WO2020147282A1 - 基于光学神经网络结构的图像识别方法、装置及电子设备 - Google Patents

基于光学神经网络结构的图像识别方法、装置及电子设备 Download PDF

Info

Publication number
WO2020147282A1
WO2020147282A1 PCT/CN2019/096458 CN2019096458W WO2020147282A1 WO 2020147282 A1 WO2020147282 A1 WO 2020147282A1 CN 2019096458 W CN2019096458 W CN 2019096458W WO 2020147282 A1 WO2020147282 A1 WO 2020147282A1
Authority
WO
WIPO (PCT)
Prior art keywords
neural network
optical
layer
image
input
Prior art date
Application number
PCT/CN2019/096458
Other languages
English (en)
French (fr)
Inventor
翁文康
张笑鸣
Original Assignee
南方科技大学
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 南方科技大学 filed Critical 南方科技大学
Publication of WO2020147282A1 publication Critical patent/WO2020147282A1/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

Definitions

  • This application belongs to the field of data processing technology, and in particular relates to an image recognition method, device and electronic equipment based on an optical neural network structure.
  • graphics processors and tensor processors can accelerate deep learning algorithms, these hardware structures are often based on electronic components, and their calculation speed cannot exceed the theoretical limit of linear polynomial growth. This is likely to affect the speed and speed of image recognition and other operations. effectiveness.
  • the present application provides an image recognition method, image recognition device, electronic device, and computer-readable storage medium based on an optical neural network structure to improve the speed of image recognition.
  • the first aspect of the application provides an image recognition method based on an optical neural network structure.
  • the optical neural network structure is composed of X-layer neural networks, where X is a positive integer;
  • the image recognition method includes:
  • optical neural network structure is used for:
  • the input vector of the i-th layer of neural network For the i-th layer of neural network, obtain the input vector of the i-th layer of neural network.
  • the above i is a positive integer greater than 0 and less than X+1.
  • the input vector of the above-mentioned layer of neural network is based on the above Generated by each pixel of the image to be recognized; when i is greater than 1, the input vector of the i-th layer of neural network is the output vector of the i-1th layer of neural network;
  • the output vector of the neural network of this layer is the output result of the above-mentioned optical neural network structure, and the output of the above i-th layer neural network
  • Each element in the vector is used to indicate the possibility that the image to be recognized belongs to each different category.
  • the second aspect of the application provides an image recognition device based on an optical neural network structure.
  • the optical neural network structure is composed of X-layer neural networks, where X is a positive integer;
  • the image recognition device includes:
  • Image acquisition module for acquiring the image to be recognized
  • An image input module for inputting the image to be recognized into the optical neural network structure
  • the result recognition module is used to determine the recognition result of the image to be recognized based on the output result of the optical neural network structure
  • each layer of neural network of the above-mentioned optical neural network structure includes:
  • the vector input unit is used to obtain the input vector of the i-th neural network for the i-th neural network.
  • the above i is a positive integer greater than 0 and less than X+1.
  • the i-th neural network is The input vector of the network is generated based on each pixel of the image to be recognized; when i is greater than 1, the input vector of the i-th neural network is the output vector of the i-1th neural network;
  • the linear transformation unit is configured to perform linear transformation on the input vector based on the Yi inner product calculation units to obtain Yi linear transformation results, where the Yi is a positive integer;
  • the activation unit is used to activate the results of the above Yi linear transformations through a nonlinear crystal to obtain Yi activation results;
  • the result output unit is configured to use the above Yi activation results as the output vector of the neural network of this layer, where, when i is equal to X, the output vector of the i-th layer of neural network is the output result of the above-mentioned optical neural network structure.
  • Each element in the output vector of the i-layer neural network is used to indicate the possibility that the image to be recognized belongs to each different category.
  • a third aspect of the present application provides an electronic device.
  • the electronic device includes a memory, a processor, and a computer program that is stored in the memory and can run on the processor.
  • the processor executes the computer program, the above The steps of the method on the one hand.
  • a fourth aspect of the present application provides a computer-readable storage medium, the above-mentioned computer-readable storage medium stores a computer program, and when the above-mentioned computer program is executed by a processor, the steps of the method of the above first aspect are implemented.
  • a fifth aspect of the present application provides a computer program product.
  • the computer program product includes a computer program, and the computer program is executed by one or more processors to implement the steps of the method of the first aspect.
  • the image to be recognized is first acquired, and then the image to be recognized is input to the optical neural network structure, and then the recognition result of the image to be recognized is determined based on the output result of the optical neural network structure;
  • the above-mentioned optical neural network structure is composed of X-layer neural networks.
  • the above-mentioned X is a positive integer.
  • the above-mentioned optical neural network structure is used to obtain the input vector of the i-th layer neural network for the i-th layer neural network.
  • FIG. 1 is a schematic diagram of an implementation process of an image recognition method based on an optical neural network structure provided by an embodiment of the present application
  • FIG. 2 is a schematic diagram of a working flow of an optical neural network structure provided by an embodiment of the present application
  • FIG. 3 is a schematic diagram of an inner product calculation unit in an optical neural network structure provided by an embodiment of the present application.
  • FIG. 4 is a schematic structural diagram of an optical neural network structure provided by an embodiment of the present application.
  • FIG. 5 is a structural block diagram of an image recognition device based on an optical neural network structure provided by an embodiment of the present application
  • FIG. 6 is a structural block diagram of a single-layer neural network in the optical neural network structure provided by an embodiment of the present application.
  • Fig. 7 is a schematic diagram of an electronic device provided by an embodiment of the present application.
  • the image recognition method based on the optical neural network structure provided by the embodiment of the application includes:
  • step 101 an image to be recognized is acquired
  • the image to be recognized may be acquired by the electronic device first.
  • the camera application of the above-mentioned electronic device can be monitored, and after monitoring that the electronic device starts the camera through the camera application to perform a shooting operation Then, the captured picture is acquired as the image to be recognized, where the aforementioned camera can be a front camera or a rear camera, which is not limited here; or, if the aforementioned electronic device is an electronic device with social functions, it can be The social application program of the above electronic device monitors, and after monitoring that the above social application program receives pictures sent by other users, the received picture is used as the image to be recognized; or, if the above electronic device has a networking function, then It is also possible to monitor the browser-type application of the above electronic device. After monitoring that the user has downloaded the picture through the above-mentioned browser-type application, the downloaded picture
  • step 102 input the image to be recognized into the optical neural network structure
  • the aforementioned optical neural network structure is composed of X-layer neural networks, and the aforementioned X is a preset positive integer.
  • the input vector of the optical neural network structure is a list of vectors, where each layer of neural network is composed of linear transformation and nonlinear transformation. Activation function composition. Among them, for the input vector of the i-th layer, it must be linearly transformed first, and then nonlinearly transformed by the activation function.
  • step 201 for the i-th neural network, an input vector of the i-th neural network is obtained;
  • the aforementioned i is a positive integer greater than 0 and less than X+1.
  • the ith layer of neural network is the first layer of neural network; the input vector of the first layer of neural network is generated based on each pixel of the image to be recognized; when i is greater than 1, the ith layer The input vector of the layer neural network is the output vector of the i-1th layer neural network.
  • the above step 201 specifically includes:
  • the single coherent light source is equally divided into N optical signals
  • the above-mentioned N paths of optical signals are respectively encoded by an optical attenuator, and the input vector of the first layer neural network is constructed based on the amplitude of the encoded N paths of optical signals.
  • the above-mentioned optical attenuator and its encoding of the optical signal may have multiple implementation methods.
  • a part of the light field is drawn out through an adjustable beam splitter.
  • the amplitude of the input light field of the optical attenuator can be preset to Constant (for example, it can be set to 1), then the input beam can be divided into two lights by the adjustable beam splitter, and the weight of the two lights can be adjusted arbitrarily, so that the first light of the two lights is The amplitude is associated with the pixels of the image to be identified.
  • the first light is used as the output of the optical attenuator, and the second light is discarded to complete the encoding of the optical signal.
  • a single coherent light source is equally divided into N optical signals, and then the N optical signals are encoded through the optical attenuator encoding process, so that the encoding
  • the latter N optical signals are respectively associated with each pixel of the image to be identified.
