CN-122021757-A - Photoelectric hybrid diffraction neural network system based on characteristic distillation and training method

CN122021757ACN 122021757 ACN122021757 ACN 122021757ACN-122021757-A

Abstract

The invention discloses a photoelectric hybrid diffraction neural network system based on characteristic distillation and a training method, and relates to the technical field of photoelectric hybrid diffraction neural networks. The system comprises an information input module, a diffraction neural network module, an optical information acquisition module and a target identification module, wherein the information input module is used for generating and loading an input light field of a target image, the diffraction neural network module is used for carrying out phase modulation coding on input light field information and comprises a phase modulation layer, the phase modulation layer loads phase modulation parameters obtained through training, the phase modulation parameters are obtained through optimization based on a joint loss function of characteristic distillation, the optical information acquisition module is used for receiving an output light field after phase modulation coding and obtaining optical intensity information, and the target identification module is used for realizing target classification by accessing the optical intensity information into an electric domain neural network. The photoelectric hybrid diffraction neural network system and the training method based on the characteristic distillation have the advantages of simple structure, low registration difficulty and capability of greatly improving the performance of the photoelectric hybrid diffraction neural network.

Inventors

CHANG WEIJIE
ZHANG ZHENGWEN
XU SHENGYAO
HUANG FENG

Assignees

福州大学

Dates

Publication Date: 20260512
Application Date: 20251209

Claims (10)

1. An opto-electronic hybrid diffractive neural network system based on feature distillation, comprising: the information input module is used for generating and loading an input light field of the target image; the diffraction neural network module is used for carrying out phase modulation coding on input light field information and comprises a phase modulation layer, wherein the phase modulation layer loads phase modulation parameters obtained through training, and the phase modulation parameters are obtained through optimization based on a joint loss function of characteristic distillation; The optical information acquisition module is used for receiving the output light field after the phase modulation and coding to acquire optical intensity information; And the target identification module is used for realizing target classification by accessing the optical intensity information into the electric domain neural network.
2. The system of claim 1, wherein the training process of the phase modulation parameters of the phase modulation layer comprises: Training an electric domain teacher network for target classification according to a training set and a testing set given by a target classification task; The output characteristics of the tail end convolution layer of the electric domain teacher network are taken as one output of the electric domain teacher network to obtain a target light field; Constructing a first loss function between the predicted light field and the target light field, constructing a second loss function between the predicted tag and the truth tag, and constructing a joint loss function based on the first loss function and the second loss function; The phase modulation parameters of the optical diffraction layer and the weight parameters of the electrical target recognition network are jointly optimized by error back propagation based on the joint loss function.
3. The system of claim 2, wherein the first loss function is a mean square error loss between the predicted light field and the target light field for calculating a characteristic distillation loss, wherein the second loss function is a cross entropy loss between the predicted label and the truth label for calculating a classification inference performance loss, and wherein the joint loss function ratio is a classification inference performance loss: characteristic distillation loss = 1:50.
4. The system of claim 3, wherein the first loss function is: Wherein, the The number of sample points is represented, Representing the target light field intensity for the ith sample point, Representing the predicted optical field strength for the i-th sample point.
5. The system of claim 3 or 4, wherein the second loss function is: Wherein N represents the number of categories, Representing the real label at the ith class, Representing the predicted probability of the model at the ith class.
6. The system of claim 1, wherein the object recognition module comprises two fully connected layers for non-linearly mapping and classifying the light intensity information.
7. The system of claim 1, wherein the information input module comprises a light source, a pinhole filter, a collimating lens and a mask or DMD carrying image information, wherein the light beam emitted by the light source is parallel to the pinhole filter and the collimating lens, and the object information is loaded on the light field through the mask or DMD carrying the image information.
8. The system of claim 1, wherein the optical information acquisition module is a photodetector or a CCD camera.
9. The photoelectric hybrid diffraction neural network training method based on characteristic distillation is characterized by comprising the following steps of: Training an electric domain teacher network for target classification according to a training set and a testing set given by a target classification task; The output characteristics of the tail end convolution layer of the electric domain teacher network are taken as one output of the electric domain teacher network to obtain a target light field; Constructing a first loss function between the predicted light field and the target light field, constructing a second loss function between the predicted tag and the truth tag, and constructing a joint loss function based on the first loss function and the second loss function; The phase modulation parameters of the optical diffraction layer and the weight parameters of the electrical target recognition network are jointly optimized by error back propagation based on the joint loss function.
10. The method of claim 9, wherein the electrical domain teacher network comprises six convolution modules, two max-pooling layers, a channel-flattening layer, and a full-connection layer, wherein one convolution module comprises one convolution layer, one normalization layer, and one nonlinear activation layer, and the electrical target recognition network of the photoelectric hybrid student network comprises two full-connection layers.

