WO2019218895A1 - 一种应用于深度学习的特征提取方法及系统 - Google Patents

一种应用于深度学习的特征提取方法及系统 Download PDF

Info

Publication number
WO2019218895A1
WO2019218895A1 PCT/CN2019/085836 CN2019085836W WO2019218895A1 WO 2019218895 A1 WO2019218895 A1 WO 2019218895A1 CN 2019085836 W CN2019085836 W CN 2019085836W WO 2019218895 A1 WO2019218895 A1 WO 2019218895A1
Authority
WO
WIPO (PCT)
Prior art keywords
image information
feature
feature extraction
extracted
optical
Prior art date
Application number
PCT/CN2019/085836
Other languages
English (en)
French (fr)
Inventor
周旭东
姚长呈
宋海涛
闫超
Original Assignee
成都理想境界科技有限公司
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 成都理想境界科技有限公司 filed Critical 成都理想境界科技有限公司
Publication of WO2019218895A1 publication Critical patent/WO2019218895A1/zh

Links

Images

Classifications

    • 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
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/067—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using optical means
    • G06N3/0675—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using optical means using electro-optical, acousto-optical or opto-electronic means
    • 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
    • G06N3/045—Combinations of networks

Definitions

  • the invention relates to the field of artificial intelligence, and in particular to a feature extraction method and system applied to deep learning.
  • Deep learning stems from the study of artificial neural networks.
  • a multilayer perceptron with multiple hidden layers is a deep learning structure. Deep learning combines low-level features to form more abstract high-level representation attribute categories or features to discover distributed feature representations of data. Deep learning is a method based on the representation and learning of data in machine learning. In the fast-growing technology environment, it is increasingly used in the field of machine learning such as artificial intelligence, facial recognition, and iris recognition. Humans are also constantly exploring how to extract features faster in deep learning.
  • the convolutional neural network in deep learning consists of input layer, convolution layer, activation function, pooling layer and fully connected layer.
  • Convolution layer is used to extract technical features.
  • deep learning convolution operation the basic principle of feature extraction A matrix convolution operation on a spatial domain is performed using different filters and a set of data to be extracted by features. If the data to be extracted by the feature is X and the filter is Y, the information after feature extraction is
  • the calculation of the matrix convolution operation is complicated, computationally intensive, and computationally time consuming.
  • the data to be extracted by the feature and the filter are generally Fourier transformed in the computer, thereby converting the spatial/space domain information of the two sets of data into spectral information; correspondingly, the convolution operation on the spatial/space domain is converted.
  • the calculation of the point multiplication operation is much smaller.
  • the conversion time is determined by the computer performance and the amount of data. The stronger the computer performance, the faster the conversion speed; the larger the amount of data, the slower the conversion speed. If you want to improve the deep learning power, you only have to invest more and more computer hardware, which is costly and inefficient.
  • the object of the present invention is to provide a feature extraction method and system for deep learning, which solves the problem of how to improve the efficiency of deep learning calculation under the condition of limited computer performance and quantity.
  • the present invention provides a feature extraction system for deep learning, comprising: a display module, an optical Fourier transform device, and an optical filter, wherein: the display module is configured to be characterized The extracted digital signal is converted into spatial domain image information to be feature extracted; the optical Fourier transform device is configured to convert the spatial domain image information to be feature extracted into spectral image information to be feature extracted; the optical filter is used for The spectral image information extracted by the feature is converted into spectral image information after feature extraction.
  • different optical regions of the optical filter have different light transmittances.
  • the optical filter is a liquid crystal light valve array or a lithographic lens.
  • the display module comprises one of a fiber scanning display module, an OLED display module, an LCD display module, an LCoS display module, and a DLP display module.
  • the system further comprises an optical Fourier inverter device, configured to convert the feature extracted spectral image information into feature extracted spatial image information.
  • the system further comprises a photodetector for converting the spatial image information extracted by the feature into an electrical signal.
  • the photodetector comprises one of a CCD detector, a photodiode, a photon detector, and a photomultiplier tube.
  • the optical Fourier transform device and the optical Fourier inverter device are both Fourier transform lenses.
  • the present invention also provides a feature extraction method applied to deep learning, which comprises: converting a digital signal to be extracted by a feature into spatial domain image information to be extracted by a display technology; and spatial image information extracted by the feature to be extracted. Converting to spectral image information to be feature extracted by an optical Fourier transform device; inputting the spectral image information to be extracted by the feature to its corresponding optical filter to obtain spectral image information after feature extraction.
  • the optical filter is a liquid crystal light valve array or a lithographic lens, and different regions have different light transmittances.
  • the present invention has the following beneficial effects:
  • the invention converts the digital signal to be feature extracted into a spatial domain signal by using display technology, and then transforms the time domain signal into a frequency domain signal by using an optical Fourier transform device, and then implements frequency domain dot multiplication by an optical scheme to realize feature extraction.
  • the invention can effectively reduce the amount of deep learning calculation, and after the partial operation is converted into the light calculation, the calculation speed of the part becomes the speed of light, and the calculation time is greatly reduced.
