Method and system for reconstructing optical phase penetrating through scattering medium
By constructing the transfer matrix of the scattering system and a deep empirical neural network, and training the network using a single image, the problem of data dependency in optical fiber communication was solved, and efficient optical phase reconstruction was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-24
- Publication Date
- 2026-03-17
AI Technical Summary
Existing technologies in fiber optic communication rely on large datasets for supervised learning, which leads to time-consuming data collection and wasted computing resources, and is heavily dependent on the diversity and representativeness of training data.
A scattering system is constructed and a transfer matrix is obtained. A deep empirical neural network is built based on the transfer matrix. The deep empirical neural network is trained using a single original image and a speckle image. The network parameters are optimized using the total loss function to achieve the reconstruction of the original image.
It achieves high-quality and fast training speed optical phase reconstruction without the need for a large dataset, thus improving the efficiency and effectiveness of optical phase reconstruction.
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention relates to the field of image reconstruction technology, and in particular to an optical phase reconstruction method and system that uses a scattering medium. Background Technology
[0002] Supervised learning is a machine learning method that trains and optimizes multi-layer neural network models to extract high-level abstract features and patterns from large-scale data. Supervised learning has achieved significant results in fields such as computer vision, natural language processing, and speech recognition, and has also brought new opportunities to the field of fiber optic communication.
[0003] Fiber optic communication is a high-bandwidth, low-loss information transmission method widely used in long-distance communication and data center networks. However, with the continuous growth of data demand, the process of collecting supervised learning data is time-consuming and wastes computing resources, heavily relying on the diversity and representativeness of training data. Summary of the Invention
[0004] To overcome the above-mentioned reliance on large datasets, this invention provides an optical phase reconstruction method and system that uses a scattering medium.
[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: This invention provides an optical phase reconstruction method through a scattering medium, comprising: Construct a scattering system and obtain the transfer matrix of the scattering system, and construct a deep empirical neural network based on the transfer matrix; A single raw image is acquired and input into the scattering system to obtain the corresponding raw speckle image; The original speckle image is input into the deep empirical neural network to obtain a preliminary reconstructed image of the original image; The preliminary reconstructed image of the original image is input into the transfer matrix to obtain the predicted speckle image; A total loss function is constructed based on the original speckle image and the predicted speckle image, and the deep empirical neural network is trained to obtain a trained deep empirical neural network. The original speckle image is input into a trained deep empirical neural network to obtain the final reconstructed image of the original image.
[0006] Preferably, the scattering system includes a laser, a polarizer, a laser beam expander, a first single-aperture aperture, a phase-type spatial light modulator, a filtering system, a first objective lens, a multimode fiber, a second objective lens, and a detector arranged sequentially. The filtering system includes a first lens, a second single-aperture aperture, and a second lens arranged sequentially.
[0007] Preferably, acquiring a single original image and inputting the original image into the scattering system to obtain the corresponding original speckle image includes: The laser receives the original image and emits a beam of light to the polarizer; The polarizer converts the light beam into horizontally polarized light; The laser beam expander expands the horizontally polarized light into a plane wave with uniform intensity; The first single-aperture aperture confines the uniformly intense plane wave as a single-beam plane wave; A phase-type spatial light modulator receives and modulates the single-beam plane wave to obtain a modulated plane wave; The filtering system filters out the 0th order diffracted light from the modulated plane wave to obtain the filtered plane wave; The first objective lens couples the filtered plane wave to obtain a coupled plane wave; The multimode fiber receives the coupled plane wave and emits speckle. The second objective lens receives the speckle, collects and magnifies it to obtain magnified speckle; The detector generates the original speckle image based on the magnified speckle.
[0008] Preferably, constructing a deep empirical neural network based on the transfer matrix includes: Establish the mapping relationship of the fully connected layer based on the mapping relationship in the transmission matrix; The deep empirical neural network includes a first encoding block, a second encoding block, a third encoding block, a fourth encoding block, a fifth encoding block, a sixth encoding block, a first connection point, a first decoding block, a second connection point, a second decoding block, a third connection point, a third decoding block, a fourth connection point, a fourth decoding block, a fifth connection point, a fifth decoding block, a feature connection point, and the fully connected layer. The first coding block, the second coding block, the third coding block, the fourth coding block, the fifth coding block, the sixth coding block, the first connection point, the first decoding block, the second connection point, the second decoding block, the third connection point, the third decoding block, the fourth connection point, the fourth decoding block, the fifth connection point, the fifth decoding block, and the feature connection point are connected in sequence; The output of the first coded block is also connected to the input of the fifth connection point; The output of the second coded block is also connected to the input of the fourth connection point; The output of the third coding block is also connected to the input of the third connection point; The output of the fourth coding block is also connected to the input of the second connection point; The output of the fifth coding block is also connected to the input of the first connection point; The outputs of the first, second, third, and fourth decoding blocks are connected to the input of the feature connection point.
