Spatial diffraction deep neural network multifunctional optical calculation method and system

Abstract

The invention discloses a space diffraction deep neural network multifunctional optical calculation method and system, and is suitable for the field of optical signal processing. According to the invention, a feature construction technology is combined, a multi-phase surface structure is utilized, various numerical calculation problems are solved, and the method has the characteristics of low loss and high efficiency. According to the method, spatial features are created through regional division of an output plane, the spatial features are fused with digital information of an input light field, new interaction features are constructed to serve as training labels, a mapping relation between the digital information and operation results is established by training a spatial diffraction deep neural network, and the spatial diffraction depth neural network is obtained. And finally, the trained model parameters are synchronized to a spatial diffraction deep neural network component, and various numerical operation problems are solved.

Description

Technical Field

[0001] The present invention belongs to the field of optical signal processing, and more specifically, to a spatial diffraction deep neural network multifunctional optical computing method and system. Background Art

[0002] Spatial diffraction deep neural network represents an important intersection of optics and computing, which combines the principle of optical diffraction and neural network algorithms. Optical diffraction is a natural phenomenon in which light waves are deflected and overlapped when passing through the edge of an object or an obstacle. This phenomenon plays a key role in optical imaging and information processing, because the analysis of diffraction patterns can obtain information about the shape, size and structure of an object. Neural network is a mathematical model inspired by the biological nervous system, which realizes feature extraction and learning of input data through the connection and weight adjustment between neurons. As one of the important branches of machine learning, deep learning has achieved great success in the past few years, especially in the fields of medical image analysis, speech recognition, natural language processing and image classification. The innovation of spatial diffraction deep neural network is to combine the principle of optical diffraction with deep learning. By utilizing the information processing ability of optical diffraction and the learning ability of neural network, spatial diffraction deep neural network can process and analyze optical signals and images more quickly. However, despite its great potential, spatial diffraction deep neural network still faces some challenges in practical application. First, it has high requirements for the environment and is easily affected by external factors such as light, temperature and humidity. This may lead to unstable performance of the network, thus limiting its reliability and stability in practical applications. Secondly, although in theory the spatial diffraction deep neural network can handle complex nonlinear problems, in practical situations, it is still a challenge to design effective network structures and algorithms. However, despite these challenges, researchers are still confident in the application potential of spatial diffraction deep neural networks. Future research directions include further optimizing the network structure and algorithm, improving the robustness and stability of the network, and exploring a wider range of application scenarios. With the continuous advancement of technology and the deepening of theory, it is believed that the spatial diffraction deep neural network will play an increasingly important role in optical imaging, information processing and other fields. Summary of the invention

[0003] The present invention provides a multifunctional optical computing method and system of spatial diffraction deep neural network, which is unique in that the spatial diffraction neural network can solve multiple types of linear and nonlinear operations. This design enables the optical computing method to not only flexibly respond to different computing requirements, but also perform multiple computing operations at the speed of light. This innovation fills the research gap in the related technology in the direction of spatial diffraction deep neural network, and brings new development opportunities to the field of optical signal processing.

[0004] To achieve the above object, the present invention provides a spatial diffraction deep neural network multifunctional optical computing method, comprising the following steps:

[0005] A spatial diffraction neural network is constructed. The spatial diffraction neural network is used as a carrier of optical computing. The input data is a light field containing two digital information generated by a spatial light modulator. After being converted into a light beam, it is transmitted to the spatial diffraction neural network to generate a light spot. The imaging position of the light spot represents the operation result of the two numbers in the input light field. The spatial diffraction neural network includes n phase layers, and the phase combination of all phase layers is (θ 1 ,θ 2 ...θ n ), where θ i is the phase of the i-th phase layer, 1≤i≤n; the output plane is composed of multiple grids, each grid represents a different calculation result, the size of the spot is smaller than the grid size, and the distance between the spot and the edge in the grid is determined by the input digital information to avoid crosstalk;

[0006] The input light field and the imaging position of the light spot on the output plane are used as training data, and the spatial diffraction neural network is trained by a machine learning method to optimize the phase combination to obtain a trained spatial diffraction neural network;

[0007] The digital information to be calculated is converted into a light field and input into a trained spatial diffraction neural network. The imaging position of the output light spot is the calculation result.

[0008] Furthermore, the determination of the imaging position of the light spot includes: dividing the output plane into a plurality of grid areas, each grid area representing a different calculation result, and the calculation result of two numbers in the input light field determines the grid into which the light spot should fall; establishing a local coordinate system in the grid, and further determining the specific position of the light spot in the grid based on the two numbers in the input light field as horizontal and vertical coordinate information.