  • FIG. 1 is a schematic structural diagram of a feature extraction system applied to deep learning according to an embodiment of the present invention
  • FIG. 2 is another schematic structural diagram of a feature extraction system applied to deep learning according to an embodiment of the present invention.
  • FIG. 3 is a schematic flowchart of a feature extraction method applied to deep learning according to an embodiment of the present invention.
  • a feature extraction system applied to deep learning includes: a display module 1, an optical Fourier transform device 2, and an optical filter 3.
  • the display module 1 is configured to convert the digital signal to be extracted by the feature into the spatial image information to be extracted by the feature.
  • the display module 1 may be a fiber scanning display module, an OLED display module, or an LCD display module. , LCoS display module or DLP display module.
  • the optical Fourier transform device 2 is configured to convert the spatial domain image information to be feature extracted into spectral image information to be feature extracted.
  • the spatial image information to be extracted by the feature is illuminated at one end of the optical Fourier transform device, and the other end of the optical Fourier transform outputs its corresponding spectral image information, that is, the spectral image information to be extracted by the feature.
  • the optical filter 3 is configured to convert the spectral image information to be feature extracted into the spectral image information after feature extraction.
  • the optical filter 3 is based on a digital filter, and a certain transmittance optical device corresponding to the spectrum of the digital filter is placed on the spectrum surface of the input image, and the spectrum surface is A dot multiplication of the two spectra is performed, which is performed on the optical domain.
  • the optical filter may be a liquid crystal light valve array or a lithographic lens.
  • the optical filter can be produced as follows:
  • each element of the matrix is made into an optical device having a certain transmittance, that is, a matrix of modulation regions having the same number of rows and columns as the normalized matrix is formed on one optical device, wherein if the value of the element in the normalized matrix is 1, The transmittance of the modulation area of the corresponding position is 100%, and if the value of the element is 0.5, the transmittance of the modulation area of the corresponding position is 50%, and other modulation areas are produced according to the same principle (transmission rate setting and change) This can be achieved by conversion of the liquid crystal state. If the element is negative, it corresponds to a phase change which is modulated by the thickness of the modulation region.
  • the spectral image information extracted by the feature obtained by the embodiment of the present invention can be directly used for subsequent feature extraction. In other embodiments, as shown in FIG. 2, it can be input to the optical Fourier inverter device 4, and converted. The spatial image information after the feature extraction.
  • the system further includes a photodetector 5 for converting the spatial image information extracted by the feature into an electrical signal for subsequent analysis, storage, use, and the like.
  • the photodetector 5 may be a CCD detector, a photodiode, a photon detector or a photomultiplier tube.
  • both the optical Fourier transform device 2 and the optical Fourier inverter device 4 may be a Fourier transform lens or an optical device having the same function as the Fourier transform lens.
  • the method mainly includes: converting a digital signal to be extracted by a feature into a spatial image information to be extracted by a display technology;
  • the spatial image information to be extracted by the feature is converted into the spectral image information to be extracted by the feature by the optical Fourier transform device; and the spectral image information extracted by the feature is input to the optical filter corresponding to the spectral image information to be extracted by the feature.
  • the spectral image information after feature extraction is obtained.
  • the optical filter is a liquid crystal light valve array or a lithographic lens, and different regions have different light transmittances, and the fabrication may be as follows:
  • each element of the matrix is made into an optical device having a certain transmittance, that is, a matrix of modulation regions having the same number of rows and columns as the normalized matrix is formed on one optical device, wherein if the value of the element in the normalized matrix is 1, The transmittance of the modulation area of the corresponding position is 100%, and if the value of the element is 0.5, the transmittance of the modulation area of the corresponding position is 50%, and other modulation areas are produced according to the same principle (transmission rate setting and change) This can be achieved by conversion of the liquid crystal state. If the element is negative, it corresponds to a phase change which is modulated by the thickness of the modulation region.
  • feature extraction in deep learning requires convolution between a set of digital signals to be extracted by the feature and the digital signal of the filter.
  • the digital signal to be extracted by the feature is converted into an optical signal form.
  • the filter in the form of digital signal is made into the corresponding optical device.
  • the optical signal form can be easily converted into the frequency domain.
  • the convolution in the time domain is equivalent to the dot multiplication in the corresponding frequency domain, and the convolution in the time domain is complicated.
  • the operation of the frequency domain dot multiplication is much simpler, so the present invention can effectively reduce the amount of deep learning calculation, and the Fourier transform process in the embodiment of the present invention uses an optical device of the same function such as a Fourier transform lens to perform conversion.
  • the transformation speed is the speed of light, which greatly reduces the time consuming of the Fourier transform, and realizes the point multiplication of the spectrum by the optical scheme.
  • the multiplication operation speed is also the speed of light, which greatly reduces the time consumption of the spectrum points.
  • the invention is not limited to the specific embodiments described above.
  • the invention extends to any new feature or any new combination disclosed in this specification, as well as any novel method or process steps or any new combination disclosed.

