Noise space heterodyne interferometric spectral information correction method based on deep neural network
A method for correcting spectral information in noisy spatial heterodyne interferometry is constructed by using deep neural networks. This method solves the problem of spectral information loss under noise interference and enables efficient extraction of denoised spectra from noisy spatial heterodyne interferograms while preserving the integrity of spectral information.
Patent Information
- Application Number
- CN202410116551.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-29
- Publication Date
- 2026-08-25
- Estimated Expiration
- 2044-01-29
AI Technical Summary
Existing technologies struggle to effectively extract corrected denoised spatial heterodyne spectra from noisy spatial heterodyne interferograms under noise interference, resulting in spectral information loss.
A method for correcting noise spatial heterodyne interferometric spectral information is constructed using a deep neural network. By learning the mapping relationship between noise and ideal spectrum through the training set, a nonlinear implicit mapping relationship is established, and the corrected noise-reduced spatial heterodyne spectrum is directly extracted from the noise spatial heterodyne interferogram.
It can quickly and efficiently extract the corrected denoised spatial heterodyne spectrum from the noisy spatial heterodyne interferogram, simplify the processing of noise interference spectral information, and retain the spectral information of the measured object to a great extent.
Smart Images
Figure CN117928734B_ABST
Abstract
Description
(I) Technical Field
[0001] This invention relates to a method for correcting noise spatial heterodyne interferometric spectral information based on deep neural networks. It can be used to directly obtain the corrected noise-reduced spatial heterodyne spectrum from the noise spatial heterodyne interferogram, and belongs to the field of spatial heterodyne spectroscopy technology. (II) Background Technology
[0002] Spatial heterodyne spectroscopy is a novel spatial modulation Fourier interferometry technique. It inherits the high signal-to-noise ratio, multiplexing, and high resolution of traditional Fourier interferometry while eliminating the moving mirrors and scanning components. Within a defined spectral range, it achieves ultra-high spectral resolution. Furthermore, it boasts advantages such as no moving parts, compact size, light weight, and relatively high sensitivity, making it uniquely advantageous in applications such as atmospheric remote sensing, astronomical observation, mineral exploration, and weather forecasting. Although spatial heterodyne spectrometers offer significant advantages, interference from complex experimental environments and the inherent characteristics of electronic components can lead to signal noise. This noise can mask the spectral signal's features, preventing the acquisition of valuable research data. Therefore, correcting noisy spatial heterodyne interferometry spectral information is crucial for achieving high-precision target detection.
[0003] Current research on denoising spatial heterodyne interferograms includes traditional algorithms such as median filtering and wavelets, as well as deep learning algorithms like convolutional neural network denoising. These algorithms remove noise from the image itself, attempting to visually restore the original appearance of the image and recover the spectral information of the measured object recorded by the interference fringes. However, these algorithms inevitably alter the interference fringes to some extent, resulting in the loss of some spectral information. Therefore, a more efficient method is needed to extract the corrected, denoised spatial heterodyne spectrum from noisy spatial heterodyne interferograms.
[0004] As a rapidly developing field over the past decade, deep learning excels at extracting increasingly abstract feature representations from raw input data. It learns hidden relationships from a large number of input and output data samples by building neural networks, and through feedback parameter adjustment, establishes nonlinear implicit mapping relationships to fit the input and output. In the field of optical information processing, deep neural networks have provided a novel approach to solving problems in digital holography, fringe analysis, phase unwrapping, ghost imaging, Fourier layered imaging, super-resolution microscopy, scattering medium imaging, optical tomography, and metasurface optimization design, achieving significant results. Therefore, how to fit the mapping relationship between noisy spatial heterodyne interferograms and ideal spatial heterodyne spectra using deep neural networks, enabling the network to automatically learn to remove noise and directly extract corrected denoised spatial heterodyne spectra from noisy spatial heterodyne interferograms, has significant engineering application value and theoretical guiding significance in the field of spatial heterodyne spectroscopy and even atmospheric remote sensing. (III) Summary of the Invention
[0005] The purpose of this invention is to provide a method for correcting noisy spatial heterodyne interferometric spectral information based on deep neural networks. This method can simplify the processing of noisy interferometric spectral information and quickly and efficiently obtain the corrected denoised spatial heterodyne spectrum from the noisy spatial heterodyne interferogram.
