A method and system for interferometric spectral imaging restoration based on degradation simulation

Abstract

The present invention relates to an interference spectral imaging restoration method and system based on degradation simulation, and belongs to the technical field of remote sensing image processing. In the degradation modeling calibration, the degradation elements in the interference imaging are divided into trend items and noise, and the degradation process is modeled into three stages: instrument degradation, photosensitivity degradation, and signal-independent degradation. The interference imaging degradation model and the corresponding model calibration method are used to extract the degradation characteristics and model parameters of each stage for a specific instrument, generate corresponding calibration results, and use them to generate highly realistic interference data. In the restoration process, a multi-stage restoration algorithm framework is constructed to process the trend items, transformation calculations and other elements and noise in the degradation process separately. At the same time, a multi-stage simulation data set generation method is designed, which can generate a high-fidelity simulation interference pattern data set dedicated to training the entire algorithm from the spectral data set according to the calibration results. The present invention achieves more efficient and better-effective interference imaging restoration.

Description

Technical Field

[0001] The present invention relates to an interference spectrum imaging restoration method and system based on degraded data simulation and multi-stage deep learning, belonging to the technical field of remote sensing image processing. Background Art

[0002] Spectral imaging technology is a core perception technology in many fields. It can detect and obtain a data cube consisting of two-dimensional spatial information and one-dimensional spectral information of the target. After data processing, it can obtain spectral curves of different objects. Spectral imaging technology is divided into multiple categories. Among them, interferometric spectral imaging is an indirect spectral imaging method with a wide spectral range and high spectral resolution. It has the advantages of large light flux and high resolution. At present, interferometric imaging is mainly used in long-distance remote sensing, carried on satellites and spacecraft, and plays a key role in tasks such as environmental monitoring and aerospace remote sensing.

[0003] The main medium of interferometric imaging is an interferometric imaging spectrometer composed of an interferometer and a sensor. During the imaging process, light undergoes Fourier interference transformation through the optical path of the interferometric imaging spectrometer and is converted into an interference pattern. It then enters the sensor and is converted into a digital signal (hardware encoding). The downstream solution device executes a restoration algorithm to restore the acquired interference pattern to a spectrum through computational transformation (software decoding), completing indirect imaging. Interferometric imaging spectrometers can be divided into time, space, and joint time and space modulation according to different optical path principles.

[0004] As an indirect imaging method, interferometric imaging faces a complex link degradation problem. In the multiple imaging stages from optical path, sensor, electronic circuit to digital signal, different degradation components are introduced in each stage. These error terms are continuously transmitted and amplified in the link, resulting in obvious degradation and distortion of the interferometric data, which cannot be directly restored to spectral data. Therefore, in the calculation and restoration process, these degradation terms need to be actively restored.

[0005] However, existing restoration technologies are mainly concentrated in the fields of optics and signal processing, and have two limitations. First, most existing methods are based on the framework of model-driven algorithms, which have limited potential in performance and efficiency. They require a large number of complex iterative optimizations and serial calculations. Even in a hardware-accelerated environment, the efficiency is still insufficient and far from the level of real-time restoration. In addition, existing methods are relatively loose, with different algorithms distributed in different restoration stages and serialized with each other, lacking an end-to-end overall solution.

[0006] In the field of image processing and computer vision, there are a series of denoising and enhancement methods for spectral images, ranging from classic model-driven algorithms (such as BM3D, BM4D, etc.) to various recent deep learning algorithms. Deep learning technology (such as QRNN-3D, Uformer, SCUNet) breaks the limitations of model-driven algorithms and has achieved outstanding results in many fields. However, pure deep learning enhancement methods are difficult to train and require a large amount of paired low-quality / high-quality data for network learning. These data are extremely expensive to collect and difficult to meet training requirements. If data simulation methods are used for training, they will be very different from real data training, and it is difficult to achieve ideal results. In addition, deep learning technology generally faces the problem of being out of touch with actual instrument deployment and application. It is difficult to directly adapt to interferometric imaging spectrometers and cannot directly handle the link-type data degradation during the entire imaging process.