Abstract

一种应用于深度学习的特征提取方法，其包括：将待特征提取的数字信号通过显示技术转换为待特征提取的空域图像信息；将所述待特征提取的空域图像信息通过光学傅里叶变换装置转换为待特征提取的频谱图像信息；将所述待特征提取的频谱图像信息输入至其对应的光学滤波器，得到特征提取后的频谱图像信息。该方法及对应的特征提取系统将现有技术中时域上的卷积转换为频域上的点乘，有效减小深度学习计算量，且傅里叶变换过程采用傅里叶变换透镜等相同功效的光学器件进行变换，其变换速度为光速，极大地缩减傅里叶变换的耗时，且通过光学方案实现频谱的点乘运算，该点乘运算速度也为光速，极大缩减频谱点乘耗时。

Description

一种应用于深度学习的特征提取方法及系统
本申请要求享有2018年5月15日提交的名称为“一种应用于深度学习的特征提取方法及系统”的中国专利申请CN201810460811.8的优先权，其全部内容通过引用并入本文中。
技术领域
本发明涉及人工智能领域，尤其涉及一种应用于深度学习的特征提取方法及系统。
背景技术
深度学习的概念源于人工神经网络的研究。含多隐层的多层感知器就是一种深度学习结构。深度学习通过组合低层特征形成更加抽象的高层表示属性类别或特征，以发现数据的分布式特征表示。深度学习是机器学习中一种基于对数据进行表征学习的方法，在高速发展的科技环境中，其越来越多地被应用在人工智能、面部识别、虹膜识别等机器学习领域。人类也在不断探索在深度学习中如何更快地进行特征提取。
目前深度学习中卷积神经网络由输入层、卷积层、激活函数、池化层、全连接层组成，卷积层用于提取技术特征，在深度学习卷积运算里，特征提取的基本原理是采用不同的滤波器与一组待特征提取的数据进行空域上的矩阵卷积运算。若待特征提取的数据为X，滤波器为Y，则特征提取后的信息则为
Figure PCTCN2019085836-appb-000001
但是，矩阵卷积运算过程的计算复杂、计算量大、计算耗时长。因此，目前一般在计算机中将待特征提取的数据与滤波器进行傅里叶变换，从而将两组数据的空域/空域信息变换为频谱信息；相应地，空域/空域上的卷积运算便转换为频谱上的点乘运算，而点乘运算的计算量则会小很多。但即便进行傅里叶变换，也是在计算机的程序中完成，其变换时间也是由计算机性能及数据量决定。计算机性能越强，变换速度越快；数据量越大，变换速度越慢。如果要使深度学习计算能力提高，只有不断投入性能更强更多的计算机硬件，其成本高，效率低。
发明内容
本发明的目的是提供一种应用于深度学习的特征提取方法及系统，解决在有限计算机性能及数量的情况下，如何实现提升深度学习计算效率的问题。
为了实现上述发明目的，本发明提供了一种应用于深度学习的特征提取系统，包括：显示模组、光学傅里叶变换装置和光学滤波器，其中：所述显示模组用于将待特征提取的数字信号转换为待特征提取的空域图像信息；所述光学傅里叶变换装置用于将所述待特征提取的空域图像信息转换为待特征提取的频谱图像信息；所述光学滤波器用于将所述待特征提取的频谱图像信息转换为特征提取后的频谱图像信息。
优选的，所述光学滤波器不同区域光透过率不同。
优选的，所述光学滤波器为液晶光阀阵列或光刻透镜。
优选的，所述显示模组包括光纤扫描显示模组、OLED显示模组、LCD显示模组、LCoS显示模组和DLP显示模组中的一种。