[0006] The objective of this invention is achieved as follows:
[0007] A method for correcting spatial heterodyne interferometric spectral information of noise based on deep neural networks includes the following steps:
[0008] S1. Obtain noise-free spatial heterodyne interferograms using spatial heterodyne spectroscopy.
[0009] S2. Simultaneously, spatial heterodyne spectroscopy is used to obtain a noisy spatial heterodyne interferogram corresponding to the above noiseless spatial heterodyne interferogram.
[0010] S3. Obtain the ideal spatial heterodyne spectrum by performing Fourier transform and other methods on the noise-free spatial heterodyne interferogram;
[0011] S4. Use the light intensity information in the noisy spatial heterodyne interferogram as the input of the training set, and use the spectral information of the ideal spatial heterodyne spectrum extracted from the corresponding noiseless spatial heterodyne interferogram as the output of the training set. Then, import them into the constructed deep neural network for training.
[0012] S5. The deep neural network undergoes multiple parameter adjustments and iterative optimizations to obtain a trained network model.
[0013] S6. When applying this technique, spatial heterodyne spectroscopy is used to detect the target. If noise interference is present, a noise spatial heterodyne interferogram is obtained.
[0014] S7. Import the noise spatial heterodyne interferogram into the trained network model;
[0015] S8. The trained network model directly outputs the denoised spatial heterodyne spectrum from the noisy spatial heterodyne interferogram.
[0016] Furthermore, the specific process of S1 includes the following steps:
[0017] For simulated data, a simulation program for a spatial heterodyne spectrometer is written based on the principle of spatial heterodyne spectroscopy. The spectrum is imported into the simulation program to obtain the corresponding simulated noiseless spatial heterodyne interferogram. For measured data, a spatial heterodyne spectrometer designed and manufactured using the principle of spatial heterodyne spectroscopy is used to detect the target and obtain the corresponding measured noiseless spatial heterodyne interferogram.
[0018] The principle of spatial heterodyne spectroscopy is as follows: First, the light from the object being measured passes through the aperture and collimating lens of the pre-collimation system. The collimating lens collimates the incident light into light parallel to the optical axis, and then the light enters the beam splitter after passing through the incident wavefront. The beam splitter splits the incident light into two coherent beams of the same energy. The reflected light and the transmitted light are directed to different blazed gratings. After diffraction by the blazed gratings, these two parts of light are directed to the beam splitter again. Since the two coherent beams are emitted at different angles, spatial interference fringes are generated on the output wavefront, and finally a spatial heterodyne interferogram is formed on the electronic imaging detector.
[0019] Furthermore, the specific process of S2 includes the following steps:
[0020] For simulated data, a program can be written to add corresponding noise to the simulated spatial heterodyne interferogram to obtain a noisy spatial heterodyne interferogram. For measured data, after obtaining the measured noiseless spatial heterodyne interferogram using a spatial heterodyne spectrometer, non-target elements or other noise interference factors can be mixed in to obtain the corresponding noisy spatial heterodyne interferogram again. Alternatively, a program can be written to add corresponding noise to the measured noiseless spatial heterodyne interferogram to obtain a noisy spatial heterodyne interferogram.
[0021] Furthermore, the construction process of deep neural networks in S3 includes the following steps:
[0022] Deep neural networks mainly consist of an input layer, hidden layers, and an output layer. The number of neurons in the input layer is determined by the pixel values of the electron imaging detector in the spatial heterodyne spectrometer; the number of neurons in the output layer is determined by the detection band and spectral accuracy of the spatial heterodyne spectrometer; and the depth of the hidden layers is adjusted according to the complexity of the extracted spectral information.