[0007] In summary, there is an urgent need for an interference imaging restoration and enhancement method with excellent results, real-time and high efficiency, which can break the limitations of existing model-driven algorithms and solve the difficulties of deep learning algorithms in training overhead and application methods. Summary of the invention

[0008] The purpose of the present invention is to start from the performance requirements of current interferometric imaging restoration methods, and to address the difficulties of current classical algorithms and deep learning technologies in performance and application, integrate the computational ideas of deep learning into the restoration framework of interferometric imaging, and creatively propose an interferometric spectral imaging restoration method and system based on degradation simulation.

[0009] The innovative features of the present invention include:

[0010] For the first time, the ideas of degradation mathematical modeling, high-fidelity data generation, and multi-stage restoration algorithms are applied to the process of interferometric imaging. In the degradation modeling calibration, the present invention divides the degradation elements in interferometric imaging into two parts: trend terms and noise, and models the degradation process into three stages: instrument degradation, photosensitivity degradation, and signal-independent degradation. Through the mining and experimentation of relative and absolute calibration data, a set of interferometric imaging degradation models and corresponding model calibration methods are proposed, which can extract the degradation characteristics and model parameters of each stage for a specific instrument, generate corresponding calibration results, and use them to generate highly realistic interferometric data. The model concisely and accurately expresses various degradation mechanisms and their corresponding performances in imaging results, ensuring the effect of data simulation. The calibration method is efficient and accurate, and can solve all model parameters with very low time overhead.

[0011] In addition, during the restoration process, considering the performance characteristics of the deep learning algorithm, the present invention constructs a multi-stage restoration algorithm framework. The algorithm framework processes trend items, transformation calculations and other elements in the degradation process separately from the noise. On the one hand, the framework contains a certain data path design, which can actively perform operations such as removing trend items and restoration based on the calibration results; on the other hand, between the above-mentioned active operation stages, a multi-stage plug-and-play deep learning denoising network is embedded to suppress the noise explosion in the restoration process, achieve high-quality restoration, and form a multi-stage framework structure. According to the above-mentioned algorithm architecture, considering that the multi-stage denoising network needs to be trained separately, the present invention designs a matching multi-stage simulation data set generation method, which can generate a high-fidelity simulation interference pattern data set dedicated to training the entire algorithm from a common spectral data set based on the calibration results.

[0012] The present invention is implemented by adopting the following technical solutions.

[0013] On the one hand, the present invention proposes an interference spectrum imaging restoration method based on degradation simulation, comprising the following steps:

[0014] Step 1: Interferometric imaging degradation modeling based on imaging mechanism and calibration data analysis.

[0015] Step 1.1: Use the interferometric imaging spectrometer to capture calibration data, and perform data sorting and preprocessing.

[0016] Among them, the calibration data is taken by using an interferometric imaging spectrometer, including relative calibration data and absolute calibration data. In addition, it is necessary to use the classical restoration method and absolute calibration data to obtain uniform spectrum restoration results. The above results are sorted out for use in the subsequent calibration process.

[0017] Step 1.2: Construct a calibration solution for the degradation model on the actual machine calibration data to determine the model parameters.

[0018] Specifically, according to the design of the calibration method, the calibration data is matched with the three-stage degradation model, and parameter calibration is performed to decompose the degradation elements layer by layer.

[0019] First, the signal-independent degradation parameters are determined by the black frame in the relative calibration data, and then the photosensitivity degradation parameters are determined by the uniform light data. Finally, the instrument degradation parameters are determined by the absolute calibration data and spectral data to obtain a set of complete degradation model parameters for the instrument.

[0020] Step 2: Staged interferometric imaging restoration based on deep learning.

[0021] Step 2.1: Use the calibration results to perform multi-stage training data simulation to generate the algorithm framework training data set.

