Specific space heterodyne spectrum filtering method based on deep learning

Abstract

The invention provides a specific space heterodyne spectrum filtering method based on deep learning. Comprising the following steps: firstly, collecting a spatial heterodyne interferogram with clutters and a corresponding spatial heterodyne interferogram without clutters by using a spatial heterodyne spectrometer; then, Fourier transform is adopted to convert the spatial heterodyne interference pattern with the clutter and the spatial heterodyne interference pattern without the clutter into a spatial heterodyne spectrum with the clutter and a spatial heterodyne spectrum without the clutter; then, building a training set by taking the spatial heterodyne spectrum with clutter and the spatial heterodyne spectrum without clutter as input and output, and importing the training set into a deep neural network for training to obtain a trained filtering network model; and finally, performing Fourier transform on a space heterodyne interference pattern with clutter in actual measurement collected by the space heterodyne spectrometer in an actual detection process to obtain a space heterodyne spectrum with clutter in actual measurement, and importing the space heterodyne spectrum into the trained filtering network model to obtain a filtered space heterodyne spectrum. According to the invention, filtering processing can be rapidly and effectively carried out on a specific spatial heterodyne spectrum, spectral information of specific clutters is removed, and spectral information of a characteristic wave band of a measured target object is greatly reserved.

Description

(I) Technical field

[0001] The present invention relates to a specific spatial heterodyne spectrum filtering method based on deep learning, which can be used to efficiently filter out clutter information in a specific band of a specific spatial heterodyne spectrum and retain the spectral information of the detection target to the greatest extent, belonging to the field of spatial heterodyne spectrum technology. (II) Background technology

[0002] Spatial heterodyne spectroscopy is a new type of spatial modulation Fourier interferometry spectroscopy technology. It inherits the high signal-to-noise ratio, multi-channel transmission and high resolution of Fourier interferometry spectroscopy technology, while abandoning the moving mirror and scanning components of traditional Fourier interferometry spectroscopy technology. It can obtain ultra-high spectral resolution within a certain spectral range. At the same time, it has the advantages of no moving parts, compact size, light weight, and relatively high sensitivity. It has unique advantages in application fields such as atmospheric remote sensing, astronomical observation, ore detection and weather forecasting.

[0003] In the target detection performed by some spatial heterodyne spectrometers, in order to eliminate the interference of other substances, only the band where the target is sensitive to light absorption is usually selected for detection, so the acquired spectral shape is relatively fixed and only changes in strength at specific wavelengths. Therefore, this type of spatial heterodyne spectrometer is designed for spectral detection of specific targets. However, in actual detection, when the spatial heterodyne spectrometer has multiple detection tasks and the detection band is designed to be too large, there may be spectral information of clutter in other bands in addition to the spectral information of the characteristic band of the target, which will affect the subsequent concentration inversion. When this happens, the conventional filtering method of adding filters may be too inefficient and affect the spectral information of the target. Therefore, it is necessary to detect a new type of spatial heterodyne spectral filtering method, which can efficiently filter the specific band of a specific spatial heterodyne spectrum while retaining the spectral information of the detection target to the greatest extent.

[0004] In the past decade, the rapid development of deep learning has led to a climax in artificial intelligence technology, and it has achieved good applications in various fields. Deep learning is data-driven. By establishing a neural network to learn its hidden relationship from a large number of input and output data samples, and by feedback adjustment parameters, a nonlinear implicit mapping relationship is established to fit the input and output. In the field of optical information processing, it has provided a new idea for solving some problems such as digital holography, fringe analysis, phase unwrapping, ghost imaging, Fourier stack imaging, super-resolution microscopy, scattering medium imaging, optical tomography, super-surface optimization design, and spectral regression prediction, and has achieved remarkable results.

[0005] Therefore, a deep neural network is used to construct a nonlinear mapping relationship between heterodyne spectrum with clutter and heterodyne spectrum without clutter, so as to achieve efficient filtering of heterodyne spectrum with clutter and retain the spectral information of target characteristic wavelengths to a great extent. It can provide a new solution for spatial heterodyne spectrum filtering, and has important engineering use value in the field of spatial heterodyne spectrum technology and even in the field of atmospheric remote sensing. (III) Summary of the invention

[0006] The purpose of the present invention is to provide a specific space heterodyne spectrum filtering method based on deep learning, which can quickly and efficiently filter the specific space heterodyne spectrum and retain the spectral information of the characteristic band of the detection target to the greatest extent.