优选的，所述系统还包括光学傅里叶逆变装置，用于将所述特征提取后的频谱图像信息转换为特征提取后的空域图像信息。
优选的，所述系统还包括光探测器，用于将所述特征提取后的空域图像信息转换为电信号。
优选的，所述光探测器包括CCD探测器、光电二极管、光子型探测器、光电倍增管中的一种。
优选的，所述光学傅里叶变换装置和光学傅里叶逆变装置均为傅里叶变换透镜。
相应的，本发明还提出一种应用于深度学习的特征提取方法，包括：将待特征提取的数字信号通过显示技术转换为待特征提取的空域图像信息；将所述待特征提取的空域图像信息通过光学傅里叶变换装置转换为待特征提取的频谱图像信息；将所述待特征提取的频谱图像信息输入至其对应的光学滤波器，得到特征提取后的频谱图像信息。
优选的，所述光学滤波器为液晶光阀阵列或光刻透镜，其不同区域光透过率不同。
与现有技术相比，本发明具有如下有益效果：
本发明将待特征提取的数字信号用显示技术变成空域信号，然后再用光学傅里叶变换装置将时域信号变成频域信号，然后通过光学方案实现频域点乘，实现特征提取。本发明可有效减小深度学习计算量，且部分运算转换为光计算后，这部分计算速度即变为光速，计算时间大幅缩小。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技 术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图：
图1为本发明实施例应用于深度学习的特征提取系统的一种结构示意图；
图2为本发明实施例应用于深度学习的特征提取系统的另一种结构示意图；
图3为本发明实施例一种应用于深度学习的特征提取方法流程示意图。
具体实施方式
下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
现有深度学习的所有运算都是在计算机中通过运行程序所完成，即运算是全软件运算，本发明在现有技术深度学习的基础上，提出将计算机运算中的部分或全部运算通过本发明方案转换为光计算，实现提升深度学习计算效率。下面结合附图对本发明实施例进行详细介绍。
参见图1，为本发明实施例一种应用于深度学习的特征提取系统，包括：显示模组1、光学傅里叶变换装置2和光学滤波器3。其中：所述显示模组1用于将待特征提取的数字信号转换为待特征提取的空域图像信息，所述显示模组1可以为光纤扫描显示模组、OLED显示模组、LCD显示模组、LCoS显示模组或DLP显示模组等。所述光学傅里叶变换装置2用于将所述待特征提取的空域图像信息转换为待特征提取的频谱图像信息。待特征提取的空域图像信息照射在光学傅里叶变换装置的一端，光学傅里叶变换另一端便会输出其对应的频谱图像信息，即待特征提取的频谱图像信息。所述光学滤波器3用于将所述待特征提取的频谱图像信息转换为特征提取后的频谱图像信息。所述光学滤波器3是基于数字滤波器制作的，数字滤波器的频谱所对应的一定透过率的光器件，将该光器件置于输入图像的频谱面上，即可在该频谱面上实现两者频谱的点乘运算，该运算是在光域上进行的。
本发明实施例中，所述光学滤波器不同区域光透过率不同，所述光学滤波器可以为液晶光阀阵列或光刻透镜。光学滤波器的制作可采取如下方式：