[0022] Firstly, the noise-free hyperspectral dataset is preprocessed, including clipping transformation and spectral domain interpolation, to obtain a spectrogram in a specific format.

[0023] Then, the spectrum and calibration results are substituted into the degradation model, and the imaging process of the instrument is simulated through the calculation of degradation and transformation to generate simulated degradation interference data. According to the algorithm training requirements, the trend term and noise are processed separately, and the data simulation process is divided into multiple stages according to whether the simulated data has / has no trend term and noise, and multi-stage simulation data sets are generated in sequence.

[0024] Step 2.2: Embed the denoising network into the multi-stage restoration algorithm framework and train it using the dataset.

[0025] Through the aforementioned multi-stage simulation training set, the denoising network of different stages is trained respectively. The network is trained separately according to different stages, and the latter stage uses the intermediate results generated in the previous stage as the training source and uses the simulation data set as the label. After the training is completed, the denoising model is embedded in the algorithm framework to form a multi-stage restoration algorithm.

[0026] Step 3: Reconstruct a high-quality spectral image from the real interference data captured by the interferometer.

[0027] The real interference data taken by the interferometer under normal circumstances is input into the trained restoration algorithm, and high-quality spectral image results are obtained through restoration calculation.

[0028] Furthermore, objective evaluation indicators such as peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) can be used to evaluate the restoration quality.

[0029] On the other hand, the present invention also proposes an interference spectrum imaging restoration system based on degradation simulation, including a denoising network A, a denoising network B, a denoising network C, a trend term removal module, an FFT (fast Fourier transform) module, and a spectral domain efficiency restoration module.

[0030] Denoising network A is used for preliminary denoising and restoration of data;

[0031] The trend item removal module is used to estimate and remove the main trend items of the data under the interference mode;

[0032] The denoising network B is used to post-process the first two stages and suppress the residual and amplified errors;

[0033] The FFT module is used to perform Fourier transform, spectral domain adjustment, interpolation and other operations on the data to convert the interference pattern into a spectrum pattern;

[0034] The spectrum domain efficiency restoration module is used to process the trend item of data in the spectrum mode;

[0035] The denoising network C is used to post-process the transformation restoration process, suppress the residual and amplified errors, and enhance the final restoration result.

[0036] The connection relationship between the above modules is:

[0037] The output end of the denoising network A is connected to the input end of the trend item removal module;

[0038] The output end of the trend item removal module is connected to the input end of the denoising network B;

[0039] The output of the denoising network B is connected to the input of the FFT module;

[0040] The output end of the FFT module is connected to the input end of the spectrum domain efficiency restoration module;

[0041] The output end of the spectral domain efficiency restoration module is connected to the input end of the denoising network C. Beneficial Effects

[0042] Compared with the prior art, the present invention has the following advantages:

[0043] 1. This invention proposes a multi-stage degradation mechanism model for interferometric imaging. The total number of parameters of this model is small, but it can accurately describe the components and characteristics of interferometric imaging degradation. The overall structure is clear and reasonable, ensuring the convenience of subsequent data simulation in execution and the fidelity of the results.

[0044] 2. The present invention designs a parameter calibration and data simulation method based on a degradation model. The method is simple and efficient, can complete large-scale calculations in a short time, and can accurately evaluate model parameters and generate high-fidelity simulation data. In addition, the method is closely related to the actual imaging process of the instrument, overcoming the problem of the general disconnection between deep learning algorithms and actual instruments and applications at the training level.

[0045] 3. This paper designs a multi-stage restoration algorithm for interferometric imaging. The algorithm is based on deep learning, breaking the limitations of existing algorithms in performance and efficiency; using simulation data training to solve the difficulties of training; and through framework design, it provides a new idea of ​​integrating deep learning into interferometric imaging restoration.