[0007] The object of the present invention is achieved in that:

[0008] A specific spatial heterodyne spectral filtering method based on deep learning comprises the following steps:

[0009] S1. First, a spatial heterodyne spectrometer is used to collect a spatial heterodyne interferogram with clutter and a corresponding spatial heterodyne interferogram without clutter.

[0010] S2. Then, Fourier transform is used to convert the spatial heterodyne interference pattern with clutter and the spatial heterodyne interference pattern without clutter into the spatial heterodyne spectrum with clutter and the spatial heterodyne spectrum without clutter;

[0011] S3, then, the spatial heterodyne spectrum with clutter and the spatial heterodyne spectrum without clutter are used as input and output to form a training set, and the training set is imported into the deep neural network for training to obtain a trained filter network model;

[0012] S4. Finally, the spatial heterodyne interference pattern with clutter collected by the spatial heterodyne spectrometer during the actual detection process is Fourier transformed into the spatial heterodyne spectrum with clutter, and then imported into the trained filtering network model to obtain the filtered spatial heterodyne spectrum.

[0013] Furthermore, the specific process of S1 includes the following steps:

[0014] The spatial heterodyne interferogram with clutter and the spatial heterodyne interferogram without clutter are in one-to-one correspondence, that is, after the spatial heterodyne interferogram with clutter and the spatial heterodyne interferogram without clutter are converted into the spatial heterodyne spectrum with clutter and the spatial heterodyne spectrum without clutter, the signal intensity at the target characteristic wavelength excluding clutter in the spatial heterodyne spectrum with clutter is consistent with the signal intensity at the target characteristic wavelength in the spatial heterodyne spectrum without clutter.

[0015] The optical path principle of the spatial heterodyne spectrometer is as follows: first, the light of the object to be measured passes through the aperture and collimating lens of the pre-collimation system. The collimating lens collimates the incident light to form light parallel to the optical axis, and then passes through the incident wavefront to enter the beam splitter. The beam splitter divides the incident light into two coherent light beams with the same energy. The reflected light and the transmitted light are respectively emitted to different blazed gratings. These two parts of light are emitted to the beam splitter again after diffraction by the blazed grating. Since the two coherent lights are emitted at different angles, spatial interference fringes are generated on the output wavefront, and finally an interference pattern appears on the electronic imaging detector.

[0016] Furthermore, the construction process of the deep neural network in S3 includes the following steps:

[0017] The constructed deep neural network includes input layer, hidden layer and output layer. The number of points in the input layer and output layer is determined according to the wavelength range and accuracy of the spectrum; the number of points and depth of the hidden layer are adjusted according to the complexity of the spectral information; the connection method between the layers is selected according to the specific task, and can be full connection, convolution and other methods.

[0018] The beneficial effects of this specific spatial heterodyne spectral filtering method based on deep learning are:

[0019] The present invention can train different network models according to a plurality of bands that need to be filtered of a specific spatial heterodyne spectrum. The trained model can filter the spatial heterodyne spectrum quickly and efficiently, and retain the spectral information of the characteristic band of the detection target to the greatest extent. (IV) Description of the drawings

[0020] Figure 1 It is a flowchart of a specific spatial heterodyne spectral filtering method based on deep learning;

[0021] Figure 2 It is a simplified diagram of the structure of a spatial heterodyne spectrometer;

[0022] Figure 3 It is a schematic diagram of a deep neural network for spatial heterodyne spectral filtering;

[0023] Figure 4 The figure is an example of the effect of a specific spatial heterodyne spectral filtering method based on deep learning; wherein, (a) is a spatial heterodyne interference pattern with noise; (b) is a spatial heterodyne interference pattern without noise; (c) is a spatial heterodyne spectrum with noise; (d) is a spatial heterodyne spectrum without noise; (e) is a filtered spatial heterodyne spectrum; (f) is the average spectral difference between all filtered spatial heterodyne spectra and the spatial heterodyne spectrum without noise in the test set. (V) Specific implementation methods

[0024] The present invention will be further described below in conjunction with specific embodiments.