[0009] Preferably, machine learning is used to train the spatial diffraction neural network, and a loss function is used to measure the difference between the real target and the network output result, and based on the difference, the gradient of the network parameters is obtained by the gradient descent method, and then the parameters are updated by the back propagation method, wherein the loss function can select the mean square error function (MSE):

[0010]

[0011] Among them, y pre is the network output result, and y is the target value.

[0012] The present invention also provides a spatial diffraction deep neural network multifunctional optical computing system, including a signal input module, a spatial diffraction deep neural network, and a signal output module;

[0013] The signal input module is used to provide a light field containing two digital information;

[0014] The spatial diffraction deep neural network is used as a carrier of optical calculation, so that the light field generates a light spot after being transmitted through the spatial diffraction neural network; the imaging position of the light spot represents the operation result of two numbers in the input light field; the spatial diffraction neural network includes n phase layers, and the phase combination of all phase layers is (θ 1 ,θ 2 ...θ n ), where θ i is the phase of the i-th phase layer, 1≤i≤n; the output plane is composed of multiple grids, each grid represents a different calculation result, the size of the spot is smaller than the grid size, and the distance between the spot and the edge in the grid is determined by the input digital information;

[0015] The signal output module is used to display the light spot.

[0016] Preferably, the phase layers are arranged according to a preset interlayer spacing, and the phase combination (θ 1 ,θ 2 ...θ n ) is trained using a machine learning method. The layer spacing does not participate in the optimization during the training process. Instead, the spacing value with the best effect is finally selected as a fixed parameter by comparing the experimental results under different spacing conditions.

[0017] Preferably, each phase layer includes m×m pixels, and each pixel is used to achieve phase modulation in the range from 0 to 2π.

[0018] Preferably, the processing material system of each phase layer includes any one of gallium nitride, aluminum oxide, metal, silicon, and photosensitive polymer.

[0019] Preferably, the medium between each phase layer is vacuum, gas, liquid or solid.

[0020] Preferably, the signal output module comprises an output plane, each area in the output plane represents a different calculation result, and the area division method can be dynamically adjusted according to actual task requirements.

[0021] Through the above technical scheme conceived by the present invention, compared with the prior art, the present invention proposes a spatial diffraction deep neural network multifunctional optical computing method and system, and successfully realizes the application of spatial diffraction neural network in various numerical calculation problems by constructing spatial diffraction deep neural network, feature construction technology and machine learning optimization and other technical means, with low loss, high efficiency and scalability and other significant technical effects. On the basis of building a spatial diffraction deep neural network, the feature construction technology is used to expand its function, and multiple types of computing tasks can be flexibly performed. This is essentially different from the previous invention. On the other hand, the spatial diffraction deep neural network is widely used in the fields of optical imaging, optical communication, information security, pattern recognition, optical sensing and detection, etc. The spatial diffraction deep neural network multifunctional optical computing method and system proposed by the present invention can realize the solution of multiple types of operations, which provides a good research basis for the application of spatial diffraction deep neural network. Therefore, the spatial diffraction deep neural network multifunctional optical computing method and system provided by the present invention will play a very important application value in various fields in the future. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a schematic diagram of a spatial diffraction deep neural network multifunctional optical computing method provided by the present invention.

[0023] Figure 2 This is a physical experiment schematic diagram of a spatial diffraction deep neural network multifunctional optical computing system provided by the present invention.

[0024] Figure 3 It is a simulation flow chart of an all-optical diffraction deep neural network module in a spatial diffraction deep neural network multifunctional optical computing method provided by an embodiment of the present invention.

[0025] Figure 4(a) is a schematic diagram of an output plane in a spatial diffraction deep neural network multifunctional optical computing method provided by the present invention.

[0026] FIG4( b ) is a feature engineering design diagram provided by an embodiment of the present invention.

[0027] FIG4( c ) is a schematic diagram of output results provided by an embodiment of the present invention.