  • the above N is set based on the number of pixels in the image to be recognized, that is, if there are N pixels in the image to be recognized, in the first layer of neural network, the coded light There are also N channels of signals, so that the N pixels in the image to be identified and the N channels of optical signals present a one-to-one correspondence.
  • step 202 linear transformation is performed on the input vector based on Yi inner product calculation units to obtain Yi linear transformation results
  • the above Yi is a preset positive integer, and when the Yi is set, it is related to the neural network of this layer. For example, when i is 1, Y1 is the number of inner product calculation units in the first layer of neural network; when i is 2, Y2 is the number of inner product calculation units in the second layer of neural network.
  • each inner product calculation unit contains M optical components, and the working process of a single inner product calculation unit is as follows:
  • one inner product calculation unit since one inner product calculation unit includes M optical components, the above-mentioned M input optical signals and M optical components also have a one-to-one correspondence; for example, the first input optical signal represents the input vector The first element is to input the first input optical signal into the first optical component to obtain the first output optical signal.
  • Mach-Zehnder interferometers can be used to combine the output light signals of M channels.
  • the Mach-Zehnder interferometer since the Mach-Zehnder interferometer has two input ports and two output ports, the output optical signals of two adjacent channels can be used as the input of a Mach-Zehnder interferometer, respectively.
  • one output port will be proportional to the sum of the amplitudes of the two input signals.
  • the inner product calculation unit proposed in the embodiment of this application only the above-mentioned proportional to two The output of the sum of the amplitudes of the input signals, the other output is discarded.
  • this series of Mach-Zehnder interferometers constitute a binary tree, the total input of which is M output optical signals, and only one final combined result is obtained in the end.
  • i is greater than 1
  • the output vector has Yi-1 elements; and this Yi-1 elements will constitute the input vector of the i-th layer neural network.
  • M is the number of elements contained in the input vector of the above-mentioned i-th layer neural network, the inner product of the non-first layer neural network
  • the calculation unit, the number of optical components contained in it M Yi-1.
  • M is the number of elements contained in the above-mentioned input vector, and the number of elements contained in the input vector of this layer is compared with the inner product calculation unit contained in the upper layer neural network The number is the same, therefore, the value of M can be determined by the number of neurons (that is, the inner product calculation unit) of the upper neural network.
  • Figure 3 shows that an inner product calculation unit is composed of two parts, assuming the value of M is 8.
  • the above inner product calculation unit can be seen as composed of two parts:
  • ⁇ 1, ⁇ 2 to ⁇ 8 are the 8 optical signals encoded by each element in the input vector, and ⁇ 1, ⁇ 2 to ⁇ 8 are the corresponding 8 optical components; first, the 8 optical signals are passed through the corresponding optical After the assembly, 8 output optical signals are obtained. In this way, the output of the first part will have 8 lights, and the amplitude of each light is ⁇ j* ⁇ j, and the value of j is greater than 0 and less than 9 (ie M+1) A positive integer.
  • the combination of light is completed by a preset number of Mach-Zehnder interferometers.
  • Fig. 3 it can be seen that every two adjacent paths of light need to be combined by retaining the Mach-Zehnder interferometer. Specifically, after the two output lights are obtained through the Mach-Zehnder interferometer, they are The reserved optical route is represented by a solid line, and the discarded output is represented by a dashed line.
  • the number of elements in the input vector of the inner product calculation unit must be 2z, where z is a positive integer. When the number of elements of the input vector is insufficient, it needs to be filled with 0 until the number of elements of the input vector reaches 2z.
  • the above ⁇ out is the amplitude of the output optical signal obtained by a single inner product calculation unit.
  • the inner product calculation unit can be used as a component of an optical circuit that deals with more complex problems, and the output optical signal is further processed in other parts.
  • the main advantages of this optical circuit are as follows: the circuit depth of the inner product calculation unit only grows in logarithmic form, which means that its calculation error only grows in logarithmic form, which greatly increases the robustness of the circuit .
  • step 203 the results of the above Yi linear transformations are activated through a nonlinear crystal to obtain Yi activation results
  • step 204 the above Yi activation results are used as the output vector of the neural network of this layer.
  • the output result of the i-th layer of neural network has no special meaning and is only used as the input vector of the i+1-th layer of neural network.
  • the output result of the i-th layer of neural network can be regarded as the intermediate transition parameter of the above-mentioned optical neural network structure; and when i is equal to X, the above-mentioned first
  • the output vector of the i-layer neural network (that is, the X-th layer neural network) is the final total output result of the above-mentioned optical neural network structure.
  • each element is used to indicate the possibility that the image to be recognized belongs to each different category.
  • optical neural network structure since all calculations are completed by optical elements, its energy consumption is extremely low and the processing speed is faster.
  • the initial input of the optical neural network structure is a list of vectors
  • Each layer of neural network is composed of linear transformation and nonlinear activation function. Assuming that the input of the nth layer is a vector The nth layer of neural network will first pass multiple inner product calculation units (encoded with corresponding calculation matrices) Make a linear transformation, and then through the nonlinear crystal by the activation function Perform a nonlinear transformation, and the output of this layer (that is, multiple F in the figure) is used as the input vector of the neural network of the next layer of the nth layer (that is, the n+1th layer) Of the various elements.
  • step 103 the recognition result of the image to be recognized is determined based on the output result of the optical neural network structure
  • the output result of the above-mentioned optical neural network structure represents the possibility that the above-mentioned image to be recognized belongs to different categories. For example, assuming that the above-mentioned optical neural network is used to determine whether the object in the image is a cat or a dog, the output result of the above-mentioned optical neural network structure will have two data, namely data A1 representing a dog and data A2 representing a cat. If A2 > A1, it can be determined that the recognition result of the image to be recognized is a cat; if A1>A2, it can be determined that the recognition result of the image to be recognized is a dog.
  • the foregoing image recognition method further includes:
  • Obtaining a preset conversion matrix where the preset conversion matrix is related to the i-th layer of neural network, and the dimension of the preset conversion matrix is M*Yi;
  • the above-mentioned inputting the foregoing M channels of input optical signals to the corresponding optical components respectively to obtain M channels of output optical signals includes:
  • the M channels of input optical signals are respectively input to the corresponding optical components, so that the input vectors are respectively subjected to inner product operations with the calculation matrix coded by the inner product calculation unit to obtain M channels of output optical signals.
  • each layer of neural network has its own conversion matrix, and the dimension of the conversion matrix is M*Yi, where M is the number of elements of the input vector of the layer of neural network;
  • the above-mentioned optical component includes an optical attenuator and a polarizer;
  • the above-mentioned encoding of the above-mentioned Yi calculation matrices into the corresponding inner product calculation unit includes: for any calculation matrix, the absolute value of each element of the calculation matrix The value is encoded on the optical attenuator of the corresponding optical component, and the symbol of each element of the above calculation matrix is respectively encoded on the polarizer of the corresponding optical component.
  • the symbol of each element above refers to the sign of each element.
  • the element is a positive number, the "+" sign is encoded on the polarizer of the corresponding optical component of the element; if the element is a negative number, the The "-" sign is encoded on the polarizer of the corresponding optical component of the element.
  • An embodiment of the application provides an image recognition device based on an optical neural network structure.
  • the above-mentioned optical neural network structure is composed of X-layer neural networks, where X is a positive integer; please refer to FIG. 5, the image recognition device 500 includes:
  • the image acquisition module 501 is used to acquire the image to be recognized
  • the image input module 502 is configured to input the image to be recognized into the optical neural network structure
  • the result recognition module 503 is configured to determine the recognition result of the image to be recognized based on the output result of the optical neural network structure
  • each layer of the neural network of the above optical neural network structure includes:
  • the vector input unit 601 is used to obtain the input vector of the i-th neural network for the i-th neural network.
  • the above i is a positive integer greater than 0 and less than X+1.
  • the i-th layer The input vector of the neural network is generated based on each pixel of the image to be recognized; when i is greater than 1, the input vector of the i-th layer of neural network is the output vector of the i-1th layer of neural network;
  • the linear transformation unit 602 is configured to perform linear transformation on the input vector based on the Yi inner product calculation units to obtain Yi linear transformation results, where the Yi is a positive integer;
  • the activation unit 603 is configured to activate the results of the above Yi linear transformations through a nonlinear crystal to obtain Yi activation results;