[0023] The beneficial effects of this noise spatial heterodyne interferometric spectral information correction method based on deep neural networks are as follows:
[0024] This invention can train different network models according to different noise levels. The trained model can quickly and efficiently extract the corrected denoised spatial heterodyne spectrum from the noise spatial heterodyne interferogram, simplifying the noise interferometric spectral information processing process and preserving the spectral information of the measured object to a great extent. (iv) Description of the attached drawings
[0025] Figure 1 This is a flowchart illustrating the noise spatial heterodyne interferometric spectral information correction method based on deep neural networks;
[0026] Figure 2 This is a schematic diagram of the principle of spatial heterodyne spectroscopy.
[0027] Figure 3 This is a schematic diagram of a deep neural network for correcting spatial heterodyne interference spectral information in noisy environments;
[0028] Figure 4 The diagram shows the effect of a noise spatial heterodyne interference spectral information correction method based on deep neural networks on monochromatic light; where (a) is the ideal and noise spectra of monochromatic light; (b) is the ideal and denoised spectra of monochromatic light; and (c) is the spectral difference between the noise and ideal spectra of monochromatic light and the spectral difference between the denoised and ideal spectra.
[0029] Figure 5 The diagram shows the effect of a noise spatial heterodyne interference spectral information correction method based on deep neural networks on continuous light; where (a) is the ideal and noise spectra of continuous light; (b) is the ideal and denoised spectra of continuous light; and (c) is the spectral difference between the noise and ideal spectra of continuous light and the spectral difference between the denoised and ideal spectra. (V) Detailed Implementation
[0030] The present invention will be further illustrated below with reference to specific embodiments.
[0031] Example 1:
[0032] The noise spatial heterodyne interferometric spectral information correction method based on deep neural networks described in this embodiment, such as... Figure 1 As shown, it includes the following steps:
[0033] W1. Obtain noise-free spatial heterodyne interferograms using spatial heterodyne spectroscopy.
[0034] Furthermore, the specific process of W1 includes the following steps:
[0035] The spatial heterodyne interferogram in this embodiment is produced based on spatial heterodyne spectroscopy, the specific principle of which is as follows: Figure 2As shown: First, the light from the object being measured passes through the aperture and collimating lens of the pre-collimation system. The collimating lens collimates the incident light, forming light parallel to the optical axis, which then enters the beam splitter after passing through the incident wavefront. The beam splitter splits the incident light into two coherent beams of equal energy. The reflected and transmitted light are directed towards different blazed gratings. After diffraction by the blazed gratings, these two beams are directed back towards the beam splitter. Because the two coherent beams exit at different angles, spatial interference fringes are generated on the output wavefront, and finally, an interference pattern appears on the electronic imaging detector. Based on this principle, a simulation program for a spatial heterodyne spectrometer with a detection wavelength of 756.8 nm-771.648 nm and a spectral resolution of 0.029 nm was written. Then, the spectrum of simulated monochromatic light was input into the simulation program to obtain 1000 corresponding 1024×1024 pixel spatial heterodyne interferograms.
[0036] W2. Write a program to add Gaussian noise of sigma=25 to the above 1000 spatial heterodyne interferograms to obtain 1000 noisy spatial heterodyne interferograms.
[0037] W3. Obtain 1000 ideal spatial heterodyne spectra by Fourier transforming 1000 noise-free spatial heterodyne interferograms;
[0038] W4. Use the light intensity information of the above 1000 noisy spatial heterodyne interferograms as the input of the training set, and the spectral information of 1000 ideal spatial heterodyne spectra as the output of the training set, and import them into the deep neural network for training.
[0039] Furthermore, the construction process of the deep neural network in W4 is as follows:
[0040] The overall architecture of the deep neural network in this embodiment is as follows: Figure 3 As shown, each layer is connected using a fully connected architecture. Specifically, the input layer has 1024 neurons based on the detector's 1024×1024 pixel size; the input layer also has 512×2=1024 neurons based on the detection wavelength of 756.8nm-771.648nm and the spectral resolution of 0.029nm; the hidden layers initially consist of 3 layers, each with 512 neurons.