[0028] Figure 4(d) is a schematic diagram of a spatial diffraction deep neural network multifunctional optical computing method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0030] To achieve the above object, a spatial diffraction deep neural network multifunctional optical computing method provided by the present invention includes the following steps:

[0031] A spatial diffraction neural network is constructed, and the spatial diffraction neural network is used as a carrier of optical computing. The input data is a light field containing two digital information generated by a spatial light modulator, which is transmitted to the spatial diffraction neural network to generate a light spot. The imaging position of the light spot represents the operation result of the two numbers in the input light field. The spatial diffraction neural network includes n phase layers, and the phase combination of all phase layers is (θ 1 ,θ 2 ...θ n ), where θ i is the phase of the i-th phase layer, 1≤i≤n; the output plane is composed of multiple grids, each grid represents a different calculation result, the size of the spot is smaller than the grid size, and the distance between the spot and the edge in the grid is determined by the input digital information to avoid crosstalk;

[0032] The input light field and the imaging position of the light spot on the output plane are used as training data, and the spatial diffraction neural network is trained by a machine learning method to optimize the phase combination to obtain a trained spatial diffraction neural network;

[0033] The digital light field to be calculated is converted into a light beam and input into the trained spatial diffraction neural network. The imaging position of the output light spot is the calculation result.

[0034] Specifically, the determination of the imaging position of the light spot includes: dividing the output plane into multiple grid areas, each grid area represents a different calculation result, and the calculation results of two numbers in the input light field determine the grid into which the light spot should fall; establishing a local coordinate system in the grid, and further determining the specific position of the light spot in the grid based on the two numbers in the input light field as horizontal and vertical coordinate information.

[0035] Specifically, machine learning is used to train the spatial diffraction neural network, and the difference between the real target and the network output result is measured by the loss function. Based on the difference, the gradient of the network parameters is obtained by the gradient descent method, and then the parameters are updated by the back propagation method. The loss function can be the mean square error function (MSE):

[0036]

[0037] Among them, y pre is the network output result, and y is the target value.

[0038] The present invention also provides a spatial diffraction deep neural network multifunctional optical computing system, including a signal input module, a spatial diffraction deep neural network, and a signal output module;

[0039] The signal input module is used to provide a light field containing two digital information;

[0040] The spatial diffraction deep neural network is used as a carrier of optical calculation, so that the light field generates a light spot after being transmitted through the spatial diffraction neural network; the imaging position of the light spot represents the operation result of two numbers in the input light field; the spatial diffraction neural network includes n phase layers, and the phase combination of all phase layers is (θ 1 ,θ 2 ...θ n ), where θ i is the phase of the i-th phase layer, 1≤i≤n; the output plane is composed of multiple grids, each grid represents a different calculation result, the size of the spot is smaller than the grid size, and the distance between the spot and the edge in the grid is determined by the input digital information;

[0041] The signal output module is used to display the light spot.

[0042] Specifically, the phase layers are arranged according to a preset layer spacing, and the phase combination (θ 1 ,θ 2 ...θ n ) is trained using a machine learning method. The layer spacing does not participate in the optimization during the training process. Instead, the spacing value with the best effect is finally selected as a fixed parameter by comparing the experimental results under different spacing conditions.

[0043] Specifically, each phase layer includes m×m pixels, and each pixel is used to achieve phase modulation in the range from 0 to 2π.

[0044] Specifically, the processing material system of each phase layer includes any one of gallium nitride, aluminum oxide, metal, silicon, and photosensitive polymer.

[0045] Specifically, the medium between each phase layer is vacuum, gas, liquid or solid.

[0046] Specifically, the signal output module includes an output plane, each area in the output plane represents a different calculation result, and the area division method can be dynamically adjusted according to actual task requirements.

[0047] The following is a description with reference to specific embodiments and drawings.

[0048] like Figure 1 As shown, the present invention provides a schematic diagram of a spatial diffraction deep neural network multifunctional optical computing method, and the specific implementation method is as follows:

[0049] This embodiment uses a spatial diffraction deep neural network to extract signal features, which is composed of multiple phase layers. During the simulation process of the model, the output data is feature constructed. The specific construction method is: regional division is performed on the output plane to construct spatial features for characterizing the calculation results. Then, the spatial features are combined with the digital information on the input light field to construct new interactive features. Specifically, a light spot can be generated in the region based on the numbers on the input light field as horizontal and vertical coordinate information, and the light spot map is used as a training label for a specific calculation task. After that, a mapping relationship between the digital information of the input light field and the training label is established through a spatial diffraction deep neural network. Finally, the trained model parameters are synchronized to the components of the network, so that the optical computing system finally prepared can complete multiple types of linear and nonlinear operations relying only on linear models.