[0046] The present invention provides a data basis for deep learning restoration through degradation modeling, calibration and simulation methods; and gradually removes the degradation components through the construction of a multi-stage restoration algorithm to achieve more efficient and effective interference imaging restoration operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1It is a schematic diagram of the process of the present invention;

[0048] Figure 2 Schematic diagram of the degradation model in the method of the present invention;

[0049] Figure 3 Schematic diagram of the calibration and data simulation method in the method of the present invention;

[0050] Figure 4 Schematic diagram of the restoration algorithm in the method of the present invention. DETAILED DESCRIPTION

[0051] The present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0052] Example

[0053] like Figure 1 As shown, an interference spectrum imaging restoration method based on degradation simulation includes the following steps:

[0054] Step 1: Interferometric imaging degradation modeling based on imaging mechanism and calibration data analysis.

[0055] Step 1.1: Capture calibration data using an interferometric imaging spectrometer and perform data organization and preprocessing.

[0056] In order to calibrate the degradation model of a specific interferometric imaging spectrometer, it is necessary to capture the calibration data through the instrument. For the sake of convenience and uniformity, the present invention uses calibration data consistent with the calibration process of the classical restoration algorithm, and can directly use the existing calibration results without the need to capture new data.

[0057] In this embodiment, the calibration data is divided into three parts, such as Figure 2 -a as shown.

[0058] The first is the relative calibration data. The relative calibration data is obtained by removing the interferometer optical path in the instrument and only using the uniform light field data captured by the push-scanning sensor in the instrument. Since there is no interference process, the data does not contain interference curves in the interference domain. The purpose is to extract the degradation components of the instrument and only retain the degradation information related to the sensor. Under a uniform light field, the ideal imaging result is a completely consistent and uniform picture in both the spatial domain and the interference domain, while the various noises and textures in the data at this time are information related to data degradation. The relative calibration data includes one with 0 brightness (darkness) and several with different uniform brightness, which can be used to extract and calibrate signal-independent degradation and photosensitivity degradation components, respectively.

[0059] The second is the absolute calibration data. The absolute calibration data is obtained by shooting the same uniform light field with the entire instrument. The difference from the relative calibration data is that an interferometer is installed in the instrument, so the ideal imaging result should be a completely consistent interference curve in the spatial domain. The absolute calibration data also needs to contain several interference images with different uniform brightness (no need for 0 brightness), and the instrument degradation parameters are extracted by comparing with the relative calibration data.

[0060] The third is the spectral restoration result from the absolute calibration data. The data is obtained from the absolute calibration data by the classical restoration method and is used to extract the two parameters of the instrument degradation, the background image and the spectral efficiency.

[0061] After the above shooting and sorting, the calibration data set is obtained and the calibration process begins.

[0062] Step 1.2: Perform calibration and solution of the degradation model on the actual machine calibration data to determine the model parameters.

[0063] Specifically, the degradation model used in the present invention is designed based on the basic imaging theory and the mining process of calibration data, and is synchronously designed with the calibration method of the model itself. Figure 2 -b, the degradation model is divided into three parts: instrument degradation, photosensitivity degradation and signal-independent degradation, which correspond to the three main stages of interference imaging: the light signal passes through the interference optical path (converted from spectral light signal to interference light signal), is converted into an electrical signal by the photoelectric effect in the sensor, and is gradually converted from an electrical signal to a digital signal in the sensor circuit.

[0064] In these three stages, degradation is carried out in a chain-like manner of function compounding, corresponding to the imaging link. All degradation components are defined as noise and trend terms, where noise refers to random terms in the image, manifested as noise points, random strips / surfaces (at different scales), etc., while trend terms refer to degradation terms with fixed patterns, which are attached to the source data through operations such as addition / multiplication in different directions.

[0065] Furthermore, the calibration process of the degradation model can be regarded as a simple manual restoration process, which requires constantly stripping the composite degradation function from the calibration data and trying to determine the information of the degradation component. The order of the calibration process is opposite to that of the degradation model, and the signal-independent, photosensitive and instrument degradation parameters are determined in turn. In the calibration process, two different methods are mainly used to calculate the trend term and noise parameter. One is to weaken the amplitude of the noise data and extract the trend term component by taking the average of a large number of similar / related samples; the other is that for noise, the characteristic parameters can be extracted by fitting the probability distribution.

