Image classification system and method based on asymmetric pumping

Abstract

The invention belongs to the technical field of optical signal processing, and discloses an image classification system and method based on asymmetric pumping, and the system comprises an image obtaining module which is used for obtaining a handwritten digital image and transmitting the handwritten digital image to a nonlinear XOR module; the nonlinear XOR module is used for generating a plurality of serial signals according to the received handwritten digital image; delaying the serial signal to obtain a first signal and a second signal; sending the first signal and the second signal to a BIC resonant cavity to obtain a coupled XOR signal, and sending the coupled XOR signal to a linear classification module; the first signal is delayed by a preset time length than the second signal, and the difference between the power of the first signal and the power of the second signal is equal to a preset threshold; and the linear classification module is used for classifying the handwritten digital images according to the coupled XOR signals and outputting digital classification labels. According to the invention, all-optical nonlinear transformation is realized in the optical neural network, and the data separability is remarkably improved.

Description

Technical Field

[0001] This application relates to the field of optical signal processing technology, and particularly to an image classification system and method based on asymmetric pumping. Background Art

[0002] In recent years, optical neural network technology has received much attention. Researchers have gradually developed a series of network layers constructed based on components such as optical waveguides and lasers to circumvent the speed limitations of traditional electronic neural networks. Currently, typical optical neural network architectures include silicon optical neural networks based on on-chip Mach-Zehnder interferometers, diffraction neural networks based on 3D printed phase plates, and hybrid optical neural networks incorporating phase change materials. Most of them rely on linear optical responses such as interference, diffraction, and wavelength division multiplexing.

[0003] However, existing optical neural networks fail to introduce non-linear terms, and the arithmetic relationships between network layers are limited to linear operations. Although multi-layer diffraction networks can exhibit certain depth features, the constructed models can always only handle linear input-output relationships, significantly reducing the generalization ability of diffraction neural networks and making it difficult to apply them to more complex machine learning tasks such as image classification. Summary of the Invention

[0004] This application provides an image classification system and method based on asymmetric pumping, which can achieve all-optical non-linear transformation in an optical neural network, significantly improve data separability, and make the classification accuracy of the optical neural network in image classification tasks much higher than that of traditional linear classifiers, promoting the further development of the field of optical computing.

[0005] In the first aspect, an embodiment of this application provides an image classification system based on asymmetric pumping, including: An image acquisition module, configured to acquire a handwritten digit image and send it to the non-linear exclusive OR module; A non-linear exclusive OR module, configured to generate multiple serial signals according to the received handwritten digit image; delay the serial signals to obtain a first signal and a second signal; send the first signal and the second signal to the BIC resonator to obtain a coupled exclusive OR signal and send it to the linear classification module; wherein, the first signal is delayed by a preset time duration compared to the second signal, and the difference between the power of the first signal and the power of the second signal is equal to a preset threshold; A linear classification module, configured to receive the coupled exclusive OR signals corresponding to the serial signals, classify the handwritten digit image according to the coupled exclusive OR signals, and output a digital classification label.

[0006] Further, the non-linear exclusive OR module includes a binarization unit, a signal generation unit, a delay unit, and a BIC resonator; The binarization unit is used to binarize the received handwritten digit image to obtain a digital grayscale image; The signal generation unit is used to randomly select pixel pairs in the digital grayscale image and generate corresponding serial signals according to the pixel values of the two pixel points in the pixel pair; The delay unit is used to delay the serial signal to obtain a first signal and a second signal, and send the first signal and the second signal to the coupled XOR unit; The BIC resonator is used to couple the first signal and the second signal to obtain a coupled XOR signal.

[0007] Furthermore, the BIC resonator adopts an etched-free BIC quasi-two-dimensional perovskite photonic crystal structure.

[0008] Furthermore, the non-linear XOR module further includes a frequency doubling unit; the frequency doubling unit is arranged after the signal generation unit; The frequency doubling unit is used to perform frequency doubling processing on the serial signal and send it to the delay unit.

[0009] Furthermore, the linear classification module is a fully connected linear neural network.

[0010] Furthermore, the first signal is delayed by 30 picoseconds compared with the second signal.

[0011] Furthermore, the difference between the power of the first signal and the power of the second signal is equal to 0.3Pth.

[0012] Furthermore, the frequency doubling unit is a β-phase barium metaborate crystal.