[0050] The following is a specific embodiment of the spatial diffraction deep neural network multifunctional optical computing system provided by the present invention, to illustrate the Figure 2 The structural diagram shown in the figure is as follows:

[0051] The device includes: a laser 1, a collimator 2, a spatial light field control device (SLM) 3, a first lens 4, a first objective lens 5, a first optical phase modulation layer 6, a second optical phase modulation layer 7, a third optical phase modulation layer 8, a fourth optical phase modulation layer 9, a second objective lens 10, a second lens 11, a CMOS camera 12, and a computer 13. The Gaussian light output by the laser 1 is transmitted through the collimator 2 to ensure its stable transmission and precise alignment, and then transmitted to the SLM3, which loads the set signal information to modulate the light field, and then through the 4f system composed of the first lens 4 and the first objective lens 5, the signal light carrying the pre-modulation information is reduced and imaged into the all-optical diffraction neural network composed of the first optical phase modulation layer 6, the second optical phase modulation layer 7, the third optical phase modulation layer 8, and the fourth optical phase modulation layer 9, and then expanded and collimated through the second objective lens 10 and the second lens 11 to be transmitted to the CMOS camera 12 for collection, and the collection results are uploaded to the computer for storage.

[0052] The following is a specific embodiment of the spatial diffraction deep neural network multifunctional optical computing method provided by the present invention, to illustrate the establishment process of the all-optical deep diffraction neural network module. The various parameters of the module in the embodiment of the present invention are obtained by establishing a simulation model and optimizing it using a deep learning method. Specifically, Figure 3A simulation flow chart of a multifunctional optical computing method according to an embodiment of the present invention is provided. The specific process is as follows:

[0053] A numerical simulation model of the multifunctional optical computing method of spatial diffraction deep neural network is established. During the simulation process, the output plane is divided into regions to construct spatial features for characterizing the calculation results. Then, the spatial features are combined with the digital information on the input light field to construct new interactive features, which are used as training labels for specific calculation tasks. After that, a spatial diffraction deep neural network model is established to construct the mapping relationship between input and label. The model uses the angular spectrum method to numerically simulate the propagation and diffraction of light between phase modulation layers. Based on the principle of fast Fourier transform (FFT), the diffraction of light waves during propagation and the scene after diffraction are simulated. Then, the training parameters of the all-optical deep diffraction neural network are optimized using the gradient descent method and the back propagation method.

[0054] Specifically, assuming that the input light field is U(x, y, z), where x and y are spatial coordinates on the plane, z = 0 represents the input surface, the propagation of the light field U at the far field z is as follows:

[0055] U N (x,y,0)=U N-1 (x,y,0)e i*2π*tanh(phi)

[0056]

[0057] U N (x,y,z)=F -1 {F{U N (x,y,0)}*H(f x ,f y ,z,λ)}

[0058]

[0059] Where F represents the Fourier transform operator, is the wave number, λ is the wavelength, H(f x ,f y ,z,λ) is the transfer function of free space, where f x and f y denote the spatial frequencies along the x and y directions respectively.

[0060] In terms of training strategy, in order to make the output result and the training label as close as possible, the mean square error can be selected as the loss function, and the calculation formula can be defined as:

[0061]

[0062] Where (x, y) represents the coordinates, V represents the x *N y The output plane area defined in pixels.

[0063] The following uses addition as an example of the calculation method to introduce in detail an embodiment of the spatial diffraction deep neural network multifunctional optical calculation method provided by the present invention.

[0064] Taking the plane area of ​​125*125 as an example, it is divided into 25 squares as the output area, and each square is 25*25 pixels in size. The output plane diagram is shown in Figure 4(a). Taking x+y=z as an example, if the size of the output light spot is n*n, then the feature construction of the output label of this method should satisfy the coordinates of the upper left corner of the target light spot (x 1 ,y 1 ) should satisfy:

[0065] (x 1 ,y 1 )=n*(x,y)+m

[0066] m is an offset, which is used to ensure that no light spot falls on the edge of the output area, thereby affecting the final decision effect. The feature construction schematic diagram is shown in Figure 4(b). Figure 4(c) is a schematic diagram of the output result provided by an embodiment of the present invention.

[0067] A specific embodiment of the spatial diffraction deep neural network multifunctional optical computing method provided by the present invention is shown in FIG4(d). The specific implementation method is as follows:

[0068] Firstly, a numerical simulation model of the spatial diffraction deep neural network is established based on the angular spectrum method. According to the calculation method and input data information, the calculation results are characterized and used as training labels. Then, the simulation model is used for training, and the trained model parameters are synchronized to the physical layer to complete the construction of the optical computing system, so that the device can realize special calculation methods based on the spatial diffraction deep neural network.