[0041] W5. The deep neural network is trained using a training set consisting of light intensity information from a noisy spatial heterodyne interferogram and spectral information from an ideal spatial heterodyne spectrum. The learning rate is 0.00001, and the training epochs are 10,000. After multiple parameter adjustments and iterative optimizations, a trained network model is obtained. This trained network model has 8 hidden layers, with 1024 neurons in each layer.
[0042] W6. Using the simulation program of the spatial heterodyne spectrometer again, 200 noiseless spatial heterodyne interferograms of monochromatic light with intensity different from that in the training set (pixel count 1024×1024) were simulated. A program was then written to add Gaussian noise (sigma=25) to these 200 noiseless spatial heterodyne interferograms, resulting in 200 noisy spatial heterodyne interferograms.
[0043] W7. Import 200 noisy spatial heterodyne interferograms directly into the trained network model.
[0044] W8. The trained network model directly obtains the corrected denoised spatial heterodyne spectrum from the noisy spatial heterodyne interferogram.
[0045] The effect of the noise spatial heterodyne interferometric spectral information correction method based on deep neural networks in this embodiment is as follows: Figure 4 As shown. Figure 4 (a) shows the ideal and noise spectra of monochromatic light, where the ideal spectrum refers to the average spectral line of the spatial heterodyne spectrum extracted from 200 noise-free spatial heterodyne interferograms, and the noise spectrum refers to the average spectral line of the noise spatial heterodyne spectrum extracted from 200 noise spatial heterodyne interferograms. As can be seen from the figure, there is a large difference between the noise spectrum and the ideal spectrum. Figure 4 (b) shows the ideal and denoised spectra of monochromatic light. The ideal spectrum refers to the average spectral line of the spatial heterodyne spectrum extracted from 200 noise-free spatial heterodyne interferograms, and the denoised spectrum refers to the average spectral line of the corrected denoised spatial heterodyne spectrum extracted from 200 noisy spatial heterodyne interferograms by the deep neural network. As can be seen from the figure, the denoised spectrum and the ideal spectrum almost overlap. Figure 4 (c) represents the spectral difference between monochromatic light noise and the ideal spectrum, and the spectral difference between the denoised light and the ideal spectrum. Wherein, Figure 4 (c) The above shows the spectral difference between the noise and the ideal spectrum. The difference is -20.63% and -15.58% at the two peaks, respectively, indicating that the noise will cause great interference to the ideal spectrum. Figure 4 (c) Below is the spectral difference between the denoised and ideal spectra. The difference is only -0.22% and -0.31% at the two peaks, respectively, indicating that deep neural networks can effectively extract near-ideal spectral information from the spatial heterodyne interferogram of monochromatic light disturbed by noise. The above results demonstrate the effectiveness of the noise spatial heterodyne interferogram spectral information correction method based on deep neural networks in monochromatic light.
[0046] Example 2:
[0047] The noise spatial heterodyne interferometric spectral information correction method based on deep neural networks described in this embodiment, such as... Figure 1 As shown, it includes the following steps:
[0048] W1. Obtain noise-free spatial heterodyne interferograms using spatial heterodyne spectroscopy.
[0049] Furthermore, the specific process of W1 includes the following steps:
[0050] The spatial heterodyne interferogram in this embodiment is produced based on spatial heterodyne spectroscopy, the specific principle of which is as follows: Figure 2 As shown: First, the light from the object being measured passes through the aperture and collimating lens of the pre-collimation system. The collimating lens collimates the incident light, forming light parallel to the optical axis, which then enters the beam splitter after passing through the incident wavefront. The beam splitter splits the incident light into two coherent beams of equal energy. The reflected and transmitted light are directed towards different blazed gratings. After diffraction by the blazed gratings, these two beams return to the beam splitter. Because the two coherent beams exit at different angles, spatial interference fringes are generated on the output wavefront, and finally, an interference pattern appears on the electronic imaging detector. Based on this principle, a simulation program for a spatial heterodyne spectrometer with a detection wavelength of 756.8 nm-771.648 nm and a spectral resolution of 0.029 nm was written. Then, the spectrum of simulated continuous light was input into the simulation program to obtain 1000 spatial heterodyne interferograms with corresponding pixels of 1024×1024.