[0025] This embodiment provides a specific spatial heterodyne spectral filtering method based on deep learning. Figure 1 As shown, the following steps are included:

[0026] W1. First, use spatial heterodyne spectroscopy to create a spatial heterodyne interferogram with clutter and a corresponding spatial heterodyne interferogram without clutter;

[0027] Furthermore, the specific process of W1 includes the following steps:

[0028] The spatial heterodyne spectrometer is designed and manufactured according to the principle of spatial heterodyne spectroscopy technology and specific detection targets. The spatial heterodyne interference pattern in this embodiment is obtained by simulating the input of the simulated spectrum into the program according to the principle of spatial heterodyne spectroscopy technology. The specific principle is as follows Figure 2 As shown: First, the light of the object to be measured passes through the aperture and collimating lens of the front collimation system. The collimating lens collimates the incident light to form light parallel to the optical axis, and then passes through the incident wavefront to enter the beam splitter. The beam splitter divides the incident light into two coherent light beams with the same energy. The reflected light and the transmitted light are respectively emitted to different blazed gratings. These two parts of light are diffracted by the blazed grating and then emitted to the beam splitter again. Since the two coherent lights are emitted at different angles, spatial interference fringes are generated on the output wavefront, and finally an interference pattern appears on the electronic imaging detector.

[0029] Based on this principle, a simulation program for a spatial heterodyne spectrometer with a detection wavelength of 756.8nm-771.648nm and a spectral resolution of 0.029nm was written. Then the simulated spectrum with noise and the corresponding spectrum without noise were input into the simulation program to obtain 800 spatial heterodyne interference patterns with noise and 800 corresponding spatial heterodyne interference patterns without noise of 1024×1024 pixels.

[0030] W2. Next, Fourier transform is used to convert the spatial heterodyne interference pattern with clutter and the spatial heterodyne interference pattern without clutter into the spatial heterodyne spectrum with clutter and the spatial heterodyne spectrum without clutter.

[0031] W3. Then, the spatial heterodyne spectrum with clutter and the spatial heterodyne spectrum without clutter are used as input and output to form a training set, and the training set is imported into the deep neural network for training to obtain the trained filtering network model.

[0032] Furthermore, the specific process of W3 includes the following steps:

[0033] The overall architecture of the deep neural network in this embodiment is as follows Figure 3As shown in the figure, each layer is connected in a fully connected manner. According to the wavelength range and spectral resolution of the spectrum, the input layer and the output layer are set to 1024 points; the number of hidden layers is set to 3, and the number of neurons in each layer is set to 1024. The learning rate is 0.00001, the training round is 10000, and the trained filter network model is obtained after the training optimization iteration.

[0034] W4. Use the simulation program of the spatial heterodyne spectrometer again, and simulate 200 spatial heterodyne interference images with different light intensity from those in the training set and the corresponding spatial heterodyne interference images without clutter as the test set. Use Fourier transform to convert the spatial heterodyne interference images with clutter and the spatial heterodyne interference images without clutter in the test set into the spatial heterodyne spectrum with clutter and the spatial heterodyne spectrum without clutter. Import the spatial heterodyne spectrum with clutter into the trained filter network model, and then the corresponding filtered spatial heterodyne spectrum can be output.