1)将滤波器的数字信号转化为空域信号，再将该空域信号计算出其对应的频谱信号；2)将频谱信号对应的数字矩阵进行归一化得到归一化矩阵；3)根据归一化矩阵的每一个 元素制作具有一定透过率的光器件，即在一个光器件上制作与归一化矩阵相同行列数的调制区域矩阵，其中归一化矩阵中元素的数值若为1，则对应位置的调制区域透过率为百分之百，元素的数值若为0.5，则对应位置的调制区域透过率为百分之五十，其他调制区域依据同理制作(透过率的设定和改变可通过液晶态的转换来实现)，若元素为负数，其对应相位变化，该相位变化通过调制区域的厚度来调制。
本发明实施例得到的特征提取后的频谱图像信息可直接用于后续特征提取，在另一些实施例中，也可以如图2所示，将其输入至光学傅里叶逆变装置4，转换为特征提取后的空域图像信息。
在图2实施例中，所述系统还包括光探测器5，用于将所述特征提取后的空域图像信息转换为电信号，用于后续分析、存储、使用等。所述光探测器5可以为CCD探测器、光电二极管、光子型探测器或光电倍增管等。
在上述实施例中，所述光学傅里叶变换装置2和光学傅里叶逆变装置4均可以为傅里叶变换透镜或与傅里叶变换透镜具有相同功效的光学器件。
参见图3，为本发明实施例一种应用于深度学习的特征提取方法流程示意图，该方法主要包括：将待特征提取的数字信号通过显示技术转换为待特征提取的空域图像信息；将所述待特征提取的空域图像信息通过光学傅里叶变换装置转换为待特征提取的频谱图像信息；将所述待特征提取的频谱图像信息输入至该待特征提取的频谱图像信息对应的光学滤波器，得到特征提取后的频谱图像信息。其中，所述光学滤波器为液晶光阀阵列或光刻透镜，其不同区域光透过率不同，其制作可采取如下方式：
1)将滤波器的数字信号转化为空域信号，再将该空域信号计算出其对应的频谱信号；2)将频谱信号对应的数字矩阵进行归一化得到归一化矩阵；3)根据归一化矩阵的每一个元素制作具有一定透过率的光器件，即在一个光器件上制作与归一化矩阵相同行列数的调制区域矩阵，其中归一化矩阵中元素的数值若为1，则对应位置的调制区域透过率为百分之百，元素的数值若为0.5，则对应位置的调制区域透过率为百分之五十，其他调制区域依据同理制作(透过率的设定和改变可通过液晶态的转换来实现)，若元素为负数，其对应相位变化，该相位变化通过调制区域的厚度来调制。
现有计算中，深度学习中的特征提取需要将一组待特征提取的数字信号与滤波器的数字信号之间的卷积，本发明中将待特征提取的数字信号转换为光信号形式，将数字信号形式的滤波器制作为对应的光器件，此时光信号形式可简便地转换为频域，时域上的卷积等同于其对应频域上的点乘，时域上的卷积很复杂，频域点乘的运算会简单很多，因此本发 明可有效减小深度学习计算量，且本发明实施例中傅里叶变换过程采用傅里叶变换透镜等相同功效的光学器件进行变换，其变换速度为光速，极大地缩减傅里叶变换的耗时，且通过光学方案实现频谱的点乘运算，该点乘运算速度也为光速，极大缩减频谱点乘耗时。
本说明书中公开的所有特征，或公开的所有方法或过程中的步骤，除了互相排斥的特征和/或步骤以外，均可以以任何方式组合。
本说明书(包括任何附加权利要求、摘要和附图)中公开的任一特征，除非特别叙述，均可被其他等效或具有类似目的的替代特征加以替换。即，除非特别叙述，每个特征只是一系列等效或类似特征中的一个例子而已。
本发明并不局限于前述的具体实施方式。本发明扩展到任何在本说明书中披露的新特征或任何新的组合，以及披露的任一新的方法或过程的步骤或任何新的组合。