[0009] Preferably, the first coding block, the second coding block, the third coding block, the fourth coding block, the fifth coding block, and the sixth coding block each include multiple downsampling modules; The first decoding block, the second decoding block, the third decoding block, the fourth decoding block, and the fifth decoding block each include multiple upsampling modules.
[0010] Preferably, a total loss function is constructed based on the original speckle image and the predicted speckle image, and the deep empirical neural network is trained to obtain a trained deep empirical neural network, including: The weight parameters of the deep empirical neural network are initialized using a Gaussian distribution; The original speckle image and the predicted speckle image are input into the total loss function to obtain the loss function value; Based on the loss function value, backpropagation updates the weight parameters of the deep empirical neural network, calculates the performance evaluation function parameters between the original speckle image and the predicted speckle image, and stops backpropagation when the performance evaluation function parameters no longer improve, thus obtaining the trained deep empirical neural network.
[0011] Preferably, the backpropagation updates the weight parameters of the deep empirical neural network, including: The formula for calculating the backpropagation is as follows:
[0012] in, These are the weight parameters of the updated deep empirical neural network. These are the network weight parameters of a deep empirical neural network. It's the learning rate. It is the gradient operation of the weight parameters of a deep empirical neural network with respect to the loss function value.
[0013] Preferably, the performance evaluation function parameters include: structural similarity, Pearson correlation coefficient, peak signal-to-noise ratio, and mutual information.
[0014] Preferably, the expression for the total loss function is as follows:
[0015] in, To predict speckle images, This is the original speckle image.
[0016] The present invention also provides an optical phase reconstruction system through a scattering medium, comprising: An initialization module is used to construct a scattering system, obtain the transfer matrix of the scattering system, and construct a deep empirical neural network based on the transfer matrix; The original speckle image acquisition module is used to acquire a single original image, input the original image into the scattering system, and obtain the corresponding original speckle image. The original image preliminary reconstruction module is used to input the original speckle image into the deep empirical neural network to obtain a preliminary reconstructed image of the original image; The predicted speckle image acquisition module is used to input the preliminary reconstructed image of the original image into the transfer matrix to obtain the predicted speckle image; The deep empirical neural network training module is used to construct a total loss function based on the original speckle image and the predicted speckle image, and to train the deep empirical neural network to obtain a trained deep empirical neural network. The original image final reconstruction module is used to input the original speckle image into a trained deep empirical neural network to obtain the final reconstructed image of the original image.
[0017] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: This invention constructs a deep empirical neural network based on the transfer matrix of a scattering system. The deep empirical neural network can be trained with only a single speckle image obtained through the scattering system and outputs the final reconstructed image of the original image. It does not rely on a large dataset, thus achieving higher reconstruction quality and faster training speed. Attached Figure Description
[0018] Figure 1 This is a schematic flowchart of an optical phase reconstruction method through a scattering medium in Example 1; Figure 2 This is a schematic diagram of the scattering system in Example 2; Figure 3 This is a schematic diagram of the training process of the deep empirical neural network in Example 2; Figure 4 This is a schematic diagram of the structure of an optical phase reconstruction system that transmits through a scattering medium in Example 3.
[0019] Explanation of reference numerals in the attached figures 1: Laser; 2: Polarizing filter; 3: Laser beam expander; 4: First single-aperture aperture; 5: Phase-type spatial light modulator; 6: Filtering system; 7: First objective lens; 8: Multimode fiber; 9: Second objective lens; 10: Detector; 61: First lens; 62: Second single-aperture stop; 63: Second lens. Detailed Implementation
[0020] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions; It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.
[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0022] Example 1 This embodiment provides an optical phase reconstruction method through a scattering medium, such as... Figure 1 As shown, it includes: Construct a scattering system and obtain the transfer matrix of the scattering system, and construct a deep empirical neural network based on the transfer matrix; A single raw image is acquired and input into the scattering system to obtain the corresponding raw speckle image; The original speckle image is input into the deep empirical neural network to obtain a preliminary reconstructed image of the original image; The preliminary reconstructed image of the original image is input into the transfer matrix to obtain the predicted speckle image; A total loss function is constructed based on the original speckle image and the predicted speckle image, and the deep empirical neural network is trained to obtain a trained deep empirical neural network. The original speckle image is input into a trained deep empirical neural network to obtain the final reconstructed image of the original image.