[0013] In a second aspect, an image classification method based on asymmetric pumping provided by an embodiment of the present application includes: The image acquisition module acquires a handwritten digit image and sends it to the non-linear XOR module; The non-linear XOR module generates a plurality of serial signals according to the received handwritten digit image; delays the serial signals to obtain a first signal and a second signal; sends the first signal and the second signal to the BIC resonator to obtain a coupled XOR signal and sends it to the linear classification module; wherein, the first signal is delayed by a preset time period compared with the second signal, and the difference between the power of the first signal and the power of the second signal is equal to a preset threshold; The linear classification module receives the coupled XOR signals corresponding to the serial signals, classifies the handwritten digit image according to the coupled XOR signals, and outputs a digital classification label.

[0014] Furthermore, the above-mentioned non-linear XOR module generates a plurality of serial signals according to the received handwritten digit image, including: Binarize the received handwritten digit image to obtain a digital grayscale image; Randomly select pixel pairs in a digital grayscale image; Generate corresponding serial signals according to the pixel values of the two pixel points in the pixel pair.

[0015] Furthermore, the method further includes: Before the delay, use a β-phase barium metaborate crystal to perform frequency doubling processing on the serial signal.

[0016] In summary, compared with the prior art, the beneficial effects brought by the technical solution provided by the embodiments of the present application at least include: An image classification system based on asymmetric pumping provided by an embodiment of the present application delays the serial signal generated by a handwritten digital image into a first signal and a second signal with different delay durations, and then uses the coupling effect of a BIC resonator to non-linearly map the first signal and the second signal into a high-dimensional space, realizing all-optical non-linear processing. At the same time, due to the delay and power gap between the first signal and the second signal, an all-optical exclusive OR (XOR) logic gate can be realized during mapping to obtain a coupled XOR signal, significantly improving data separability. Then, based on each coupled XOR signal for linear classification, its accuracy is much greater than that of traditional neural network classification methods, demonstrating the application potential of optical neural networks in the field of image classification. Description of the Drawings

[0017] Figure 1 It is a flowchart of an image classification system based on asymmetric pumping provided by an embodiment of the present application.

[0018] Figure 2 It is a schematic diagram of an optical neural network computing architecture provided by an embodiment of the present application.

[0019] Figure 3 It is an internal structure diagram of a delay unit provided by an embodiment of the present application.

[0020] Figure 4 It is an incident schematic diagram of a BIC resonator provided by an embodiment of the present application.

[0021] Figure 5 It is a schematic diagram of realizing classification based on exclusive OR logic gate operation provided by an embodiment of the present application.

[0022] Figure 6 It is a signal schematic diagram of XOR coupling under different pump powers provided by an embodiment of the present application. Detailed Embodiments

[0023] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments.

[0024] Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of this application.

[0025] Please refer to Figure 1 , an image classification system based on asymmetric pumping provided by an embodiment of this application includes: An image acquisition module, configured to acquire a handwritten digit image and send it to a non-linear exclusive OR module.

[0026] Among them, the handwritten digit image can be selected as the MNIST image, and this image set can be called using pytorch.

[0027] A non-linear exclusive OR module, configured to generate multiple serial signals according to the received handwritten digit image; delay the serial signals to obtain a first signal and a second signal; send the first signal and the second signal to a BIC resonant cavity to obtain a coupled exclusive OR signal and send it to a linear classification module; wherein, the first signal is delayed by a preset time duration compared to the second signal, and the difference between the power of the first signal and the power of the second signal is equal to a preset threshold.

[0028] Among them, the BIC resonant cavity can adopt an unetched BIC quasi-two-dimensional perovskite photonic crystal structure. The delay of the first signal and the second signal is performed after separating the serial signals, and the control of the power is achieved when generating the serial signals.

[0029] It can be understood that the problem to be solved in this application is to implement non-linear transformation in an optical neural network. The core of optical non-linearity is to achieve the transformation of light regulation from low dimension to high dimension, and the fundamental of light regulation lies in the interaction between light and matter. Therefore, the key to optical non-linearity is to solve the interaction between light and matter.

[0030] The BIC (Bound States in the Continuum, BIC) adopted in this application is a periodic photonic crystal structure that suppresses radiative losses through interference effects. When the structure supports the symmetry in a specific direction, the modes with the same symmetry degenerate at the Γ point (the symmetry point where the wave vector k = 0), resulting in the cancellation of the radiative fields in different propagation directions, thus forming a non-radiative bound state. This process not only suppresses the leakage of energy but also localizes the modes, achieving an extremely high quality factor Q.