  • the result output unit 604 is configured to use the above-mentioned Yi activation results as the output vector of the neural network of this layer, where, when i is equal to X, the output vector of the i-th layer of neural network is the output result of the above-mentioned optical neural network structure.
  • Each element in the output vector of the i-th layer neural network is used to indicate the possibility that the image to be recognized belongs to each different category.
  • the vector input unit of the first layer neural network of the foregoing optical neural network structure includes:
  • the light source acquisition subunit is used to acquire a single coherent light source
  • Molecular units such as light sources are used to equally divide a single coherent light source into N optical signals;
  • the light source encoding subunit is used to encode the above N channels of optical signals through an optical attenuator, and construct the input vector of the first layer neural network based on the amplitude of the encoded N channels of optical signals, wherein the above encoded N channels
  • the light signals are respectively associated with each pixel of the image to be identified, and the N is set based on the number of pixels of the image to be identified.
  • the linear transformation unit of each layer of the neural network of the optical neural network structure is specifically used to encode each element contained in the input vector to obtain M input optical signals, where M is the input vector The number of elements contained;
  • the linear transformation unit is specifically also used to input the M input optical signals to the corresponding optical components for any inner product calculation unit to obtain M output optical signals, wherein the inner product calculation unit includes M optical components; combine the M output optical signals to obtain an inner product calculation result as a linear transformation result obtained by linear transformation of the input vector based on the inner product calculation unit.
  • each layer of neural network also includes:
  • the conversion matrix obtaining unit is configured to obtain a preset conversion matrix, where the preset conversion matrix is related to the i-th layer of neural network, and the dimension of the preset conversion matrix is M*Yi;
  • the conversion matrix splitting unit splits the above conversion matrix into Yi M*1 matrices, which are recorded as Yi calculation matrices;
  • a calculation matrix coding unit configured to respectively encode the above Yi calculation matrices into corresponding inner product calculation units
  • the linear transformation unit is specifically configured to input the M input optical signals to the corresponding optical components, so that the input vector and the calculation matrix encoded by the inner product calculation unit perform inner product calculations to obtain M channels output optical signals.
  • the above-mentioned optical component includes an optical attenuator and a polarizer; the above-mentioned calculation matrix encoding unit is specifically configured to encode the absolute value of each element of the above-mentioned calculation matrix to the optical attenuator of the corresponding optical component for any calculation matrix.
  • the symbols of each element of the above calculation matrix are respectively coded onto the polarizers of the corresponding optical components.
  • the image recognition device recognizes images through the above-mentioned new type of optical neural network structure. Since all calculations are completed by optical elements, it can The consumption is extremely low, and the processing speed is fast, and the result of image recognition can be obtained quickly.
  • An embodiment of the present application provides an electronic device. Please refer to FIG. 7.
  • the electronic device in the embodiment of the present application includes: a memory 701, one or more processors 702 (only one is shown in FIG. 7) and stored in the memory 701 A computer program that can run on the processor.
  • the memory 701 is used to store software programs and modules, and the processor 702 executes various functional applications and data processing by running the software programs and units stored in the memory 701 to obtain resources corresponding to the aforementioned preset events.
  • the processor 702 implements the following steps when running the foregoing computer program stored in the memory 701:
  • the above-mentioned optical neural network structure is composed of X-layer neural networks, and the above-mentioned X is a positive integer; the above-mentioned optical neural network structure is used for:
  • the input vector of the i-th layer of neural network For the i-th layer of neural network, obtain the input vector of the i-th layer of neural network.
  • the above i is a positive integer greater than 0 and less than X+1.
  • the input vector of the above-mentioned layer of neural network is based on the above Generated by each pixel of the image to be recognized; when i is greater than 1, the input vector of the i-th layer of neural network is the output vector of the i-1th layer of neural network;
  • the output vector of the neural network of this layer is the output result of the above-mentioned optical neural network structure, and the output of the above i-th layer neural network
  • Each element in the vector is used to indicate the possibility that the image to be recognized belongs to each different category.
  • the above N channels of optical signals are respectively encoded by an optical attenuator, and the input vector of the first layer neural network is constructed based on the amplitude of the encoded N channels of optical signals, wherein the above encoded N channels of optical signals are respectively compared with the above to be identified Each pixel of the image is associated, and the above N is set based on the number of pixels of the image to be recognized.
  • the above-mentioned linear transformation is performed on the above-mentioned input vector based on Yi inner product calculation units respectively to obtain Yi linear transformation results, including:
  • the inner product calculation unit includes M optical components, and the M input optical signals are respectively input to the corresponding optical components to obtain M output optical signals;
  • Combining the foregoing M output optical signals to obtain an inner product calculation result is used as a linear transformation result obtained by performing linear transformation on the input vector based on the inner product calculation unit.
  • the processor 702 implements the following steps when running the foregoing computer program stored in the memory 701:
  • Obtaining a preset conversion matrix where the preset conversion matrix is related to the i-th layer of neural network, and the dimension of the preset conversion matrix is M*Yi;
  • the above-mentioned inputting the foregoing M channels of input optical signals to the corresponding optical components respectively to obtain M channels of output optical signals includes:
  • the M paths of input optical signals are respectively input to the corresponding optical components, so that the input vector and the calculation matrix encoded by the inner product calculation unit are respectively subjected to inner product operations to obtain M paths of output optical signals.
  • the foregoing optical component includes an optical attenuator and a polarizer; the foregoing Yi calculation matrices are respectively coded to the corresponding inner product calculation unit Include:
  • the absolute value of each element of the calculation matrix is coded to the optical attenuator of the corresponding optical component, and the symbol of each element of the calculation matrix is coded to the polarizer of the corresponding optical component.
  • the above electronic device may further include: one or more input devices 703 (only one is shown in FIG. 7) and one or more output devices 704 (only one is shown in FIG. 7).
  • the memory 701, the processor 702, the input device 703, and the output device 704 are connected by a bus 705.
  • the processor 702 may be a central processing unit (Central Processing Unit, CPU), and the processor may also be other general-purpose processors or digital signal processors (DSP). , Application Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc.
  • the general-purpose processor may be a microprocessor or the processor may be any conventional processor or the like.
  • the input device 703 may include a keyboard, a touch panel, a fingerprint sensor (used to collect user fingerprint information and fingerprint orientation information), a microphone, etc.
  • the output device 704 may include a display, a speaker, and the like.
  • the memory 701 may include a read-only memory and a random access memory, and provides instructions and data to the processor 702. A part or all of the memory 701 may also include a non-volatile random access memory. For example, the memory 701 may also store device type information.
  • the disclosed device and method may be implemented in other ways.
  • the system embodiment described above is only illustrative.
  • the division of the above-mentioned modules or units is only a logical function division.
  • there may be other division methods for example, multiple units or components may be combined. Or it can be integrated into another system, or some features can be ignored or not implemented.
  • the displayed or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, devices or units, and may be electrical, mechanical or other forms.
  • the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
  • the above integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium.
  • this application implements all or part of the processes in the above-mentioned embodiments and methods, and can also be completed by instructing relevant hardware through a computer program.
  • the above-mentioned computer program can be stored in a computer-readable storage medium. When executed by the processor, the steps of the foregoing method embodiments can be implemented.
  • the above-mentioned computer program includes computer program code, and the above-mentioned computer program code may be in the form of source code, object code, executable file, or some intermediate forms.
  • the above-mentioned computer-readable medium may include: any entity or device capable of carrying the above-mentioned computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random Access memory (RAM, Random Access Memory), electrical carrier signal, telecommunications signal, and software distribution media, etc.
  • ROM Read-Only Memory
  • RAM Random Access Memory
  • electrical carrier signal telecommunications signal
  • telecommunications signal and software distribution media, etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Image Analysis (AREA)