Claims (10)

  1. 一种应用于深度学习的特征提取系统，其特征在于，包括：显示模组、光学傅里叶变换装置和光学滤波器，其中：
    所述显示模组用于将待特征提取的数字信号转换为待特征提取的空域图像信息；
    所述光学傅里叶变换装置用于将所述待特征提取的空域图像信息转换为待特征提取的频谱图像信息；
    所述光学滤波器用于将所述待特征提取的频谱图像信息转换为特征提取后的频谱图像信息。
  2. 如权利要求1所述的特征提取系统，其特征在于，所述光学滤波器不同区域光透过率不同。
  3. 如权利要求2所述的特征提取系统，其特征在于，所述光学滤波器为液晶光阀阵列或光刻透镜。
  4. 如权利要求1至3任一项所述的特征提取系统，其特征在于，所述显示模组包括光纤扫描显示模组、OLED显示模组、LCD显示模组、LCoS显示模组和DLP显示模组中的一种。
  5. 如权利要求4所述的特征提取系统，其特征在于，所述系统还包括光学傅里叶逆变装置，用于将所述特征提取后的频谱图像信息转换为特征提取后的空域图像信息。
  6. 如权利要求5所述的特征提取系统，其特征在于，所述系统还包括光探测器，用于将所述特征提取后的空域图像信息转换为电信号。
  7. 如权利要求6所述的特征提取系统，其特征在于，所述光探测器包括CCD探测器、光电二极管、光子型探测器、光电倍增管中的一种。
  8. 如权利要求5所述的特征提取系统，其特征在于，所述光学傅里叶变换装置和光学傅里叶逆变装置均为傅里叶变换透镜。
  9. 一种应用于深度学习的特征提取方法，其特征在于，包括：
    将待特征提取的数字信号通过显示技术转换为待特征提取的空域图像信息；
    将所述待特征提取的空域图像信息通过光学傅里叶变换装置转换为待特征提取的频谱图像信息；
    将所述待特征提取的频谱图像信息输入至该信息对应的光学滤波器，得到特征提取后 的频谱图像信息。
  10. 如权利要求9所述的特征提取方法，其特征在于，所述光学滤波器为液晶光阀阵列或光刻透镜，其不同区域光透过率不同。
PCT/CN2019/085836 2018-05-15 2019-05-07 一种应用于深度学习的特征提取方法及系统 WO2019218895A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810460811.8 2018-05-15
CN201810460811.8A CN108805278B (zh) 2018-05-15 2018-05-15 一种应用于深度学习的特征提取方法及系统

Publications (1)

Publication Number Publication Date
WO2019218895A1 true WO2019218895A1 (zh) 2019-11-21

Family

ID=64092475

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/085836 WO2019218895A1 (zh) 2018-05-15 2019-05-07 一种应用于深度学习的特征提取方法及系统

Country Status (2)

Country Link
CN (1) CN108805278B (zh)
WO (1) WO2019218895A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112444959A (zh) * 2020-11-20 2021-03-05 中电科仪器仪表(安徽)有限公司 一种光纤图像虚影的处理装置及方法

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108805278B (zh) * 2018-05-15 2022-03-08 成都理想境界科技有限公司 一种应用于深度学习的特征提取方法及系统
CN111325198B (zh) * 2018-12-13 2023-05-16 北京地平线机器人技术研发有限公司 视频对象特征提取方法和装置、视频对象匹配方法和装置
CN110309916B (zh) 2019-06-05 2021-09-17 清华大学 多级空频域调制非线性全光学深度学习系统及方法
CN110400575B (zh) 2019-07-24 2024-03-29 腾讯科技（深圳）有限公司 通道间特征提取方法、音频分离方法和装置、计算设备