Description

Photoelectric hybrid diffraction neural network system based on characteristic distillation and training method Technical Field The invention relates to the technical field of photoelectric hybrid diffraction neural networks, in particular to a photoelectric hybrid diffraction neural network system based on characteristic distillation and a training method. Background The deep neural network is excellent in artificial intelligent tasks such as image recognition and voice processing, but the traditional electronic neural network is limited by moore's law and faces bottlenecks in terms of computational power, power consumption and integration level. The diffraction optical neural network becomes a potential alternative scheme with the advantages of light speed calculation, high parallelism, low power consumption and the like. However, the traditional all-optical diffraction neural network has low accuracy, complex structure and difficult registration. In the prior art, although the multilayer cascade diffraction network has higher theoretical precision, the experiment is difficult to realize and the error accumulation is serious. In addition, the existing photoelectric hybrid architecture can not fully exert the optical computing advantage in the target classification task, and the accuracy still has room for improvement. Therefore, a photoelectric hybrid diffraction neural network architecture with simple structure, high precision and easy realization is needed. Disclosure of Invention The invention aims to solve the technical problem of providing a photoelectric hybrid diffraction neural network system based on characteristic distillation and a training method, wherein the photoelectric hybrid architecture is constructed and only one phase modulation layer is needed, so that the architecture is simple in structure and low in registration difficulty, and meanwhile, the performance of the photoelectric hybrid diffraction neural network is greatly improved. In a first aspect, the present invention provides a photoelectric hybrid diffraction neural network system based on characteristic distillation, comprising: the information input module is used for generating and loading an input light field of the target image; the diffraction neural network module is used for carrying out phase modulation coding on input light field information and comprises a phase modulation layer, wherein the phase modulation layer loads phase modulation parameters obtained through training, and the phase modulation parameters are obtained through optimization based on a joint loss function of characteristic distillation; The optical information acquisition module is used for receiving the output light field after the phase modulation and coding to acquire optical intensity information; And the target identification module is used for realizing target classification by accessing the optical intensity information into the electric domain neural network. Further, the training process of the phase modulation parameter of the phase modulation layer includes: Training an electric domain teacher network for target classification according to a training set and a testing set given by a target classification task; The output characteristics of the tail end convolution layer of the electric domain teacher network are taken as one output of the electric domain teacher network to obtain a target light field; Constructing a first loss function between the predicted light field and the target light field, constructing a second loss function between the predicted tag and the truth tag, and constructing a joint loss function based on the first loss function and the second loss function; The phase modulation parameters of the optical diffraction layer and the weight parameters of the electrical target recognition network are jointly optimized by error back propagation based on the joint loss function. Further, the first loss function is a mean square error loss between a predicted light field and a target light field and is used for calculating characteristic distillation loss, the second loss function is a cross entropy loss between a predicted label and a truth label and is used for calculating classification inference performance loss, and the ratio of the combined loss functions is classification inference performance loss, namely the characteristic distillation loss=1:50. Further, the first loss function is: Wherein, the The number of sample points is represented,Represent the firstThe target light field intensity for the individual sample points,Represent the firstPredicted optical field strength for each sample point. Further, the second loss function is: Wherein, the The number of sample points is represented,Represent the firstThe target light field intensity for the individual sample points,Represent the firstPredicted optical field strength for each sample point. Further, the second loss function is: Wherein N represents the number of cat