Abstract

本申请公开了一种基于光学神经网络结构的图像识别方法、图像识别装置及电子设备，其中，光学神经网络结构由X层神经网络所构成；该图像识别方法包括：获取待识别图像；将待识别图像输入至光学神经网络结构；基于光学神经网络结构的输出结果确定待识别图像的识别结果；其中，光学神经网络结构用于：针对第i层神经网络，获取第i层神经网络的输入向量，i为大于0且小于X+1的正整数；分别基于Yi个内积计算单元对输入向量进行线性变换，得到Yi个线性变换的结果；将Yi个线性变换的结果通过非线性晶体进行激活，得到Yi个激活结果；将Yi个激活结果作为本层神经网络的输出向量。本申请方案应用了新型的光学神经网络结构，进一步提升了图像识别的速度。

Description

基于光学神经网络结构的图像识别方法、装置及电子设备 技术领域
本申请属于数据处理技术领域，尤其涉及一种基于光学神经网络结构的图像识别方法、装置及电子设备。
背景技术
当前，机器学习已经成为一种十分重要的工具。其中，基于深度神经网络的深度学习受到了广泛的关注，并被应用于图像识别、语音识别、自然语言翻译等重要领域。其中，基于传统中央处理器(Central Processing Unit，CPU)的深度学习并非最优的方案；研发人员发展出了多样的硬件结构，以适应深度学习算法的要求，例如图形处理器(Graphics Processing Unit，GPU)和张量处理器(Tensor Processing Unit，TPU)。
技术问题
虽然图形处理器和张量处理器都能加速深度学习算法，但这些硬件结构往往基于电子元件，其计算速度无法超越线性多项式增长的理论极限，这很可能影响到图像识别等操作时的速度及效率。
技术解决方案
有鉴于此，本申请提供了一种基于光学神经网络结构的图像识别方法、图像识别装置、电子设备及计算机可读存储介质，用以提升图像识别的速度。
本申请的第一方面提供了一种基于光学神经网络结构的图像识别方法，上述光学神经网络结构由X层神经网络所构成，上述X为正整数；上述图像识别方法包括：
获取待识别图像；
将上述待识别图像输入至上述光学神经网络结构；
基于上述光学神经网络结构的输出结果确定上述待识别图像的识别结果；
其中，上述光学神经网络结构用于：
针对第i层神经网络，获取第i层神经网络的输入向量，上述i为大于0且小于X+1的正整数，其中，当i等于1时，上述第i层神经网络的输入向量基于上述待识别图像的各个像素点而生成；当i大于1时，上述第i层神经网络的输入向量为第i-1层神经网络的输出向量；
分别基于Yi个内积计算单元对上述输入向量进行线性变换，得到Yi个线性变换的结果，其中，上述Yi为正整数；
将上述Yi个线性变换的结果通过非线性晶体进行激活，得到Yi个激活结果；
将上述Yi个激活结果作为本层神经网络的输出向量，其中，当i等于X时，上述第i层神经网络的输出向量为上述光学神经网络结构的输出结果，上述第i层神经网络的输出 向量中的各个元素用于指示上述待识别图像属于各个不同类别的可能性。
本申请的第二方面提供了一种基于光学神经网络结构的图像识别装置，上述光学神经网络结构由X层神经网络所构成，上述X为正整数；上述图像识别装置包括：
图像获取模块，用于获取待识别图像；
图像输入模块，用于将上述待识别图像输入至上述光学神经网络结构；
结果识别模块，用于基于上述光学神经网络结构的输出结果确定上述待识别图像的识别结果；
其中，上述光学神经网络结构的各层神经网络均包括：
向量输入单元，用于针对第i层神经网络，获取第i层神经网络的输入向量，上述i为大于0且小于X+1的正整数，其中，当i等于1时，上述第i层神经网络的输入向量基于上述待识别图像的各个像素点而生成；当i大于1时，上述第i层神经网络的输入向量为第i-1层神经网络的输出向量；
线性变换单元，用于分别基于Yi个内积计算单元对上述输入向量进行线性变换，得到Yi个线性变换的结果，其中，上述Yi为正整数；
激活单元，用于将上述Yi个线性变换的结果通过非线性晶体进行激活，得到Yi个激活结果；
结果输出单元，用于将上述Yi个激活结果作为本层神经网络的输出向量，其中，当i等于X时，上述第i层神经网络的输出向量为上述光学神经网络结构的输出结果，上述第i层神经网络的输出向量中的各个元素用于指示上述待识别图像属于各个不同类别的可能性。
本申请的第三方面提供了一种电子设备，上述电子设备包括存储器、处理器以及存储在上述存储器中并可在上述处理器上运行的计算机程序，上述处理器执行上述计算机程序时实现如上第一方面的方法的步骤。
本申请的第四方面提供了一种计算机可读存储介质，上述计算机可读存储介质存储有计算机程序，上述计算机程序被处理器执行时实现如上第一方面的方法的步骤。
本申请的第五方面提供了一种计算机程序产品，上述计算机程序产品包括计算机程序，上述计算机程序被一个或多个处理器执行时实现如上述第一方面的方法的步骤。
有益效果
由上可见，通过本申请方案，首先获取待识别图像，然后将上述待识别图像输入至上述光学神经网络结构，接着基于上述光学神经网络结构的输出结果确定上述待识别图像的识别结果；其中，上述光学神经网络结构由X层神经网络所构成，上述X为正整数，上述光学神经网络结构用于：针对第i层神经网络，获取第i层神经网络的输入向量，上述i为 大于0且小于X+1的正整数，其中，当i等于1时，上述第i层神经网络的输入向量基于上述待识别图像的各个像素点而生成；当i大于1时，上述第i层神经网络的输入向量为第i-1层神经网络的输出向量；分别基于Yi个内积计算单元对上述输入向量进行线性变换，得到Yi个线性变换的结果，其中，上述Yi为正整数；将上述Yi个线性变换的结果通过非线性晶体进行激活，得到Yi个激活结果；将上述Yi个激活结果作为本层神经网络的输出向量，其中，当i等于X时，上述第i层神经网络的输出向量为上述光学神经网络结构的输出结果，上述第i层神经网络的输出向量中的各个元素用于指示上述待识别图像属于各个不同类别的可能性。通过本申请方案，提出了一种新型的光学神经网络结构，通过上述新型的光学神经网络结构对图像进行识别，由于其中的所有计算皆由光学元件完成，因而能耗极低，并且处理速度快，能够快速获得图像识别的结果。
附图说明
为了更清楚地说明本申请实施例中的技术方案，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申请的一些实施例，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是本申请实施例提供的基于光学神经网络结构的图像识别方法的实现流程示意图；
图2是本申请实施例提供的光学神经网络结构的工作流程示意图；
图3是本申请实施例提供的光学神经网络结构中，内积计算单元的示意图；
图4是本申请实施例提供的光学神经网络结构的结构示意图；
图5是本申请实施例提供的基于光学神经网络结构的图像识别装置的结构框图；
图6是本申请实施例提供的光学神经网络结构中单层神经网络的结构框图；
图7是本申请实施例提供的电子设备的示意图。
具体实施方式
以下描述中，为了说明而不是为了限定，提出了诸如特定系统结构、技术之类的具体细节，以便透彻理解本申请实施例。然而，本领域的技术人员应当清楚，在没有这些具体细节的其它实施例中也可以实现本申请。在其它情况中，省略对众所周知的系统、装置、电路以及方法的详细说明，以免不必要的细节妨碍本申请的描述。
为了说明本申请上述的技术方案，下面通过具体实施例来进行说明。
实施例1
下面对本申请实施例提供的基于光学神经网络结构的图像识别方法进行描述，请参阅图1，本申请实施例中的基于光学神经网络结构的图像识别方法包括：
在步骤101中，获取待识别图像；
在本申请实施例中，可以先由电子设备获取待识别图像。可选地，若上述电子设备为智能手机、平板电脑等具备拍摄功能的电子设备，则可以对上述电子设备的相机应用程序进行监听，在监听到电子设备通过相机应用程序启动摄像头进行了拍摄操作后，获取拍摄的图片作为待识别图像，其中，上述摄像头可以为前置摄像头，也可以为后置摄像头，此处不作限定；或者，若上述电子设备为具备社交功能的电子设备，则可以对上述电子设备的社交类应用程序进行监听，在监听到上述社交类应用程序中接收到了其它用户发送的图片后，将接收到的图片作为待识别图像；或者，若上述电子设备具备联网功能，则还可以对上述电子设备的浏览器类应用程序进行监听，在监听到用户通过上述浏览器类应用程序下载了图片后，将下载得到的图片作为待识别图像；当然，也可以通过其他方式获取待识别图像，此处不作限定。
在步骤102中，将上述待识别图像输入至上述光学神经网络结构；
在本申请实施例中，需要预先训练好上述光学神经网络结构，然后将上述待识别图像输入至上述训练好的光学神经网络结构中。具体地，上述光学神经网络结构由X层神经网络所构成，上述X为预设的正整数。为了更好的说明本申请实施例的方案，首先对上述光学神经网络结构进行说明，具体地，该光学神经网络结构的输入向量为一列向量，其中，每一层神经网络由线性变换和非线性激活函数组成。其中，针对第i层的输入向量,须先对其作线性变换，然后由激活函数对进行非线性变换，得到该第i层的输出结果后，若该第i层之后还有第i+1层神经网络，则将该第i层的输出结果作为第i+1层神经网络的输入向量的各个元素。具体地，此处对上述光学神经网络结构中每一层神经网络的工作流程作出说明，请参阅图2：
在步骤201中，针对第i层神经网络，获取第i层神经网络的输入向量；
在本申请实施例中，上述i为大于0且小于X+1的正整数。当i等于1时，上述第i层神经网络即为第1层神经网络；该第1层神经网络的输入向量基于上述待识别图像的各个像素点而生成；当i大于1时，上述第i层神经网络的输入向量为第i-1层神经网络的输出向量。
具体地，针对第1层神经网络，上述步骤201具体包括：
A1、获取单一相干光源；
A2、将单一相干光源等分为N路光信号；
A3、将上述N路光信号分别通过光学衰减器进行编码，并基于编码后的N路光信号的振幅构建第1层神经网络的输入向量。