[0023] In its specific implementation, this invention first constructs a scattering system and obtains the transfer matrix of the scattering system, then constructs a deep empirical neural network based on the transfer matrix. Next, it acquires a single original image and inputs it into the scattering system to obtain a corresponding original speckle image. Then, it inputs the original speckle image into the deep empirical neural network to obtain a preliminary reconstructed image of the original image. Next, it inputs the preliminary reconstructed image of the original image into the transfer matrix to obtain a predicted speckle image. Based on the original speckle image and the predicted speckle image, it constructs a total loss function and trains the deep empirical neural network to obtain a trained deep empirical neural network. Finally, it inputs the original speckle image into the trained deep empirical neural network to obtain the final reconstructed image of the original image.
[0024] Example 2 This embodiment provides an optical phase reconstruction method through a scattering medium, including: Construct a scattering system and obtain the transfer matrix of the scattering system, and construct a deep empirical neural network based on the transfer matrix; A single raw image is acquired and input into the scattering system to obtain the corresponding raw speckle image; The original speckle image is input into the deep empirical neural network to obtain a preliminary reconstructed image of the original image; The preliminary reconstructed image of the original image is input into the transfer matrix to obtain the predicted speckle image; A total loss function is constructed based on the original speckle image and the predicted speckle image, and the deep empirical neural network is trained to obtain a trained deep empirical neural network. The original speckle image is input into a trained deep empirical neural network to obtain the final reconstructed image of the original image.
[0025] It should be noted that, in this embodiment, the scattering system is as follows: Figure 2 As shown, it includes a laser, a polarizer, a laser beam expander, a first single-aperture aperture, a phase-type spatial light modulator, a filtering system, a first objective lens, a multimode fiber, a second objective lens, and a detector arranged in sequence. The filtering system includes a first lens, a second single-aperture aperture, and a second lens arranged sequentially.
[0026] It should be noted that, in this embodiment, acquiring a single original image and inputting the original image into the scattering system to obtain the corresponding original speckle image includes: The laser receives the original image and emits a beam of light to the polarizer; The polarizer converts the light beam into horizontally polarized light; The laser beam expander expands the horizontally polarized light into a plane wave with uniform intensity; The first single-aperture aperture confines the uniformly intense plane wave as a single-beam plane wave; A phase-type spatial light modulator receives and modulates the single-beam plane wave to obtain a modulated plane wave; The filtering system filters out the 0th order diffracted light from the modulated plane wave to obtain the filtered plane wave; The first objective lens couples the filtered plane wave to obtain a coupled plane wave; The multimode fiber receives the coupled plane wave and emits speckle. The second objective lens receives the speckle, collects and magnifies it to obtain magnified speckle; The detector generates the original speckle image based on the magnified speckle.
[0027] It should be noted that, in this embodiment, constructing a deep empirical neural network based on the transfer matrix includes: The empirical evidence is a combination of a fixed empirical matrix (a transfer matrix constructed based on actual observation data, experimental results, or real-world scene samples). Establish the mapping relationship of the fully connected layer based on the mapping relationship in the transmission matrix; The deep empirical neural network includes a first encoding block, a second encoding block, a third encoding block, a fourth encoding block, a fifth encoding block, a sixth encoding block, a first connection point, a first decoding block, a second connection point, a second decoding block, a third connection point, a third decoding block, a fourth connection point, a fourth decoding block, a fifth connection point, a fifth decoding block, a feature connection point, and the fully connected layer. The first coding block, the second coding block, the third coding block, the fourth coding block, the fifth coding block, the sixth coding block, the first connection point, the first decoding block, the second connection point, the second decoding block, the third connection point, the third decoding block, the fourth connection point, the fourth decoding block, the fifth connection point, the fifth decoding block, and the feature connection point are connected in sequence; The output of the first coded block is also connected to the input of the fifth connection point; The output of the second coded block is also connected to the input of the fourth connection point; The output of the third coding block is also connected to the input of the third connection point; The output of the fourth coding block is also connected to the input of the second connection point; The output of the fifth coding block is also connected to the input of the first connection point; The outputs of the first, second, third, and fourth decoding blocks are connected to the input of the feature connection point.
[0028] It should be noted that, in this embodiment, the first coding block, the second coding block, the third coding block, the fourth coding block, the fifth coding block, and the sixth coding block each include multiple downsampling modules; The first decoding block, the second decoding block, the third decoding block, the fourth decoding block, and the fifth decoding block each include multiple upsampling modules.