[0066] After the model calibration process, the model parameters can be obtained. The model parameters contain all the degradation components of the current instrument and can be used to simulate the instrument's shooting process with high accuracy.

[0067] Step 2: Staged interferometric imaging restoration based on deep learning.

[0068] Step 2.1: Perform multi-stage training data simulation based on the calibration results to generate the algorithm framework training data set.

[0069] Data simulation is the process of generating degraded interferograms based on ideal spectral source data without degradation and model parameters. First, collect existing noise-free multispectral / hyperspectral data sets, crop, scale, and interpolate the data sets in the spectral domain to make the data format consistent with the imaging format of the instrument to obtain source data. Then, substitute the source data and calibration parameters into the degradation model, and generate a simulated data set by calculation, such as Figure 3 -a as shown.

[0070] According to the staged training strategy design of the algorithm framework, the data set required by the present invention has a multi-stage structure, including four parts, such as Figure 3 -b. The first is the source data, which is used as the training target of the final restoration result network C; the second is the interference data without trend terms and noise, which is used as the training target of network B. The third is the interference data with trend terms and without noise, which is used as the training target of network A; the fourth is the high-fidelity degradation data with both, which is used as the training source of the entire algorithm.

[0071] Furthermore, in order to generate a multi-stage data set, intermediate results are generated in the model simulation calculation and output as the second part. At the same time, the simulation of the trend term and the noise are distinguished and output as the third and fourth parts respectively.

[0072] The above steps can be completed by only making some detailed modifications to the basic model operations. The generated multi-stage simulation data set can be directly used in the training process of the algorithm network.

[0073] Step 2.2: Embed the denoising network into the multi-stage restoration algorithm framework and train it using the dataset.

[0074] Specifically, the algorithm framework has a multi-stage structure, covering the complete process from interference imaging results to ideal spectral restoration results, such as Figure 4 As shown in the figure, there are three modules in the framework that need to be embedded in the denoising network, denoted as A, B, and C, and trained.

[0075] The denoising algorithm and its corresponding training process in the present invention are introduced below in the order of data paths.

[0076] After the degraded data enters the algorithm structure, it is directly input into the denoising network A. A is an interference pattern denoising network, which is used to remove the main noise in the interference pattern degradation process. The training source of A is the degraded data containing all noise and trend term components, and the training target is the degraded data containing only trend terms. After denoising by A, the data quality is improved and the noise amplitude is reduced. Subsequently, the trend term in the data is estimated and eliminated in combination with the calibration data. Ideally, through the above two steps, the trend term and noise under the interference pattern mode are removed, and interference data that can be directly used for Fourier transform can be obtained.

[0077] However, in reality, the quality of the data obtained through the above steps is still seriously damaged. There are two reasons for this: on the one hand, there is residual noise in the data after passing through A, and these noises will be amplified due to errors in the trend image calculation (noise explosion); on the other hand, the estimation and elimination of the trend image is not completely accurate like the simulation process, and there will be residual components. For this reason, the interference pattern denoising network B is used to perform subsequent processing on the above process to suppress the error amplification problem in the calculation. The training source of B is the intermediate result obtained after the data passes through A and the trend item elimination module, and the training target is the interference pattern without trend items and noise.

[0078] Furthermore, the data is directly transformed and efficiently restored in the spectral domain after passing through B. Similarly, the degenerate components left by B will be amplified in the restoration operation (especially Fourier transform). For this reason, a denoising network C is introduced after all calculations as the post-processing of the entire algorithm. C is a spectral image denoising network. The training source is the intermediate result of the data after passing through B and the restoration operation. The training target is the ideal spectral data. The output of C is the output result of the entire restoration algorithm.