[0031] Therefore, the BIC structure has two significant characteristics: a theoretically near-infinite Q factor and a strong field confinement ability, which will significantly enhance the interaction between light and matter. By using the BIC photonic crystal as a resonant cavity and combining with unetched micro-nano technology, the interaction between light and gain medium can be further improved.

[0032] The BIC resonator can analyze the dipole-dipole interaction therein through the dipole approximation. During the optical excitation process, the collective resonance of the dipoles excited by the pump light is usually confined to a small area near the excitation region. In order to establish the correlation between the collective dipoles, two pump lasers (the first signal and the second signal) are used in this application for excitation. When the two pump beams excite the sample, a coupling channel is formed between the BIC dipole resonances, realizing the dipole-dipole correlation coupling between different collective dipoles. In a non-Hermitian system, the relationship between gain and loss can be analyzed by studying PT symmetry (Parity-Time symmetry). When using two spatially separated laser beams with finite-sized spots to pump the BIC metasurface, a coupled PT-symmetric system can be formed between the discrete quasi-BIC nodes, and the coupled mode equations in the PT-symmetric system can be used for modeling to accurately describe the transient coupling process of the dipole oscillations in the BIC resonator under femtosecond pulsed laser pumping: The Hamiltonian is as follows: Where, and respectively represent two asymmetric pump lasers, represents the natural frequency, the gain term and as well as the loss term and respectively affect the amplitude evolution of each mode, while the coupling terms and reflect the interaction between the two modes.

[0033] A linear classification module is used to receive the coupled XOR signals corresponding to each serial signal, classify the handwritten digit images according to each coupled XOR signal, and output digital classification labels. Among them, the linear classification module is a fully connected linear neural network.

[0034] Specifically, a simple fully connected linear neural network model can be trained using the PyTorch framework. During the training process, the model sequentially performs forward propagation, loss calculation, backpropagation, and parameter update.

[0035] In the model structure design, the number of neurons in the input layer is equal to the total number of image pixels, and the output layer corresponds to different classification labels. A linear layer self.fc1 = nn.Linear(input_size, num_classes) is used, where input_size is the size of the input image. If the input handwritten digit image only includes 10 digits from 0 to 9, then the number of classes of the num_classes output labels is 10, and hyperparameters such as EPOCH, BATCH_SIZE, and learning rate are set.

[0036] Please refer to Figure 2 , it can be understood that both the serial signal and the coupled XOR signal are optical signals. Each of the coupled XOR signals output after being processed by the non - linear XOR module can be regarded as an image containing non - linear information after the XOR operation. Input it into the trained linear classification module to complete the classification task of digits 0 - 9 in the image.

[0037] An image classification system based on asymmetric pumping provided in the above - mentioned embodiment delays the serial signal generated from the handwritten digit image into a first signal and a second signal with two different delay durations, and then uses the coupling effect of the BIC resonator to non - linearly map the first signal and the second signal into a high - dimensional space, realizing all - optical non - linear processing. At the same time, due to the delay and power gap between the first signal and the second signal, an all - optical exclusive - OR (XOR) logic gate can be realized during the mapping to obtain the coupled XOR signal, significantly improving the data separability. Then, based on each coupled XOR signal for linear classification, its accuracy is much greater than that of the traditional neural network classification method, demonstrating the application potential of the optical neural network in the field of image classification.

[0038] In addition, currently, some individual optical non - linear studies focus on intensity - type non - linear transformations, which usually rely on devices such as semiconductor optical amplifiers and photodetectors to achieve the non - linear relationship between input and output, or achieve it through the non - linear absorption of materials. However, intensity - type non - linear transformations are often accompanied by optical power loss, which also limits the application of optical neural networks in complex learning tasks such as image classification. And this application combines the coupling dynamic behavior between quasi - BIC micro - lasers with optical signal processing, proposes a new idea of using optical loss as a control degree of freedom, solves the problem of power loss in non - linear transformation, and also provides a new technical path for the practical application of all - optical non - linear signal processing.

[0039] In some embodiments, the non - linear XOR module includes a binarization unit, a signal generation unit, a delay unit, and a BIC resonator; the binarization unit is used to binarize the received handwritten digit image to obtain a digital grayscale image.