在本申请实施例中，上述光学衰减器及其对光信号的编码可以有多种实现方法。例如， 通过可调分束器将一部分的光场引出。为了更好的进行举例说明，避免丧失一般性，假设所需要编码的输入向量中的元素为0到1之间的某个值，则光学衰减器的输入光场的振幅大小可以被预先设置为常数(例如，可以被设置为1)，随后通过可调分束器可以将输入光束分为两路光，并且两路光的权重可以任意调整，使得两路光的中的第一路光的振幅与上述待识别图像的像素点相关联。将第一路光作为光学衰减器的输出，将第二路光舍弃，即可完成对光信号的编码。也即是说，在第1层神经网络中，首先将单一相干光源等分为N路光信号后，通过上述光学衰减器对光信号的编码过程实现对N路光信号的编码，使得上述编码后的N路光信号分别与上述待识别图像的各个像素点相关联。具体地，上述N基于上述待识别图像的像素点的个数而设定，也即，若上述待识别图像中的像素点有N个，则在第1层神经网络中，其所编码得到光信号也有N路，使得待识别图像中N个像素点与N路光信号呈现一一对应的关系。
在步骤202中，分别基于Yi个内积计算单元对上述输入向量进行线性变换，得到Yi个线性变换的结果；
在本申请实施例中，上述Yi为预设的正整数，且该Yi在设置时，与本层神经网络相关。例如，i为1时，Y1即为第1层神经网络中内积计算单元的个数；i为2时，Y2即为第2层神经网络中内积计算单元的个数。具体地，在同一层神经网络中，每一个内积计算单元均包含有M个光学组件，其中，单一内积计算单元的工作过程如下：
B1、分别对上述输入向量所包含的各个元素进行编码，得到M路输入光信号，其中，M为上述输入向量所包含的元素的个数；
B2、将上述M路输入光信号分别输入至相应的光学组件上，得到M路输出光信号；
其中，由于一个内积计算单元均包含有M个光学组件，因而，上述M路输入光信号与M个光学组件也呈现一一对应的关系；例如，第一路输入光信号代表输入向量中的第一个元素，将该第一路输入光信号输入至第一个光学组件中，得到第一路输出光信号。
B3、将上述M路输出光信号进行合束后得到内积计算结果，作为基于上述内积计算单元对上述输入向量进行线性变换所得到的线性变换的结果。
其中，可以通过若干个马赫-曾德耳干涉仪来实现M路输出光信号的合束。具体地，由于马赫-曾德耳干涉仪有两个输入口和两个输出口，因而，可以将相邻两路的输出光信号分别作为一马赫-曾德耳干涉仪的输入，所得到的该马赫-曾德耳干涉仪的两个输出中，有一个输出口将正比于两个输入信号的振幅之和，在本申请实施例所提出的内积计算单元中，仅保留上述正比于两个输入信号的振幅之和的输出，将另一输出丢弃。可以认为，这一系列的马赫-曾德耳干涉仪构成一个二叉树，其总输入为即为M路输出光信号，并最终只得到一个最终的合束结果。具体地，当i大于1时，由于针对第i-1层神经网络来说，该层神 经网络中有Yi-1个内积计算单元，因而其输出向量中有Yi-1个元素；而这Yi-1个元素又将构成第i层神经网络的输入向量，由于上述M为上述第i层神经网络的输入向量所包含的元素的个数，因而，针对非首层神经网络中的内积计算单元，其所包含的光学组件的数量M＝Yi-1。也即是说，针对非首层神经网络，由于M为上述输入向量所包含的元素的个数，而该层输入向量所包含的元素的个数与上层神经网络中包含的内积计算单元的个数相同，因而，此处可以通过上层神经网络的神经元(也即内积计算单元)的个数而确定M的取值。
为了更好的说明内积计算单元的工作过程，请参阅图3：
图3示出了一个内积计算单元由两个部分所组成，假定M的取值为8。上述内积计算单元可看作由两部分所构成：
在第一部分中，α1、α2至α8为输入向量中的各个元素所编码得到的8路光信号，ω1、ω2至ω8为相应的8个光学组件；首先依次将8路光信号通过相应的光学组件后，得到8路输出光信号，如此，第一部分的输出将有8路光，每一路光的振幅分别为αj*ωj，j的取值为大于0，小于9(也即M+1)的正整数。
在第二部分中，通过预设数量的马赫-曾德耳干涉仪完成光线的合束。根据图3可知，每相邻两路光线之间均需要通过保留马赫-曾德耳干涉仪实现光线的合束，具体地，在通过马赫-曾德耳干涉仪得到两路输出光之后，被保留的光路由实线表示，被舍弃的输出由虚线表示。需要注意的是，内积计算单元的输入向量的元素个数需为2z，其中，z为正整数。当出现输入向量的元素个数不足时，需要以0进行补齐，直至输入向量的元素个数达到2z为止。
在通过上述第一部分及第二部分后，该内积计算单元所最终得到的输出光信号的振幅为
Figure PCTCN2019096458-appb-000001
上述βout为单个内积计算单元所得到的输出光信号的振幅。通过对输出光信号的相位以及振幅进行测量，并乘以系数
Figure PCTCN2019096458-appb-000002
便可得到β的数值大小；或者，内积计算单元可以作为处理更复杂问题的光学线路的元件，其输出光信号在其它部分进行进一步处理。对比现有技术，这种光学线路的主要优势如下：内积计算单元的线路深度仅仅以对数形式增长，这意味着其计算误差也仅以对数形式增长，大大增加了线路的鲁棒性。
在步骤203中，将上述Yi个线性变换的结果通过非线性晶体进行激活，得到Yi个激活结果；
在步骤204中，将上述Yi个激活结果作为本层神经网络的输出向量。
在本申请实施例中，当i为大于0且小于X的整数时，上述第i层神经网络的输出结 果没有特殊的含义，仅作为第i+1层神经网络的输入向量。也即是说，当i为大于0且小于X的整数时，上述第i层神经网络的输出结果可以被看做是上述光学神经网络结构的中间过渡参数；而当i等于X时，上述第i层神经网络(也即第X层神经网络)的输出向量即为上述光学神经网络结构的最终总的输出结果，上述第i层神经网络(也即第X层神经网络)的输出向量中的各个元素用于指示上述待识别图像属于各个不同类别的可能性。在上述光学神经网络结构中，由于所有计算皆由光学元件完成，因而其能耗极低，并且处理速度更快。
为了更好的说明上述光学神经网络结构，下面对上述光学神经网络结构的工作流程进行概述，请参阅图4：
光学神经网络结构的初始输入为一列向量
Figure PCTCN2019096458-appb-000003
每一层神经网络由线性变换和非线性激活函数组成。假定第n层的输入为向量
Figure PCTCN2019096458-appb-000004
该第n层神经网络将首先通过多个内积计算单元(分别编码有相应的计算矩阵)对
Figure PCTCN2019096458-appb-000005
作线性变换，随后通过非线性晶体由激活函数对
Figure PCTCN2019096458-appb-000006
进行非线性变换，所得到的该层的输出(也即图中的多个F)作为第n层的下一层(也即第n+1层)神经网络的输入向量
Figure PCTCN2019096458-appb-000007
的各个元素。
在步骤103中，基于上述光学神经网络结构的输出结果确定上述待识别图像的识别结果；
在本申请实施例中，上述光学神经网络结构的输出结果代表了上述待识别图像属于各个不同类别的可能性。例如，假定通过上述光学神经网络判断图像中物体是猫还是狗，则上述光学神经网络结构的输出结果将有两个数据，分别为代表属于狗的数据A1及代表属于猫的数据A2，若A2＞A1，则可以确定上述待识别图像的识别结果为猫；若A1＞A2，则可以确定上述待识别图像的识别结果为狗。
可选地，上述图像识别方法还包括：
获取预设的转换矩阵，其中，上述预设的转换矩阵与第i层神经网络相关，上述预设的转换矩阵的维度为M*Yi；
将上述转换矩阵拆分为Yi个M*1的矩阵，记为Yi个计算矩阵；
将上述Yi个计算矩阵分别编码至相应的内积计算单元中；
相应地，上述将上述M路输入光信号分别输入至相应的光学组件上，得到M路输出光信号，包括：
将上述M路输入光信号分别输入至相应的光学组件上，以使得上述输入向量分别与上 述内积计算单元所编码的计算矩阵进行内积运算，得到M路输出光信号。
在本申请实施例中，每一层神经网络都有各自的转换矩阵，将且上述转换矩阵的维度为M*Yi，其中M即为该层神经网络的输入向量的元素个数；将上述转换矩阵拆分为Yi个M*1的矩阵后，由于将Yi个计算矩阵分别编码至相应的内积计算单元中，因而，会有Yi个内积计算单元进行内积计算；也即是说，上述Yi个内积计算单元的输入虽然一致(均为本层神经网络的输入向量)，但由于Yi个内积计算单元的计算矩阵并不一定相同，因而会得到不同的输出结果；具体地，由于一个内积计算单元只有一个输出，因而，Yi个内积计算单元将会有Yi个输出，也即是说，一层神经网络将会得到Yi个输出结果。通过上述过程，即实现了维度的转换。具体地，上述光学组件包括光学衰减器及偏振片；上述将上述Yi个计算矩阵分别编码至相应的内积计算单元中，包括：针对任一计算矩阵，分别将上述计算矩阵的各个元素的绝对值编码至相应光学组件的光学衰减器上，分别将上述计算矩阵的各个元素的符号编码至相应光学组件的偏振片上。具体地，上述各个元素的符号指的是各个元素的正负号，如该元素为正数，则将“+”号编码至该元素相应光学组件的偏振片上；如该元素为负数，则将“-”号编码至该元素相应光学组件的偏振片上。
由上可见，通过本申请实施例，提出了一种新型的光学神经网络结构，通过上述新型的光学神经网络结构对图像进行识别，由于其中的所有计算皆由光学元件完成，因而能耗极低，并且处理速度快，能够快速获得图像识别的结果。
应理解，上述实施例中各步骤的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定，而不应对本申请实施例的实施过程构成任何限定。
实施例2
本申请实施例提供了一种基于光学神经网络结构的图像识别装置，上述光学神经网络结构由X层神经网络所构成，上述X为正整数；请参阅图5，上述图像识别装置500包括：
图像获取模块501，用于获取待识别图像；
图像输入模块502，用于将上述待识别图像输入至上述光学神经网络结构；
结果识别模块503，用于基于上述光学神经网络结构的输出结果确定上述待识别图像的识别结果；
其中，请参阅图6，上述光学神经网络结构的各层神经网络均包括：
向量输入单元601，用于针对第i层神经网络，获取第i层神经网络的输入向量，上述i为大于0且小于X+1的正整数，其中，当i等于1时，上述第i层神经网络的输入向量基于上述待识别图像的各个像素点而生成；当i大于1时，上述第i层神经网络的输入向量为第i-1层神经网络的输出向量；
线性变换单元602，用于分别基于Yi个内积计算单元对上述输入向量进行线性变换，得到Yi个线性变换的结果，其中，上述Yi为正整数；
激活单元603，用于将上述Yi个线性变换的结果通过非线性晶体进行激活，得到Yi个激活结果；
结果输出单元604，用于将上述Yi个激活结果作为本层神经网络的输出向量，其中，当i等于X时，上述第i层神经网络的输出向量为上述光学神经网络结构的输出结果，上述第i层神经网络的输出向量中的各个元素用于指示上述待识别图像属于各个不同类别的可能性。