[0069] The present invention is not limited to the above-mentioned specific implementation modes. A person skilled in the art can implement the present invention in various other specific implementation modes based on the contents disclosed in the present invention. Therefore, any design that adopts the design structure and concept of the present invention and makes some simple changes or modifications shall fall within the scope of protection of the present invention.

Claims

1. A spatial diffraction deep neural network multifunctional optical computing method, characterized in that: The method comprises the following steps: A spatial diffraction neural network is constructed. The spatial diffraction neural network is used as a carrier of optical computing. The input data is a light field including two digital information, which is transmitted to the spatial diffraction neural network to generate a light spot. The imaging position of the light spot represents the operation result of the two numbers in the input light field. The spatial diffraction neural network includes n phase layers, and the phase combination of all phase layers is (θ1, θ2...θ n ), where θ i is the phase of the i-th phase layer, 1≤i≤n; the output plane is composed of multiple grids, each grid represents a different calculation result, the size of the spot is smaller than the grid size, and the distance between the spot and the edge in the grid is determined by the input digital information; The input light field and the imaging position of the light spot on the output plane are used as training data, and the spatial diffraction neural network is trained by a machine learning method to optimize the phase combination to obtain a trained spatial diffraction neural network; The digital information to be calculated is converted into a light field and input into a trained spatial diffraction neural network. The imaging position of the output light spot is the calculation result.

2. The spatial diffraction deep neural network multifunctional optical computing method according to claim 1 is characterized in that: The determination of the imaging position of the light spot includes: dividing the output plane into a plurality of grid areas, each grid area represents a different calculation result, and the calculation result of two numbers in the input light field determines the grid into which the light spot should fall; establishing a local coordinate system in the grid, and determining the specific position of the light spot in the grid based on the two numbers in the input light field as horizontal and vertical coordinate information.

3. The spatial diffraction deep neural network multifunctional optical computing method according to claim 1 is characterized in that: The spatial diffraction neural network is trained by adopting a machine learning method, and the difference between the real target and the network output result is measured by a loss function, and the gradient of the network parameters is obtained by a gradient descent method based on the difference, and then the parameters are updated by a back propagation method, wherein the loss function is a mean square error function (MSE): Among them, y pre is the network output result, and y is the target value.

4. A spatial diffraction deep neural network multifunctional optical computing system, characterized in that: It includes a signal input module, a spatial diffraction deep neural network, and a signal output module; The signal input module is used to provide a light field containing two digital information; The spatial diffraction deep neural network is used as a carrier of light calculation, so that the light field generates a light spot after being transmitted through the spatial diffraction neural network; The imaging position of the light spot represents the operation result of two numbers in the input light field; the spatial diffraction neural network includes n phase layers, and the phase combination of all phase layers is (θ1, θ2...θ n ), where θ i is the phase of the i-th phase layer, 1≤i≤n; the output plane is composed of multiple grids, each grid represents a different calculation result, the size of the spot is smaller than the grid size, and the distance between the spot and the edge in the grid is determined by the input digital information; The signal output module is used to display the light spot.

5. The spatial diffraction deep neural network multifunctional optical computing system according to claim 4, characterized in that: The phase layers are arranged according to a preset layer spacing, and the phase combinations (θ1, θ2...θ n ) is trained using a machine learning method, and the layer spacing does not participate in the optimization during the training process.

6. The spatial diffraction deep neural network multifunctional optical computing system according to claim 4, characterized in that: Each phase layer includes m×m pixels, and each pixel is used to achieve phase modulation in the range from 0 to 2π.

7. The spatial diffraction deep neural network multifunctional optical computing system according to claim 4, characterized in that: The processing material system of each phase layer includes any one of gallium nitride, aluminum oxide, metal, silicon, and photosensitive polymer.

8. The spatial diffraction deep neural network multifunctional optical computing system according to claim 4, characterized in that: The medium between each phase layer is vacuum, gas, liquid or solid.

9. The spatial diffraction deep neural network multifunctional optical computing system according to claim 4, characterized in that: The signal output module comprises an output plane, and each area in the output plane represents a different calculation result.

10. The spatial diffraction deep neural network multifunctional optical computing system according to claim 4, characterized in that: The input light field enters the spatial diffraction neural network after passing through a propagation medium, wherein the propagation medium includes air or a complex medium, and the complex medium includes a scattering medium, an absorbing medium or a nonlinear medium.

Images