[0040] The signal generation unit is used to randomly select pixel pairs in a digital grayscale image and generate corresponding serial signals according to the pixel values of the two pixel points in the pixel pair. Among them, since the handwritten digital image has been binarized, there are only two pixel values in the digital grayscale image, namely 0 and 1. The signals corresponding to the two pixel values are the first signal and the second signal respectively, which are combined to form a serial signal. The difference between the power of the first signal and the power of the second signal is equal to 0.3Pth, where Pth = 250 nJ / cm².

[0041] Specifically, if the pixel value is "0", the pump power of the corresponding signal is 0.8 Pth, and if the pixel value is "1", the pump power of the corresponding signal is 1.2 Pth. That is, the present application uses the pump power of the signal to represent different pixel values.

[0042] The delay unit is used to delay the serial signal to obtain the first signal and the second signal, and send the first signal and the second signal to the coupled XOR unit. Among them, the first signal is delayed by 30 picoseconds compared to the second signal.

[0043] Furthermore, the non-linear XOR module further includes a frequency doubling unit; the frequency doubling unit is arranged after the signal generation unit; the frequency doubling unit is used to perform frequency doubling processing on the serial signal and send it to the delay unit.

[0044] The frequency doubling unit can specifically be a β-phase barium metaborate crystal (β-BaB2O4), abbreviated as BBO.

[0045] Specifically, BBO is used to double the frequency of 800 nm light to 400 nm, that is, the subsequent processes all use 400 nm light.

[0046] Please refer to Figure 3 , the delay unit can be implemented by using a beam splitting prism and multiple mirrors to separate and delay the first signal and the second signal in the serial signal. After the serial signal is separated into the first signal and the second signal by the beam splitting prism, they enter different reflection paths respectively and are reflected different times to achieve different delay durations of the signals.

[0047] The BIC resonator is used to couple the first signal and the second signal to obtain a coupled XOR signal.

[0048] Specifically, please refer to Figure 4 , Figure 5 and Figure 6, first, the first signal and the second signal are respectively labeled as A and B, and the incident light spot is optimized to a Gaussian distribution with a diameter of about 5 μm. When the preset time delay of the second signal relative to the first signal is 30 picoseconds and the power difference is 0.3Pth, the formed all-optical exclusive-OR logic gate is the most reasonable and accurate. Because in the coupled-mode theory, the coupling coefficient of the double resonator depends on the relative distance and relative time between the cavities. By adjusting the relative time delay of the two pump beams, the coupling of the double-cavity system can be effectively controlled. In the image classification system of the present application, the delay unit converts the time delay into spatial resolution, adjusts the relative intensity of the pump light, and realizes the control of the system gain and asymmetric pump coupling.

[0049] Specifically, pixel pairs in the image are randomly selected, and an exclusive-OR (XOR) operation is performed based on their values to generate a coupled XOR signal; if both pixel pairs are 0, [0, 0, 0] is output; if one is 0 and the other is 1, [0, 1, 1] is output; if one is 1 and the other is 0, [1, 0, 1] is output; when both are 1, [1, 1, 0] is output, that is, the output coupled XOR signal exists in vector form. This logic gate can effectively classify in a two-dimensional plane by introducing non-linear features, thereby separating the output results of "0" and "1". Combined with the trained linear classification module, the construction of the all-optical neural network for image classification can be realized.

[0050] In some embodiments, the non-linear XOR module further includes a white light generating unit; the white light generating unit is used to generate a white light signal and add the white light signal to the first signal and the second signal before they are sent to the BIC resonator.

[0051] Adding the white light generating unit before the BIC resonator in the above embodiments enables researchers to observe the transient optical response of the BIC sample at different time points, which helps to understand the transient characteristics of the BIC sample.

[0052] Another embodiment of the present application provides an image classification method based on asymmetric pumping, including: Step S1, an image acquisition module acquires a handwritten digit image and sends it to the non-linear XOR module.

[0053] Step S2, the non-linear XOR module generates a plurality of serial signals according to the received handwritten digit image; delays the serial signals to obtain a first signal and a second signal; sends the first signal and the second signal to the BIC resonator to obtain a coupled XOR signal and sends it to the linear classification module; wherein, the first signal is delayed by a preset time duration relative to the second signal, and the difference between the power of the first signal and the power of the second signal is equal to a preset threshold.

[0054] Step S3, the linear classification module receives the coupled XOR signals corresponding to the serial signals, classifies the handwritten digit images according to the coupled XOR signals, and outputs digital classification labels.