[0029] It should be noted that, in this embodiment, a total loss function is constructed based on the original speckle image and the predicted speckle image, and the deep empirical neural network is trained accordingly. The training process is as follows: Figure 3 As shown, firstly, the original image is acquired and input into a scattering system to obtain an original speckle image. This original speckle image is then input into a deep empirical neural network to obtain a reconstructed image of the original image. The reconstructed image is then input into a transfer matrix to obtain a predicted speckle image. Based on the predicted speckle image and the original speckle image, a loss function is constructed to train the deep empirical neural network, resulting in a trained deep empirical neural network, including: The weight parameters of the deep empirical neural network are initialized using a Gaussian distribution; The original speckle image and the predicted speckle image are input into the total loss function to obtain the loss function value; Based on the loss function value, backpropagation updates the weight parameters of the deep empirical neural network, calculates the performance evaluation function parameters between the original speckle image and the predicted speckle image, and stops backpropagation when the performance evaluation function parameters no longer improve, thus obtaining the trained deep empirical neural network.
[0030] It should be noted that, in this embodiment, the backpropagation updates the weight parameters of the deep empirical neural network, including: The formula for calculating the backpropagation is as follows:
[0031] in, These are the weight parameters of the updated deep empirical neural network. These are the network weight parameters of a deep empirical neural network. It's the learning rate. It is the gradient operation of the weight parameters of a deep empirical neural network with respect to the loss function value.
[0032] It should be noted that, in this embodiment, the performance evaluation function parameters include: structural similarity, Pearson correlation coefficient, peak signal-to-noise ratio, and mutual information.
[0033] It should be noted that, in this embodiment, the expression for the total loss function is as follows:
[0034] in, To predict speckle images, This is the original speckle image.
[0035] Example 3 This embodiment provides an optical phase reconstruction system that transmits through a scattering medium, used to implement the method described in Embodiment 1 or 2, such as... Figure 4 As shown, it includes: An initialization module is used to construct a scattering system, obtain the transfer matrix of the scattering system, and construct a deep empirical neural network based on the transfer matrix; The original speckle image acquisition module is used to acquire a single original image, input the original image into the scattering system, and obtain the corresponding original speckle image. The original image preliminary reconstruction module is used to input the original speckle image into the deep empirical neural network to obtain a preliminary reconstructed image of the original image; The predicted speckle image acquisition module is used to input the preliminary reconstructed image of the original image into the transfer matrix to obtain the predicted speckle image; The deep empirical neural network training module is used to construct a total loss function based on the original speckle image and the predicted speckle image, and to train the deep empirical neural network to obtain a trained deep empirical neural network. The original image final reconstruction module is used to input the original speckle image into a trained deep empirical neural network to obtain the final reconstructed image of the original image.
[0036] The same or similar labels correspond to the same or similar parts; The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent. Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.
Claims
1. An optical phase retrieval method through a scattering medium, characterized in that, The method comprises the following steps: constructing a scattering system and obtaining a transmission matrix of the scattering system, constructing a deep empirical neural network based on the transmission matrix; obtaining a single original image, inputting the original image into the scattering system to obtain a corresponding original speckle image; inputting the original speckle image into the deep empirical neural network to obtain a preliminary reconstruction image of the original image; inputting the preliminary reconstruction image of the original image into the transmission matrix to obtain a predicted speckle image; constructing a total loss function according to the original speckle image and the predicted speckle image, training the deep empirical neural network to obtain a trained deep empirical neural network; inputting the original speckle image into the trained deep empirical neural network to obtain a final reconstruction image of the original image.
2. The optical phase retrieval method through a scattering medium according to claim 1, wherein, The scattering system comprises a laser, a polarizer, a laser beam expander, a first single-hole diaphragm, a phase spatial light modulator, a filtering system, a first objective lens, a multimode optical fiber, a second objective lens and a detector arranged in sequence. The filtering system comprises a first lens, a second single-hole diaphragm and a second lens arranged in sequence.
3. The optical phase retrieval method through a scattering medium according to claim 2, wherein, The method for obtaining a single original image and inputting the original image into the scattering system to obtain a corresponding original speckle image comprises the following steps: the laser receives the original image and emits a light beam to the polarizer; the polarizer converts the light beam into horizontally polarized light; the laser beam expander expands the horizontally polarized light into a plane wave with uniform intensity; the first single-hole diaphragm limits the plane wave with uniform intensity into a single plane wave; the phase spatial light modulator receives the single plane wave and modulates the single plane wave to obtain a modulated plane wave; the filtering system filters out the 0-order diffracted light of the modulated plane wave to obtain a filtered plane wave; the first objective lens couples the filtered plane wave to obtain a coupled plane wave; the multimode optical fiber receives the coupled plane wave and emits speckles; the second objective lens receives the speckles, collects and amplifies the speckles to obtain amplified speckles; the detector generates an original speckle image according to the amplified speckles.