[0079] The training process of the above three-stage network can all use the L1 denoising loss function. The commonly used loss functions in the field of image denoising are mainly L1 loss function and L2 loss function. Among them, the L1 loss function is the sum of the absolute value of the true value and the predicted value of each pixel in the image. Let the predicted value be The training objective is , expressed as:

[0080] in, represents the loss function value of the target network, Represents the total size of the denoising network (the product of the number of pixels and the spectral resolution).

[0081] After the above multi-stage algorithm training process, the algorithm model is ready and can be used to verify experiments and restore real data.

[0082] The above three denoising networks A, B, and C can be embedded in various 3D denoising models, and the training process is separate. Therefore, under the premise of saving the intermediate results, the trained model can be plug-and-play and has flexibility.

[0083] Step 3: Reconstruct a high-quality spectral image from the real interference data captured by the interferometer.

[0084] After the above algorithm framework is trained, the trained model is embedded in A, B, and C to form a complete restoration algorithm.

[0085] The real interference data taken by the interferometer under normal circumstances is input into the trained restoration algorithm, and high-quality spectral image results are obtained through restoration calculation.

[0086] Furthermore, an interferometric spectral imaging restoration system based on degradation simulation inputs the training data into the algorithm for restoration. During the restoration process, the data passes through all modules and stages of the network in sequence, and the final restoration result is obtained, such as Figure 4 shown.

[0087] Denoising network A is used for preliminary denoising and restoration of data;

[0088] The trend item removal module is used to estimate and remove the main trend items of the data under the interference mode;

[0089] The denoising network B is used to post-process the first two stages and suppress the residual and amplified errors;

[0090] The FFT module is used to perform Fourier transform, spectral domain adjustment, interpolation and other operations on the data to convert the interference pattern into a spectrum pattern;

[0091] The spectrum domain efficiency restoration module is used to process the trend item of data in the spectrum mode;

[0092] The denoising network C is used to post-process the transformation restoration process, suppress the residual and amplified errors, and enhance the final restoration result.

[0093] The connection relationship between the above modules is:

[0094] The input imaging interference image is connected to the input of the denoising network A.

[0095] The output end of the denoising network A is connected to the input end of the trend item removal module;

[0096] The output end of the trend item removal module is connected to the input end of the denoising network B;

[0097] The output of the denoising network B is connected to the input of the FFT module;

[0098] The output end of the FFT module is connected to the input end of the spectrum domain efficiency restoration module;

[0099] The output end of the spectrum domain efficiency restoration module is connected to the input end of the denoising network C;

[0100] The output of the denoising network C is connected to the restored spectral image.

[0101] Furthermore, the quality of the restored results is evaluated by calculating the PSNR and SSIM indicators compared with the corresponding ideal images. Since the network structure in the algorithm saves intermediate results and is trained separately, the verification results should be equivalent to the restoration effects obtained in training. Since deep learning is used as the main computing body, the algorithm is highly efficient and can complete the same restoration task in a few tenths of the time of the classical algorithm.

[0102] Furthermore, after the verification is completed, the real machine interference data captured by the interferometer can be input into the algorithm for restoration. After going through the same restoration process, a spectral imaging result with higher accuracy is obtained.