可选地，上述光学神经网络结构的第1层神经网络的向量输入单元，包括：
光源获取子单元，用于获取单一相干光源；
光源等分子单元，用于将单一相干光源等分为N路光信号；
光源编码子单元，用于将上述N路光信号分别通过光学衰减器进行编码，并基于编码后的N路光信号的振幅构建第1层神经网络的输入向量，其中，上述编码后的N路光信号分别与上述待识别图像的各个像素点相关联，上述N基于上述待识别图像的像素点的个数而设定。
可选地，上述光学神经网络结构的各层神经网络的线性变换单元，具体用于分别对上述输入向量所包含的各个元素进行编码，得到M路输入光信号，其中，M为上述输入向量所包含的元素的个数；
上述线性变换单元，具体还用于针对任一内积计算单元，将上述M路输入光信号分别输入至相应的光学组件上，得到M路输出光信号，其中，上述内积计算单元中包含有M个光学组件；将上述M路输出光信号进行合束后得到内积计算结果，作为基于上述内积计算单元对上述输入向量进行线性变换所得到的线性变换的结果。
可选地，各层神经网络中还包括：
转换矩阵获取单元，用于获取预设的转换矩阵，其中，上述预设的转换矩阵与第i层神经网络相关，上述预设的转换矩阵的维度为M*Yi；
转换矩阵拆分单元，将上述转换矩阵拆分为Yi个M*1的矩阵，记为Yi个计算矩阵；
计算矩阵编码单元，用于将上述Yi个计算矩阵分别编码至相应的内积计算单元中；
相应地，上述线性变换单元，具体用于将上述M路输入光信号分别输入至相应的光学组件上，以使得上述输入向量分别与上述内积计算单元所编码的计算矩阵进行内积运算，得到M路输出光信号。
可选地，上述光学组件包括光学衰减器及偏振片；上述计算矩阵编码单元，具体用于针对任一计算矩阵，分别将上述计算矩阵的各个元素的绝对值编码至相应光学组件的光学 衰减器上，分别将上述计算矩阵的各个元素的符号编码至相应光学组件的偏振片上。
由上可见，通过本申请实施例，提出了一种新型的光学神经网络结构，图像识别装置通过上述新型的光学神经网络结构对图像进行识别，由于其中的所有计算皆由光学元件完成，因而能耗极低，并且处理速度快，能够快速获得图像识别的结果。
实施例3
本申请实施例提供一种电子设备，请参阅图7，本申请实施例中的电子设备包括：存储器701，一个或多个处理器702(图7中仅示出一个)及存储在存储器701上并可在处理器上运行的计算机程序。其中：存储器701用于存储软件程序以及模块，处理器702通过运行存储在存储器701的软件程序以及单元，从而执行各种功能应用以及数据处理，以获取上述预设事件对应的资源。具体地，处理器702通过运行存储在存储器701的上述计算机程序时实现以下步骤：
获取待识别图像；
将上述待识别图像输入至上述光学神经网络结构；
基于上述光学神经网络结构的输出结果确定上述待识别图像的识别结果；
其中，上述光学神经网络结构由X层神经网络所构成，上述X为正整数；上述光学神经网络结构用于：
针对第i层神经网络，获取第i层神经网络的输入向量，上述i为大于0且小于X+1的正整数，其中，当i等于1时，上述第i层神经网络的输入向量基于上述待识别图像的各个像素点而生成；当i大于1时，上述第i层神经网络的输入向量为第i-1层神经网络的输出向量；
分别基于Yi个内积计算单元对上述输入向量进行线性变换，得到Yi个线性变换的结果，其中，上述Yi为正整数；
将上述Yi个线性变换的结果通过非线性晶体进行激活，得到Yi个激活结果；
将上述Yi个激活结果作为本层神经网络的输出向量，其中，当i等于X时，上述第i层神经网络的输出向量为上述光学神经网络结构的输出结果，上述第i层神经网络的输出向量中的各个元素用于指示上述待识别图像属于各个不同类别的可能性。
假设上述为第一种可能的实施方式，则在第一种可能的实施方式作为基础而提供的第二种可能的实施方式中，当i等于1时，上述获取第i层神经网络的输入向量，包括：
获取单一相干光源；
将单一相干光源等分为N路光信号；
将上述N路光信号分别通过光学衰减器进行编码，并基于编码后的N路光信号的振幅构建第1层神经网络的输入向量，其中，上述编码后的N路光信号分别与上述待识别图像的各个像素点相关联，上述N基于上述待识别图像的像素点的个数而设定。
在上述第一种可能的实施方式作为基础而提供的第三种可能的实施方式中，上述分别基于Yi个内积计算单元对上述输入向量进行线性变换，得到Yi个线性变换的结果，包括：
分别对上述输入向量所包含的各个元素进行编码，得到M路输入光信号，其中，M为上述输入向量所包含的元素的个数；
针对任一内积计算单元：
上述内积计算单元中包含有M个光学组件，将上述M路输入光信号分别输入至相应的光学组件上，得到M路输出光信号；
将上述M路输出光信号进行合束后得到内积计算结果，作为基于上述内积计算单元对上述输入向量进行线性变换所得到的线性变换的结果。
在上述第三种可能的实施方式作为基础而提供的第四种可能的实施方式中，处理器702通过运行存储在存储器701的上述计算机程序时实现以下步骤：
获取预设的转换矩阵，其中，上述预设的转换矩阵与第i层神经网络相关，上述预设的转换矩阵的维度为M*Yi；
将上述转换矩阵拆分为Yi个M*1的矩阵，记为Yi个计算矩阵；
将上述Yi个计算矩阵分别编码至相应的内积计算单元中；
相应地，上述将上述M路输入光信号分别输入至相应的光学组件上，得到M路输出光信号，包括：
将上述M路输入光信号分别输入至相应的光学组件上，以使得上述输入向量分别与上述内积计算单元所编码的计算矩阵进行内积运算，得到M路输出光信号。
在上述第四种可能的实施方式作为基础而提供的第五种可能的实施方式中，上述光学组件包括光学衰减器及偏振片；上述将上述Yi个计算矩阵分别编码至相应的内积计算单元中，包括：
针对任一计算矩阵，分别将上述计算矩阵的各个元素的绝对值编码至相应光学组件的光学衰减器上，分别将上述计算矩阵的各个元素的符号编码至相应光学组件的偏振片上。
进一步，如图7所示，上述电子设备还可包括：一个或多个输入设备703(图7中仅示 出一个)和一个或多个输出设备704(图7中仅示出一个)。存储器701、处理器702、输入设备703和输出设备704通过总线705连接。
应当理解，在本申请实施例中，所称处理器702可以是中央处理单元(Central Processing Unit，CPU)，该处理器还可以是其他通用处理器、数字信号处理器(Digital Signal Processor，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现成可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
输入设备703可以包括键盘、触控板、指纹采传感器(用于采集用户的指纹信息和指纹的方向信息)、麦克风等，输出设备704可以包括显示器、扬声器等。
存储器701可以包括只读存储器和随机存取存储器，并向处理器702提供指令和数据。存储器701的一部分或全部还可以包括非易失性随机存取存储器。例如，存储器701还可以存储设备类型的信息。
由上可见，通过本申请实施例，提出了一种新型的光学神经网络结构，电子设备通过上述新型的光学神经网络结构对图像进行识别，由于其中的所有计算皆由光学元件完成，因而能耗极低，并且处理速度快，能够快速获得图像识别的结果。
所属领域的技术人员可以清楚地了解到，为了描述的方便和简洁，仅以上述各功能单元、模块的划分进行举例说明，实际应用中，可以根据需要而将上述功能分配由不同的功能单元、模块完成，即将上述装置的内部结构划分成不同的功能单元或模块，以完成以上描述的全部或者部分功能。实施例中的各功能单元、模块可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中，上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。另外，各功能单元、模块的具体名称也只是为了便于相互区分，并不用于限制本申请的保护范围。上述系统中单元、模块的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
在上述实施例中，对各个实施例的描述都各有侧重，某个实施例中没有详述或记载的部分，可以参见其它实施例的相关描述。
本领域普通技术人员可以意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、或者计算机软件和电子硬件的结合来实现。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本 申请的范围。
在本申请所提供的实施例中，应该理解到，所揭露的装置和方法，可以通过其它的方式实现。例如，以上所描述的系统实施例仅仅是示意性的，例如，上述模块或单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通讯连接可以是通过一些接口，装置或单元的间接耦合或通讯连接，可以是电性，机械或其它的形式。
上述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
上述集成的单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请实现上述实施例方法中的全部或部分流程，也可以通过计算机程序来指令相关的硬件来完成，上述的计算机程序可存储于一计算机可读存储介质中，该计算机程序在被处理器执行时，可实现上述各个方法实施例的步骤。其中，上述计算机程序包括计算机程序代码，上述计算机程序代码可以为源代码形式、对象代码形式、可执行文件或某些中间形式等。上述计算机可读介质可以包括：能够携带上述计算机程序代码的任何实体或装置、记录介质、U盘、移动硬盘、磁碟、光盘、计算机存储器、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、电载波信号、电信信号以及软件分发介质等。需要说明的是，上述计算机可读介质包含的内容可以根据司法管辖区内立法和专利实践的要求进行适当的增减，例如在某些司法管辖区，根据立法和专利实践，计算机可读介质不包括是电载波信号和电信信号。
以上上述实施例仅用以说明本申请的技术方案，而非对其限制；尽管参照前述实施例对本申请进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围，均应包含在本申请的保护范围之内。