[0051] W2. Write a program to add Gaussian noise of sigma=25 to the above 1000 spatial heterodyne interferograms to obtain 1000 noisy spatial heterodyne interferograms.
[0052] W3. Obtain 1000 ideal spatial heterodyne spectra by Fourier transforming 1000 noise-free spatial heterodyne interferograms;
[0053] W4. Use the light intensity information of the above 1000 noisy spatial heterodyne interferograms as the input of the training set, and the spectral information of 1000 ideal spatial heterodyne spectra as the output of the training set, and import them into the deep neural network for training.
[0054] Furthermore, the construction process of the deep neural network in W4 is as follows:
[0055] The overall architecture of the deep neural network in this embodiment is as follows: Figure 3 As shown, each layer is connected using a fully connected architecture. Specifically, the input layer has 1024 neurons based on the detector's 1024×1024 pixel size; the input layer also has 512×2=1024 neurons based on the detection wavelength of 756.8nm-771.648nm and the spectral resolution of 0.029nm; the hidden layers initially consist of 3 layers, each with 512 neurons.
[0056] W5. The deep neural network is trained using a training set consisting of light intensity information from a noisy spatial heterodyne interferogram and spectral information from an ideal spatial heterodyne spectrum. The learning rate is 0.00001, and the training epochs are 10,000. After multiple parameter adjustments and iterative optimizations, a trained network model is obtained. This trained network model has 8 hidden layers, with 1024 neurons in each layer.
[0057] W6. Using the simulation program of the spatial heterodyne spectrometer again, 200 noiseless spatial heterodyne interferograms with continuous light intensity different from the 1024×1024 pixel values in the training set were simulated. A program was then written to add Gaussian noise of sigma=25 to the 200 noiseless spatial heterodyne interferograms, resulting in 200 noisy spatial heterodyne interferograms.
[0058] W7. Import 200 noisy spatial heterodyne interferograms directly into the trained network model.
[0059] W8. The trained network model directly obtains the corrected denoised spatial heterodyne spectrum from the noisy spatial heterodyne interferogram.
[0060] The effect of the noise spatial heterodyne interferometric spectral information correction method based on deep neural networks in this embodiment is as follows: Figure 5 As shown. Figure 5 (a) shows the ideal and noise spectra of continuous light, where the ideal spectrum refers to the average spectral line of the spatial heterodyne spectrum extracted from 200 noise-free spatial heterodyne interferograms, and the noise spectrum refers to the average spectral line of the noise spatial heterodyne spectrum extracted from 200 noise spatial heterodyne interferograms. As can be seen from the figure, there is a large difference between the noise spectrum and the ideal spectrum. Figure 5 (b) shows the ideal and denoised spectra of continuous light. The ideal spectrum refers to the average spectral line of the spatial heterodyne spectrum extracted from 200 noise-free spatial heterodyne interferograms, and the denoised spectrum refers to the average spectral line of the corrected denoised spatial heterodyne spectrum extracted from 200 noisy spatial heterodyne interferograms by the deep neural network. As can be seen from the figure, the denoised spectrum and the ideal spectrum are almost identical. Figure 5 (c) represents the spectral difference between continuous light noise and the ideal spectrum, and the spectral difference between the denoised spectrum and the ideal spectrum. Wherein, Figure 5 (c) The above shows the spectral difference between noise and the ideal spectrum. The average difference is 204.75%, indicating that noise can cause significant interference to the ideal spectrum. Figure 5 (c) Below is the spectral difference between the denoised and ideal spectra. The average difference is only 1.15%, indicating that deep neural networks can effectively extract near-ideal spectral information from the spatial heterodyne interferogram of continuous light that is disturbed by noise. The above results demonstrate the effectiveness of the noise spatial heterodyne interferogram spectral information correction method based on deep neural networks in continuous light.