Claims

1. An interference spectrum imaging restoration method based on degradation simulation, characterized in that: The following steps are involved: Step 1: Interferometric imaging degradation modeling based on imaging mechanism and calibration data analysis; Step 1.1: Use the interferometric imaging spectrometer to capture calibration data, and perform data sorting and preprocessing; Use an interferometric imaging spectrometer to capture calibration data, including relative calibration data and absolute calibration data; use the classical restoration method and absolute calibration data to obtain uniform spectrum restoration results; The above results are collated for use in the subsequent calibration process; Step 1.2: Construct a calibration solution for the degradation model based on the actual machine calibration data to determine the model parameters; According to the design of the calibration method, the calibration data is matched with the three-stage degradation model, and parameter calibration is performed to disassemble the degradation elements layer by layer; First, the parameters of signal-independent degradation are determined through the black frames in the relative calibration data, and then the photosensitivity degradation parameters are determined through the uniform light data; finally, the instrument degradation parameters are determined through the absolute calibration data and spectral data to obtain the complete degradation model parameters for the instrument; Step 2: Staged interferometric imaging restoration based on deep learning; Step 2.1: Use the calibration results to perform multi-stage training data simulation to generate the algorithm framework training data set; Firstly, the noise-free hyperspectral dataset is preprocessed, including clipping transformation and spectral domain interpolation, to obtain a spectrogram in a specific format; Then, the spectrum and calibration results are substituted into the degradation model, and the imaging process of the instrument is simulated through the calculation of degradation and transformation to generate simulated degradation interference data; According to the algorithm training requirements, the trend term and noise are processed separately. According to whether the simulation data has trend terms and noise, the data simulation process is divided into multiple stages, and multi-stage simulation data sets are generated in sequence. Step 2.2: Embed the denoising network into the multi-stage restoration algorithm framework and train it using the dataset; The algorithm framework has a multi-stage structure, covering the complete process from interference imaging results to ideal spectral restoration results; There are three modules in the framework that need to be embedded in the denoising network, denoted as A, B, and C, and trained; through multi-stage simulation training sets, the denoising networks at different stages are trained respectively; The network is trained separately in different stages. The latter stage uses the intermediate results generated in the previous stage as the training source and uses the simulation data set as the label. After the training is completed, the denoising model is embedded in the algorithm framework to form a multi-stage restoration algorithm. Step 3: Reconstruct a high-quality spectral image from the real interference data captured by the interferometer; The real interference data captured by the interferometer is input into the trained restoration algorithm, and high-quality spectral image results are obtained through restoration calculation.

2. The interference spectrum imaging restoration method based on degradation simulation according to claim 1, characterized in that: In step 1.1, the calibration data is divided into three parts; The first one is the relative calibration data; Relative calibration data is obtained by removing the interferometer optical path in the instrument and only using the uniform light field data taken by the push-scanning sensor in the instrument. The data does not contain interference curves in the interference domain. The relative calibration data includes one image with 0 brightness and several images with different uniform brightness, which are used to extract and calibrate the signal-independent degradation and photosensitivity degradation components respectively. The second is the absolute calibration data. The absolute calibration data is obtained by shooting the same uniform light field with the entire instrument. An interferometer is installed in the instrument. The absolute calibration data contains several interference images with different uniform brightness and no need for 0 brightness. The instrument degradation parameters are extracted by comparing with the relative calibration data. The third is the spectral restoration result from the absolute calibration data; the data is obtained from the absolute calibration data by the classical restoration method and is used to extract the two parameters of the background image and spectral efficiency in instrument degradation; After the above shooting and sorting, the calibration data set is obtained.

3. The interference spectrum imaging restoration method based on degradation simulation according to claim 1, characterized in that: In step 1.2, the degradation model is designed synchronously with the model's own calibration method based on the basic imaging theory and the mining process of calibration data; the degradation model includes instrument degradation, photosensitivity degradation and signal-independent degradation, which correspond to the three stages of the optical signal in interference imaging, namely, the light signal undergoes interference light, is converted into an electrical signal by the photoelectric effect in the sensor, and is gradually converted from an electrical signal into a digital signal in the sensor circuit; In the above three stages, degradation is carried out in a chain-like manner of function compounding, corresponding to the imaging link; all degradation components are defined as noise and trend terms, where noise refers to random terms in the image, and trend terms refer to degradation terms with fixed patterns, which are attached to the source data through operations such as addition / multiplication in different directions; The calibration process of the degradation model requires continuously stripping off the composite degradation function from the calibration data and trying to determine the information of the degradation components; the order of the calibration process is opposite to that of the degradation model, and the signal-independent, photosensitivity and instrument degradation parameters are determined in turn.