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101571919A (zh) * 2009-05-26 2009-11-04 重庆大学 基于光学非下采样轮廓波变换的人脸识别方法
CN103973976A (zh) * 2014-04-14 2014-08-06 杭州电子科技大学 一种利用光学成像的显著性提取装置及方法
CN203840438U (zh) * 2014-04-14 2014-09-17 杭州电子科技大学 一种利用光学成像的显著性提取装置
CN108805278A (zh) * 2018-05-15 2018-11-13 成都理想境界科技有限公司 一种应用于深度学习的特征提取方法及系统

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
AU2013344418B2 (en) * 2012-11-19 2017-09-07 Perkinelmer U.S. Llc Optical detectors and methods of using them
US20160026912A1 (en) * 2014-07-22 2016-01-28 Intel Corporation Weight-shifting mechanism for convolutional neural networks
CN106709441B (zh) * 2016-12-16 2019-01-29 北京工业大学 一种基于卷积定理的人脸验证加速方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101571919A (zh) * 2009-05-26 2009-11-04 重庆大学 基于光学非下采样轮廓波变换的人脸识别方法
CN103973976A (zh) * 2014-04-14 2014-08-06 杭州电子科技大学 一种利用光学成像的显著性提取装置及方法
CN203840438U (zh) * 2014-04-14 2014-09-17 杭州电子科技大学 一种利用光学成像的显著性提取装置
CN108805278A (zh) * 2018-05-15 2018-11-13 成都理想境界科技有限公司 一种应用于深度学习的特征提取方法及系统

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112444959A (zh) * 2020-11-20 2021-03-05 中电科仪器仪表(安徽)有限公司 一种光纤图像虚影的处理装置及方法
CN112444959B (zh) * 2020-11-20 2023-01-17 中电科思仪科技(安徽)有限公司 一种光纤图像虚影的处理装置及方法

Also Published As

Publication number Publication date
CN108805278B (zh) 2022-03-08
CN108805278A (zh) 2018-11-13

Similar Documents

Publication Publication Date Title
WO2019218895A1 (zh) 一种应用于深度学习的特征提取方法及系统
Monakhova et al. Learned reconstructions for practical mask-based lensless imaging
Barbastathis et al. On the use of deep learning for computational imaging
Li et al. Class-specific differential detection in diffractive optical neural networks improves inference accuracy
WO2018120329A1 (zh) 基于稀疏域重构的单帧图像超分辨重建方法及装置
JP2022056316A (ja) 文字構造化抽出方法及び装置、電子機器、記憶媒体並びにコンピュータプログラム
WO2020147282A1 (zh) 基于光学神经网络结构的图像识别方法、装置及电子设备
Choi et al. Optimal combination of image denoisers
US20210027166A1 (en) Dynamic pruning of neurons on-the-fly to accelerate neural network inferences
CN105550649A (zh) 基于全耦合局部约束表示的极低分辨率人脸识别方法及系统
WO2021179281A1 (en) Optimizing low precision inference models for deployment of deep neural networks
CN113052755A (zh) 一种基于深度学习的高分辨率图像智能化抠图方法
Fu et al. Multistage supervised contrastive learning for hybrid-degraded image restoration
CN113850298A (zh) 一种图像识别方法、装置及相关设备
CN112700000A (zh) 一种光电混合深度卷积神经网络方法
Yu et al. Gradient-descent-like ghost imaging
CN115761242B (zh) 一种基于卷积神经网络和模糊图像特征的去噪方法及终端机
WO2023201772A1 (zh) 基于迭代域内适应和自训练的跨域遥感图像语义分割方法
CN108805030B (zh) 一种应用于深度学习的特征提取方法及系统
Msuya et al. Deep Learning Model Compression Techniques: Advances, Opportunities, and Perspective
CN113919486A (zh) Vgg16衍射光子神经网络及面向情境依赖处理的方法
CN109360161B (zh) 一种基于梯度域先验的多光谱图像去模糊方法
Cao et al. A successive approach to enhancement of infrared facial images
Wang et al. Multispectral demosaicing via non-local low-rank regularization
CN113378660B (zh) 一种低数据成本的人脸识别的方法及装置

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: 19803057

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: 19803057

Country of ref document: EP

Kind code of ref document: A1