[0055] Further, the above-mentioned non-linear XOR module generates a plurality of serial signals according to the received handwritten digit images, including: Step S21, binarize the received handwritten digit image to obtain a digital grayscale image.

[0056] Step S22, randomly select pixel pairs in the digital grayscale image.

[0057] Step S23, generate corresponding serial signals according to the pixel values of the two pixel points in the pixel pair.

[0058] Further, the method further includes: Before the delay, use a β-phase barium metaborate crystal to perform frequency doubling processing on the serial signal.

[0059] Further, the method further includes: Before sending to the BIC resonator, add a white light signal to the first signal and the second signal.

[0060] The specific limitations provided in this embodiment for an image classification method based on asymmetric pumping can be referred to the embodiment of an image classification system based on asymmetric pumping in the above text, and will not be elaborated here.

[0061] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0062] The above embodiments only represent several implementation manners of the present application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several deformations and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. An image classification system based on asymmetric pumping, characterized in that Comprising: An image acquisition module, configured to acquire a handwritten digit image and send it to a non-linear XOR module; The non-linear XOR module, configured to generate a plurality of serial signals according to the received handwritten digit image; Delay the serial signals to obtain a first signal and a second signal; send the first signal and the second signal to a BIC resonator cavity to obtain a coupled XOR signal and send it to a linear classification module; wherein, the first signal is delayed by a preset duration compared to the second signal, and the difference between the power of the first signal and the power of the second signal is equal to a preset threshold; The linear classification module, configured to receive the coupled XOR signals corresponding to the respective serial signals, classify the handwritten digit image according to the respective coupled XOR signals, and output a digital classification label.

2. The image classification system based on asymmetric pumping according to claim 1, wherein The non-linear XOR module includes a binarization unit, a signal generation unit, a delay unit, and a BIC resonator cavity; The binarization unit is configured to binarize the received handwritten digit image to obtain a digital grayscale image; The signal generation unit is configured to randomly select pixel pairs in the digital grayscale image and generate corresponding serial signals according to the pixel values of the two pixel points in the pixel pair; The delay unit is configured to delay the serial signals to obtain the first signal and the second signal, and send the first signal and the second signal to a coupled XOR unit; The BIC resonator cavity is configured to couple the first signal and the second signal to obtain a coupled XOR signal.

3. The image classification system based on asymmetric pumping according to claim 1, wherein The BIC resonator cavity adopts an unetched BIC quasi-two-dimensional perovskite photonic crystal structure.

4. The image classification system based on asymmetric pumping according to claim 2, wherein The non-linear XOR module further includes a frequency doubling unit; the frequency doubling unit is arranged after the signal generation unit; The frequency doubling unit is configured to perform frequency doubling processing on the serial signals and send them to the delay unit.

5. The image classification system based on asymmetric pumping according to claim 1, characterized in that The linear classification module is a fully connected linear neural network.

6. The image classification system based on asymmetric pumping according to claim 1, wherein The first signal is delayed by 30 picoseconds compared to the second signal.

7. The image classification system based on asymmetric pumping according to claim 1, wherein The difference between the power of the first signal and the power of the second signal is equal to 0.3Pth.

8. The image classification system based on asymmetric pumping according to claim 4, wherein The frequency doubling unit is a β-phase barium metaborate crystal.

9. An image classification method based on asymmetric pumping, characterized in that, Comprising: The image acquisition module acquires a handwritten digit image and sends it to the non-linear XOR module; The non-linear XOR module generates a plurality of serial signals according to the received handwritten digit image; Delay the serial signals to obtain a first signal and a second signal; send the first signal and the second signal to a BIC resonator cavity to obtain a coupled XOR signal and send it to a linear classification module; wherein, the first signal is delayed by a preset duration compared to the second signal, and the difference between the power of the first signal and the power of the second signal is equal to a preset threshold; The linear classification module receives the coupled XOR signals corresponding to the respective serial signals, classifies the handwritten digit image according to the respective coupled XOR signals, and outputs a digital classification label.

10. The image classification method based on asymmetric pumping according to claim 9, characterized in that, The non-linear XOR module generates a plurality of serial signals according to the received handwritten digit image, including: Binarize the received handwritten digit image to obtain a digital grayscale image; Randomly select pixel pairs in the digital grayscale image; Generate the corresponding serial signal according to the pixel values of the two pixel points in the pixel pair.

Images