4. The interference spectrum imaging restoration method based on degradation simulation according to claim 3, characterized in that: The calibration process uses two methods to calculate the trend term and noise parameters. One is to take the average of similar / related samples; the other is to extract characteristic parameters for noise by fitting the probability distribution.

5. The interference spectrum imaging restoration method based on degradation simulation according to claim 1, characterized in that: In step 2.1, the data simulation generates a degraded interferogram based on the ideal spectral source data without degradation and the model parameters; first, a noise-free multispectral / hyperspectral data set is collected, and the data set is cropped, scaled, and interpolated in the spectral domain to make the data format consistent with the imaging format of the instrument to obtain the source data; Then, the source data and calibration parameters are substituted into the degradation model to generate a simulation data set through calculation; The dataset has a multi-stage structure, which includes four parts: the first is the source data, which is used as the training target of the final restoration network C; the second is the interference data without trend terms and noise, which is used as the training target of network B; the third is the interference data with trend terms and without noise, which is used as the training target of network A; the fourth is the high-fidelity degradation data with both trend terms and noise, which is used as the training source of the entire algorithm; In the model simulation calculation, intermediate results are generated and output as the second part. At the same time, the simulation of the trend term and the noise are distinguished and output as the third and fourth parts respectively.

6. The interference spectrum imaging restoration method based on degradation simulation according to claim 1, characterized in that: In step 2.2, after the degradation data enters the algorithm structure, it is directly input into the denoising network A; A is an interference pattern denoising network, which is used to remove the main noise in the interference pattern degradation process; the training source of A is the degradation data containing all noise and trend term components, and the training target is the degradation data containing only trend terms; Subsequently, the trend items in the data are estimated and eliminated in combination with the calibration data; Afterwards, the denoising network B is used to perform subsequent processing on the above process to suppress the error amplification problem in the operation; The training source of the denoising network B is the intermediate result obtained after the data passes through A and the trend item removal module, and the training target is the interference graph without trend items and noise; After the data passes through the denoising network B, it is directly transformed and efficiently restored in the spectral domain; Similarly, the degraded components remaining in the denoising network B will be amplified in the restoration operation; for this reason, a denoising network C should be introduced after all calculations as post-processing of the entire algorithm; C is a spectral image denoising network, the training source is the intermediate result of the data after passing through the denoising network B and the restoration operation, the training target is the ideal spectral data, and the output of the denoising network C is used as the output result of the entire restoration algorithm.

7. The interference spectrum imaging restoration method based on degradation simulation according to claim 6, characterized in that: The training process of the three-stage network adopts the L1 denoising loss function; let the prediction value be The training objective is , expressed as: in, represents the loss function value of the target network, Represents the total size of the denoising network.

8. An interference spectrum imaging restoration system based on degradation simulation, including denoising network A, denoising network B, denoising network C, trend item removal module, FFT module, and spectral domain efficiency restoration module; Denoising network A is used for preliminary denoising and restoration of data; The trend item removal module is used to estimate and remove the main trend items of the data under the interference mode; The denoising network B is used to post-process the first two stages and suppress the residual and amplified errors; The FFT module is used to perform operations on the data, including Fourier transform, spectral domain adjustment, and interpolation, converting the interference pattern into a spectral pattern mode; The spectrum domain efficiency restoration module is used to process the trend item of data in the spectrum mode; The denoising network C is used to post-process the transformation restoration process, suppress the residual and amplified errors, and enhance the final restoration result; The connection relationship between the above modules is: The output end of the denoising network A is connected to the input end of the trend item removal module; The output end of the trend item removal module is connected to the input end of the denoising network B; The output of the denoising network B is connected to the input of the FFT module; The output end of the FFT module is connected to the input end of the spectrum domain efficiency restoration module; The output end of the spectral domain efficiency restoration module is connected to the input end of the denoising network C.

Images