[0035] The effect of a specific spatial heterodyne spectral filtering method based on deep learning in this embodiment is as follows Figure 4 As shown. Figure 4 (a) is a randomly selected spatial heterodyne interferogram with clutter in the test set; Figure 4 (b) Figure 4 (a) The corresponding clutter-free spatial heterodyne interferogram; Figure 4 (c) Figure 4 (a) The cluttered spatial heterodyne spectrum after Fourier transform of the cluttered spatial heterodyne interferogram; Figure 4 (d) Figure 4 (b) The clutter-free spatial heterodyne spectrum after Fourier transformation of the clutter-free spatial heterodyne interferogram; Figure 4 (e) Figure 4 (c) Filtered spatial heterodyne spectrum after the clutter spatial heterodyne spectrum is processed by the filter network model; Figure 4 (f) The average spectral difference between the filtered spatial heterodyne spectra of all the test sets and the corresponding noise-free spatial heterodyne spectra. Figure 4 In (e), it can be seen that in the filtered spatial heterodyne spectrum, except for the spectral intensity information at the characteristic wavelengths of 767.5 nanometers and 781.6 nanometers, the clutter signals in the other spectral bands are all filtered out, and the peak heights at the characteristic wavelengths are almost the same as the peak heights at the characteristic wavelengths of the non-clutter spectrum. Figure 4From the average spectral difference between the spatial heterodyne spectrum after all filtering and the noise-free spatial heterodyne spectrum in the test set (f), it can be seen that the difference spectrum is close to a straight line. At the characteristic wavelengths of 767.5 nanometers and 781.6 nanometers, the average difference accounts for 0.013% and 0.011% of the peak value at the characteristic wavelength of the noise-free spatial heterodyne spectrum, respectively. This result shows that a specific spatial heterodyne spectrum filtering method based on deep learning can effectively process clutter signals, and the target feature information of the filtered spectrum is highly retained.

[0036] The network structure constructed by the present invention is clear, and the trained filtering network model can quickly and efficiently remove the clutter spectral information of a specific band for a specific spatial heterodyne spectrum, and greatly retain the spectral information of the characteristic band of the measured target.

[0037] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only effective embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A specific spatial heterodyne spectral filtering method based on deep learning, comprising the following steps: Firstly, a spatial heterodyne spectrometer is used to collect spatial heterodyne interferograms with and without clutter; Next, Fourier transform is used to convert the spatial heterodyne interference pattern with clutter and the spatial heterodyne interference pattern without clutter into the spatial heterodyne spectrum with clutter and the spatial heterodyne spectrum without clutter; Then, the spatial heterodyne spectrum with clutter and the spatial heterodyne spectrum without clutter are used as input and output to form a training set, and the training set is imported into the deep neural network for training to obtain the trained filter network model; Finally, the spatial heterodyne interference pattern with clutter collected by the spatial heterodyne spectrometer during the actual detection process is Fourier transformed into the spatial heterodyne spectrum with clutter, and then imported into the trained filter network model to obtain the filtered spatial heterodyne spectrum.

2. The method for specific spatial heterodyne spectral filtering based on deep learning according to claim 1, characterized in that: The spatial heterodyne interference pattern is collected by a spatial heterodyne spectrometer. The optical path principle of the spatial heterodyne spectrometer is as follows: first, the light of the object to be measured passes through the aperture and collimating lens of the pre-collimation system. The collimating lens collimates the incident light to form light parallel to the optical axis, and then passes through the incident wavefront to enter the beam splitter. The beam splitter divides the incident light into two coherent light beams with the same energy. The reflected light and the transmitted light are respectively emitted to different blazed gratings. These two parts of light are emitted to the beam splitter again after diffraction by the blazed grating. Since the two coherent lights are emitted at different angles, spatial interference fringes are generated on the output wavefront, and finally an interference pattern appears on the electronic imaging detector.

3. The method for specific spatial heterodyne spectral filtering based on deep learning according to claim 1, characterized in that: The spatial heterodyne interferogram with clutter and the spatial heterodyne interferogram without clutter are in one-to-one correspondence, that is, after the spatial heterodyne interferogram with clutter and the spatial heterodyne interferogram without clutter are converted into the spatial heterodyne spectrum with clutter and the spatial heterodyne spectrum without clutter, the signal intensity at the target characteristic wavelength excluding clutter in the spatial heterodyne spectrum with clutter is consistent with the signal intensity at the target characteristic wavelength in the spatial heterodyne spectrum without clutter.

4. The method for specific spatial heterodyne spectral filtering based on deep learning according to claim 1, characterized in that: The constructed deep neural network includes input layer, hidden layer and output layer. The number of points in the input layer and output layer is determined according to the wavelength range and accuracy of the spectrum; the number of points and depth of the hidden layer are adjusted according to the complexity of the spectral information; the connection method between the layers is selected according to the specific task, and can be full connection, convolution and other methods.

5. The method for specific spatial heterodyne spectral filtering based on deep learning according to claim 1, characterized in that: The trained filtering network model is a network model that is applied to a specific band of a specific spatial heterodyne spectrum to effectively remove clutter information.

Images