[0061] The network structure constructed in this invention is clear, and the trained network model can quickly and efficiently extract the corrected denoised spatial heterodyne spectrum directly from a specific noisy spatial heterodyne interferogram, simplifying the processing of noisy interferometric spectral information and preserving the spectral information of the measured object to a great extent.
[0062] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only an effective embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for correcting spatial heterodyne interferometric spectral information in the presence of noise based on deep neural networks, comprising the following steps: S1. Obtain noise-free spatial heterodyne interferograms using spatial heterodyne spectroscopy. S2. Use spatial heterodyne spectroscopy to obtain a noisy spatial heterodyne interferogram corresponding to the above noiseless spatial heterodyne interferogram; S3. Extract the ideal spatial heterodyne spectrum from the noiseless spatial heterodyne interferogram; S4. Use the light intensity information in the noisy spatial heterodyne interferogram as the input of the training set, and use the spectral information of the ideal spatial heterodyne spectrum extracted from the corresponding noiseless spatial heterodyne interferogram as the output of the training set. Then, import them into the constructed deep neural network for training. S5. The deep neural network undergoes multiple parameter adjustments and iterative optimizations to obtain a trained network model. S6. When applying this technique, spatial heterodyne spectroscopy is used to detect the target. If noise interference is present, a noise spatial heterodyne interferogram is obtained. S7. Import the noise spatial heterodyne interferogram into the trained network model; S8. The trained network model directly obtains the corrected denoised spatial heterodyne spectrum from the noise spatial heterodyne interferogram. S1 includes the following steps: Based on the principles of spatial heterodyne spectroscopy, a simulation program for a spatial heterodyne spectrometer is written. The spectrum is imported into the simulation program to obtain the corresponding simulated noiseless spatial heterodyne interferogram; or a spatial heterodyne spectrometer designed and fabricated using the principles of spatial heterodyne spectroscopy is used to detect the target and obtain the corresponding measured noiseless spatial heterodyne interferogram. S2 includes the following steps: Write a program to add corresponding noise to the simulated noiseless spatial heterodyne interferogram to obtain a noisy spatial heterodyne interferogram; or after obtaining the measured noiseless spatial heterodyne interferogram using a spatial heterodyne spectrometer, mix in non-target elements or other noise interference factors and then detect again to obtain the corresponding noisy spatial heterodyne interferogram; or write a program to add corresponding noise to the measured noiseless spatial heterodyne interferogram to obtain a noisy spatial heterodyne interferogram. The process of building a deep neural network in S4 includes the following steps: A deep neural network consists of an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is determined by the number of pixels in the electronic imaging detector of the spatial heterodyne spectrometer. The number of neurons in the output layer is determined by the detection band and spectral accuracy of the spatial heterodyne spectrometer. The depth of the hidden layer is adjusted according to the complexity of the extracted spectral information.
2. The method for correcting spatial heterodyne interferometric spectral information based on deep neural networks according to claim 1, characterized in that: The spatial heterodyne spectroscopy techniques described in S1, S2, and S6 are spatial heterodyne spectrometers or simulation algorithms based on the principle of spatial heterodyne. The specific principle is as follows: First, the light from the object being measured passes through the aperture and collimating lens of the pre-collimation system. The collimating lens collimates the incident light, forming light parallel to the optical axis, which then passes through the incident wavefront and enters the beam splitter. The beam splitter splits the incident light into two coherent beams of equal energy. The reflected and transmitted light are directed towards different blazed gratings. After diffraction by the blazed gratings, these two beams are directed back towards the beam splitter. Because the two coherent beams exit at different angles, spatial interference fringes are generated on the output wavefront, finally forming a spatial heterodyne interferogram on the electronic imaging detector.
3. The method for correcting spatial heterodyne interferometric spectral information based on deep neural networks according to claim 1, characterized in that: The noiseless spatial heterodyne interferogram described in S2 corresponds to the noisy spatial heterodyne interferogram. That is, if the noise interference in the noisy spatial heterodyne interferogram is completely removed, its light intensity information is consistent with that of the noiseless spatial heterodyne interferogram.