4. The method of phase retrieval through a scattering medium of claim 1, wherein, The method for constructing a deep empirical neural network based on the transmission matrix comprises the following steps: establishing a mapping relationship of a full connection layer according to a mapping relationship in the transmission matrix; the deep empirical neural network comprises a first encoding block, a second encoding block, a third encoding block, a fourth encoding block, a fifth encoding block, a sixth encoding block, a first connection point, a first decoding block, a second connection point, a second decoding block, a third connection point, a third decoding block, a fourth connection point, a fourth decoding block, a fifth connection point, a fifth decoding block, a feature connection point and the full connection layer; the first encoding block, the second encoding block, the third encoding block, the fourth encoding block, the fifth encoding block, the sixth encoding block, the first connection point, the first decoding block, the second connection point, the second decoding block, the third connection point, the third decoding block, the fourth connection point, the fourth decoding block, the fifth connection point, the fifth decoding block and the feature connection point are connected in sequence; the output end of the first encoding block is further connected to the input end of the fifth connection point; the output end of the second encoding block is further connected to the input end of the fourth connection point; the output end of the third encoding block is further connected to the input end of the third connection point; An output end of the fourth encoding block is further connected with an input end of the second connection point; An output end of the fifth encoding block is further connected with an input end of the first connection point; Output ends of the first decoding block, the second decoding block, the third decoding block and the fourth decoding block are connected with input ends of the feature connection point.
5. The optical phase retrieval method through a scattering medium according to claim 4, wherein, The first encoding block, the second encoding block, the third encoding block, the fourth encoding block, the fifth encoding block and the sixth encoding block each comprise a plurality of down-sampling modules. The first decoding block, the second decoding block, the third decoding block, the fourth decoding block and the fifth decoding block each comprise a plurality of up-sampling modules.
6. The optical phase retrieval method through a scattering medium according to claim 1, wherein, The total loss function is constructed according to the original speckle image and the predicted speckle image, the deep empirical neural network is trained, and a trained deep empirical neural network is obtained, including: The weight parameters of the deep empirical neural network are initialized using a Gaussian distribution; The original speckle image and the predicted speckle image are input into the total loss function to obtain a loss function value; According to the loss function value, the weight parameters of the deep empirical neural network are updated by back propagation, and a performance evaluation function parameter between the original speckle image and the predicted speckle image is calculated, when the performance evaluation function parameter no longer improves, the back propagation is stopped, and the trained deep empirical neural network is obtained.
7. The optical phase retrieval method through a scattering medium according to claim 6, wherein, The updating of the weight parameters of the deep empirical neural network by back propagation includes: The calculation formula of the back propagation is as follows: wherein, is the updated weight parameter of the deep empirical neural network, is the network weight parameter of the deep empirical neural network, is the learning rate, is the gradient operation of the weight parameter of the deep empirical neural network with respect to the loss function value.
8. The optical phase retrieval method through a scattering medium according to claim 6, wherein, The performance evaluation function parameter includes structural similarity, Pearson correlation coefficient, peak signal-to-noise ratio and mutual information.
9. The method of phase retrieval through a scattering medium of claim 1, wherein, The expression of the total loss function is as follows: wherein is the predicted speckle image, is the original speckle image.
10. An optical phase retrieval system through a scattering medium for implementing the method of any one of claims 1-9, wherein, including: An initialization module is configured to construct a scattering system, obtain a transmission matrix of the scattering system, and construct a deep empirical neural network based on the transmission matrix; An original speckle image acquisition module is configured to acquire a single original image, input the original image into the scattering system, and obtain a corresponding original speckle image; An original image preliminary reconstruction module is configured to input the original speckle image into the deep empirical neural network to obtain a preliminary reconstruction image of the original image; A predicted speckle image acquisition module is configured to input the preliminary reconstruction image of the original image into the transmission matrix to obtain a predicted speckle image; A deep empirical neural network training module is configured to construct a total loss function according to the original speckle image and the predicted speckle image, train the deep empirical neural network, and obtain a trained deep empirical neural network; An original image final reconstruction module is configured to input the original speckle image into the trained deep empirical neural network to obtain a final reconstruction image of the original image.