维度HxW生成的一个统计量的第c个元素；s＝F ex(z，W)＝σ(g(z，W))＝σ(W 2δ(W 1z))；其中，δ是线性激活函数操作，
- 根据权利要求1所述的一种基于深度学习的天线下倾角测量方法，其特征在于，所述的建立天线数据库并进行量化处理包括以下步骤：输入带有标签的天线图片样本；所述标签包括天线经纬度、天线方向角和天线下倾角。
- 根据权利要求3所述的一种基于深度学习的天线下倾角测量方法，其特征在于，所述的建立天线数据库并进行量化处理包括以下步骤：将天线图片样本按照下倾角的角度0-16度且度数跨度为0.5度分类为30个等级。
- 根据权利要求1所述的一种基于深度学习的天线下倾角测量方法，其特征在于，所述特征提取网络包括级联非线性激活函数和轻量级网络。
- 根据权利要求1所述的一种基于深度学习的天线下倾角测量方法，其特征在于，所述目标识别网络为faster CNCC的识别检测模块。
- 根据权利要求9所述的一种基于深度学习的天线下倾角测量方法，其特征 在于，所述目标识别网络的目标函数为其中，i表示网络中候选框的索引，pi表示前向分类预测值，pi*表示候选框的预测值概率。
Applications Claiming Priority (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811321824 | 2018-11-07 | ||
| CN201811538747.7 | 2018-12-14 | ||
| CN201811538747.7A CN109870138A (zh) | 2018-11-07 | 2018-12-14 | 一种基于深度学习的天线下倾角测量方法 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2020118901A1 true WO2020118901A1 (zh) | 2020-06-18 |
Family
ID=66917084
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/CN2019/075902 WO2020118901A1 (zh) | 2018-11-07 | 2019-02-22 | 一种基于深度学习的天线下倾角测量方法 |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN109870138A (zh) |
| WO (1) | WO2020118901A1 (zh) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116052003B (zh) * | 2023-02-07 | 2024-05-14 | 中科星图数字地球合肥有限公司 | 一种天线角度信息的测量方法、装置及相关设备 |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105556742B (zh) * | 2014-05-27 | 2018-05-11 | 华为技术有限公司 | 天线工程参数的获取方法和设备及系统 |
| CN108446729A (zh) * | 2018-03-13 | 2018-08-24 | 天津工业大学 | 基于卷积神经网络的鸡蛋胚胎分类方法 |
| CN108710826A (zh) * | 2018-04-13 | 2018-10-26 | 燕山大学 | 一种交通标志深度学习模式识别方法 |
| CN108805188A (zh) * | 2018-05-29 | 2018-11-13 | 徐州工程学院 | 一种基于特征重标定生成对抗网络的图像分类方法 |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108550133B (zh) * | 2018-03-02 | 2021-05-18 | 浙江工业大学 | 一种基于Faster R-CNN的癌细胞检测方法 |
-
2018
- 2018-12-14 CN CN201811538747.7A patent/CN109870138A/zh active Pending
-
2019
- 2019-02-22 WO PCT/CN2019/075902 patent/WO2020118901A1/zh active Application Filing
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105556742B (zh) * | 2014-05-27 | 2018-05-11 | 华为技术有限公司 | 天线工程参数的获取方法和设备及系统 |
| CN108446729A (zh) * | 2018-03-13 | 2018-08-24 | 天津工业大学 | 基于卷积神经网络的鸡蛋胚胎分类方法 |
| CN108710826A (zh) * | 2018-04-13 | 2018-10-26 | 燕山大学 | 一种交通标志深度学习模式识别方法 |
| CN108805188A (zh) * | 2018-05-29 | 2018-11-13 | 徐州工程学院 | 一种基于特征重标定生成对抗网络的图像分类方法 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109870138A (zh) | 2019-06-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Park et al. | Concrete crack detection and quantification using deep learning and structured light | |
| CN113705478B (zh) | 一种基于改进YOLOv5的红树林单木目标检测方法 | |
| US20210215481A1 (en) | Method for measuring antenna downtilt angle based on multi-scale deep semantic segmentation network | |
| CN111080693A (zh) | 一种基于YOLOv3的机器人自主分类抓取方法 | |
| KR20200087297A (ko) | 인공신경망 기반의 영상 분할을 이용한 불량 검출 방법 및 불량 검출 장치 | |
| Bo et al. | Particle pollution estimation from images using convolutional neural network and weather features | |
| CN113065578B (zh) | 一种基于双路区域注意力编解码的图像视觉语义分割方法 | |
| CN108537286B (zh) | 一种基于关键区域检测的复杂目标精准识别方法 | |
| CN108711172B (zh) | 基于细粒度分类的无人机识别与定位方法 | |
| CN107301376B (zh) | 一种基于深度学习多层刺激的行人检测方法 | |
| CN114973002A (zh) | 一种基于改进的YOLOv5的麦穗检测方法 | |
| CN108734200B (zh) | 基于bing特征的人体目标视觉检测方法和装置 | |
| CN112801227B (zh) | 一种台风识别模型的生成方法、装置、设备及存储介质 | |
| CN110307903B (zh) | 一种家禽特定部位无接触温度动态测量的方法 | |
| CN111274926B (zh) | 图像数据筛选方法、装置、计算机设备和存储介质 | |
| CN113052185A (zh) | 一种基于Faster R-CNN的小样本目标检测方法 | |
| CN111008576A (zh) | 行人检测及其模型训练、更新方法、设备及可读存储介质 | |
| CN111160100A (zh) | 一种基于样本生成的轻量级深度模型航拍车辆检测方法 | |
| WO2020118901A1 (zh) | 一种基于深度学习的天线下倾角测量方法 | |
| CN116563841B (zh) | 配电网设备标识牌的检测方法、检测装置和电子设备 | |
| CN106682604B (zh) | 一种基于深度学习的模糊图像检测方法 | |
| CN117456441A (zh) | 一种结合变化区域识别的锈斑面积扩大的监测方法与系统 | |
| CN115830302A (zh) | 一种多尺度特征提取融合配电网设备定位识别方法 | |
| CN112991425B (zh) | 一种水域水位提取方法、系统及存储介质 | |
| CN114463678A (zh) | 一种利用摄像头视频图像的降雨类型识别方法 |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 19897242 Country of ref document: EP Kind code of ref document: A1 |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 19897242 Country of ref document: EP Kind code of ref document: A1 |