Claims (10)

  1. 一种基于光学神经网络结构的图像识别方法，其特征在于，所述光学神经网络结构由X层神经网络所构成，所述X为正整数；所述图像识别方法包括：
    获取待识别图像；
    将所述待识别图像输入至所述光学神经网络结构；
    基于所述光学神经网络结构的输出结果确定所述待识别图像的识别结果；
    其中，所述光学神经网络结构用于：
    针对第i层神经网络，获取第i层神经网络的输入向量，所述i为大于0且小于X+1的正整数，其中，当i等于1时，所述第i层神经网络的输入向量基于所述待识别图像的各个像素点而生成；当i大于1时，所述第i层神经网络的输入向量为第i-1层神经网络的输出向量；
    分别基于Yi个内积计算单元对所述输入向量进行线性变换，得到Yi个线性变换的结果，其中，所述Yi为正整数；
    将所述Yi个线性变换的结果通过非线性晶体进行激活，得到Yi个激活结果；
    将所述Yi个激活结果作为本层神经网络的输出向量，其中，当i等于X时，所述第i层神经网络的输出向量为所述光学神经网络结构的输出结果，所述第i层神经网络的输出向量中的各个元素用于指示所述待识别图像属于各个不同类别的可能性。
  2. 如权利要求1所述的图像识别方法，其特征在于，当i等于1时，所述获取第i层神经网络的输入向量，包括：
    获取单一相干光源；
    将单一相干光源等分为N路光信号；
    将所述N路光信号分别通过光学衰减器进行编码，并基于编码后的N路光信号的振幅构建第1层神经网络的输入向量，其中，所述编码后的N路光信号分别与所述待识别图像的各个像素点相关联，所述N基于所述待识别图像的像素点的个数而设定。
  3. 如权利要求1所述的图像识别方法，其特征在于，所述分别基于Yi个内积计算单元对所述输入向量进行线性变换，得到Yi个线性变换的结果，包括：
    分别对所述输入向量所包含的各个元素进行编码，得到M路输入光信号，其中，M为所述输入向量所包含的元素的个数；
    针对任一内积计算单元：
    所述内积计算单元中包含有M个光学组件，将所述M路输入光信号分别输入至相应的光学组件上，得到M路输出光信号；
    将所述M路输出光信号进行合束后得到内积计算结果，作为基于所述内积计算单 元对所述输入向量进行线性变换所得到的线性变换的结果。
  4. 如权利要求3所述的图像识别方法，其特征在于，所述图像识别方法还包括：
    获取预设的转换矩阵，其中，所述预设的转换矩阵与第i层神经网络相关，所述预设的转换矩阵的维度为M*Yi；
    将所述转换矩阵拆分为Yi个M*1的矩阵，记为Yi个计算矩阵；
    将所述Yi个计算矩阵分别编码至相应的内积计算单元中；
    相应地，所述将所述M路输入光信号分别输入至相应的光学组件上，得到M路输出光信号，包括：
    将所述M路输入光信号分别输入至相应的光学组件上，以使得所述输入向量分别与所述内积计算单元所编码的计算矩阵进行内积运算，得到M路输出光信号。
  5. 如权利要求4所述的图像识别方法，其特征在于，所述光学组件包括光学衰减器及偏振片；所述将所述Yi个计算矩阵分别编码至相应的内积计算单元中，包括：
    针对任一计算矩阵，分别将所述计算矩阵的各个元素的绝对值编码至相应光学组件的光学衰减器上，分别将所述计算矩阵的各个元素的符号编码至相应光学组件的偏振片上。
  6. 一种基于光学神经网络结构的图像识别装置，其特征在于，所述光学神经网络结构由X层神经网络所构成，所述X为正整数；所述图像识别装置包括：
    图像获取模块，用于获取待识别图像；
    图像输入模块，用于将所述待识别图像输入至所述光学神经网络结构；
    结果识别模块，用于基于所述光学神经网络结构的输出结果确定所述待识别图像的识别结果；
    其中，所述光学神经网络结构的各层神经网络均包括：
    向量输入单元，用于针对第i层神经网络，获取第i层神经网络的输入向量，所述i为大于0且小于X+1的正整数，其中，当i等于1时，所述第i层神经网络的输入向量基于所述待识别图像的各个像素点而生成；当i大于1时，所述第i层神经网络的输入向量为第i-1层神经网络的输出向量；
    线性变换单元，用于分别基于Yi个内积计算单元对所述输入向量进行线性变换，得到Yi个线性变换的结果，其中，所述Yi为正整数；
    激活单元，用于将所述Yi个线性变换的结果通过非线性晶体进行激活，得到Yi个激活结果；
    结果输出单元，用于将所述Yi个激活结果作为本层神经网络的输出向量，其中，当i等于X时，所述第i层神经网络的输出向量为所述光学神经网络结构的输出结果， 所述第i层神经网络的输出向量中的各个元素用于指示所述待识别图像属于各个不同类别的可能性。
  7. 如权利要求6所述的图像识别装置，其特征在于，所述光学神经网络结构的第1层神经网络的向量输入单元，包括：
    光源获取子单元，用于获取单一相干光源；
    光源等分子单元，用于将单一相干光源等分为N路光信号；
    光源编码子单元，用于将所述N路光信号分别通过光学衰减器进行编码，并基于编码后的N路光信号的振幅构建第1层神经网络的输入向量，其中，所述编码后的N路光信号分别与所述待识别图像的各个像素点相关联，所述N基于所述待识别图像的像素点的个数而设定。
  8. 如权利要求6所述的图像识别装置，其特征在于，所述光学神经网络结构的各层神经网络的线性变换单元，具体用于分别对所述输入向量所包含的各个元素进行编码，得到M路输入光信号，其中，M为所述输入向量所包含的元素的个数；
    所述线性变换单元，具体还用于针对任一内积计算单元，将所述M路输入光信号分别输入至相应的光学组件上，得到M路输出光信号，其中，所述内积计算单元中包含有M个光学组件；将所述M路输出光信号进行合束后得到内积计算结果，作为基于所述内积计算单元对所述输入向量进行线性变换所得到的线性变换的结果。
  9. 一种电子设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，其特征在于，所述处理器执行所述计算机程序时实现如权利要求1至5任一项所述方法的步骤。
  10. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，其特征在于，所述计算机程序被处理器执行时实现如权利要求1至5任一项所述方法的步骤。
PCT/CN2019/096458 2019-01-16 2019-07-18 基于光学神经网络结构的图像识别方法、装置及电子设备 WO2020147282A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910038838.2A CN109871871B (zh) 2019-01-16 2019-01-16 基于光学神经网络结构的图像识别方法、装置及电子设备
CN201910038838.2 2019-01-16

Publications (1)

Publication Number Publication Date
WO2020147282A1 true WO2020147282A1 (zh) 2020-07-23

Family

ID=66917757

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/096458 WO2020147282A1 (zh) 2019-01-16 2019-07-18 基于光学神经网络结构的图像识别方法、装置及电子设备

Country Status (2)

Country Link
CN (1) CN109871871B (zh)
WO (1) WO2020147282A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112232504A (zh) * 2020-09-11 2021-01-15 联合微电子中心有限责任公司 一种光子神经网络
CN112651299A (zh) * 2020-11-27 2021-04-13 深圳第三代半导体研究院 图片处理方法及其系统、应用以及计算机可读存储介质
CN117202430A (zh) * 2023-09-20 2023-12-08 浙江炯达能源科技有限公司 用于智慧灯杆的节能控制方法及系统
CN117891023A (zh) * 2024-03-15 2024-04-16 山东云海国创云计算装备产业创新中心有限公司 光子芯片、异构计算系统、精度调整方法及产品

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109871871B (zh) * 2019-01-16 2021-08-27 南方科技大学 基于光学神经网络结构的图像识别方法、装置及电子设备
CN110334804B (zh) * 2019-06-20 2021-09-07 清华大学 基于空间部分相干光的全光深度衍射神经网络系统及方法
CN112860597B (zh) * 2019-11-27 2023-07-21 珠海格力电器股份有限公司 一种神经网络运算的系统、方法、装置及存储介质
WO2021201773A1 (en) * 2020-04-03 2021-10-07 Nanyang Technological University Apparatus and method for implementing a complex-valued neural network
CN111860822B (zh) * 2020-07-20 2023-09-26 联合微电子中心有限责任公司 光学神经网络的全光非线性激活函数实现方法、装置
CN113033797B (zh) * 2021-05-08 2022-04-12 电子科技大学 一种基于正负分离的实数域光学神经网络的模式识别方法
CN113298246B (zh) * 2021-05-27 2023-02-28 山东云海国创云计算装备产业创新中心有限公司 数据处理方法、装置及计算机可读存储介质
CN115508958B (zh) * 2022-10-08 2024-05-24 深圳中科天鹰科技有限公司 一种基于光学神经网络的光子芯片

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP0499469A2 (en) * 1991-02-13 1992-08-19 Sumitomo Cement Co. Ltd. Artificial neural network pattern recognition system
US20040107172A1 (en) * 2001-09-25 2004-06-03 Ruibo Wang Optical pulse-coupled artificial neurons
CN102043962A (zh) * 2010-09-01 2011-05-04 北京大学 数字全息三维物体识别方法与系统
US20170351293A1 (en) * 2016-06-02 2017-12-07 Jacques Johannes Carolan Apparatus and Methods for Optical Neural Network
CN109871871A (zh) * 2019-01-16 2019-06-11 南方科技大学 基于光学神经网络结构的图像识别方法、装置及电子设备

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5132813A (en) * 1988-08-18 1992-07-21 Teledyne Industries, Inc. Neural processor with holographic optical paths and nonlinear operating means
US7555155B2 (en) * 2005-01-27 2009-06-30 Cambridge Research & Instrumentation, Inc. Classifying image features
CN101311811A (zh) * 2007-05-24 2008-11-26 电子科技大学 全光模数转换器
WO2010091740A1 (en) * 2009-02-16 2010-08-19 Telefonaktiebolaget Lm Ericsson (Publ) Optical digital-to-analog conversion
US10460231B2 (en) * 2015-12-29 2019-10-29 Samsung Electronics Co., Ltd. Method and apparatus of neural network based image signal processor
CN107769856A (zh) * 2016-08-22 2018-03-06 中兴通讯股份有限公司 一种光信号发送系统、接收系统和方法以及通信系统
CN108599865B (zh) * 2018-04-13 2020-04-28 北京邮电大学 基于光子神经网络的信号调制格式识别方法、装置

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP0499469A2 (en) * 1991-02-13 1992-08-19 Sumitomo Cement Co. Ltd. Artificial neural network pattern recognition system
US20040107172A1 (en) * 2001-09-25 2004-06-03 Ruibo Wang Optical pulse-coupled artificial neurons
CN102043962A (zh) * 2010-09-01 2011-05-04 北京大学 数字全息三维物体识别方法与系统
US20170351293A1 (en) * 2016-06-02 2017-12-07 Jacques Johannes Carolan Apparatus and Methods for Optical Neural Network
CN109871871A (zh) * 2019-01-16 2019-06-11 南方科技大学 基于光学神经网络结构的图像识别方法、装置及电子设备

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
YANG, ZIXIAN: "Research and Application on Image Recognition Technology Based on Deep Learning Convolutional Neural Network", CHINA PLANT ENGINEERING, vol. 2018, no. 12, 31 December 2018 (2018-12-31), CN, pages 146 - 149, XP009521962, ISSN: 1671-0711 *

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112232504A (zh) * 2020-09-11 2021-01-15 联合微电子中心有限责任公司 一种光子神经网络
CN112232504B (zh) * 2020-09-11 2024-06-21 联合微电子中心有限责任公司 一种光子神经网络
CN112651299A (zh) * 2020-11-27 2021-04-13 深圳第三代半导体研究院 图片处理方法及其系统、应用以及计算机可读存储介质
CN117202430A (zh) * 2023-09-20 2023-12-08 浙江炯达能源科技有限公司 用于智慧灯杆的节能控制方法及系统
CN117202430B (zh) * 2023-09-20 2024-03-19 浙江炯达能源科技有限公司 用于智慧灯杆的节能控制方法及系统
CN117891023A (zh) * 2024-03-15 2024-04-16 山东云海国创云计算装备产业创新中心有限公司 光子芯片、异构计算系统、精度调整方法及产品
CN117891023B (zh) * 2024-03-15 2024-05-31 山东云海国创云计算装备产业创新中心有限公司 光子芯片、异构计算系统、精度调整方法及产品

Also Published As

Publication number Publication date
CN109871871A (zh) 2019-06-11
CN109871871B (zh) 2021-08-27

Similar Documents

Publication Publication Date Title
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
CN109002861B (zh) 联邦建模方法、设备及存储介质
WO2021164269A1 (zh) 基于注意力机制的视差图获取方法和装置
CN112016312B (zh) 数据关系抽取方法、装置、电子设备及存储介质
WO2020186887A1 (zh) 一种连续小样本图像的目标检测方法、装置及设备
CN111079764B (zh) 一种基于深度学习的低照度车牌图像识别方法及装置
JP6731529B1 (ja) 単一画素攻撃サンプルの生成方法、装置、設備及び記憶媒体
CN111369427A (zh) 图像处理方法、装置、可读介质和电子设备
WO2022242122A1 (zh) 一种视频优化方法、装置、终端设备及存储介质
CN111696029A (zh) 虚拟形象视频生成方法、装置、计算机设备及存储介质
WO2020231005A1 (ko) 영상 처리 장치 및 그 동작방법
CN111985504B (zh) 基于人工智能的翻拍检测方法、装置、设备及介质
WO2023040361A1 (zh) 一种基于改进类提升方案的图像加密方法
CN109787760A (zh) 一种优化的基于h1类哈希函数族的密钥保密增强方法及装置
WO2023096571A2 (en) Data processing for release while protecting individual privacy
CN114330565A (zh) 一种人脸识别方法及装置
CN113033448A (zh) 一种基于多尺度卷积和注意力的遥感影像去云残差神经网络系统、方法、设备及存储介质
CN113962873A (zh) 一种图像去噪方法、存储介质及终端设备
CN110705653A (zh) 图像分类方法、图像分类装置及终端设备
CN112489144B (zh) 图像处理方法、图像处理装置、终端设备及存储介质
CN114677319A (zh) 干细胞分布确定方法、装置、电子设备及存储介质
CN112507323A (zh) 基于单向网络的模型训练方法、装置和计算设备
TW201140468A (en) Image texture extraction method, image identification method and image identification apparatus
CN111382831A (zh) 加速卷积神经网络模型前向推理的方法及装置
CN115345298A (zh) 联合训练模型的方法及装置

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: 19909843

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: 19909843

Country of ref document: EP

Kind code of ref document: A1