A neural network model-based super-resolution spectral reconstruction method

Abstract

The application discloses a kind of neural network model-based super-resolution spectrum reconstruction method, the neural network of the full convolutional coding-decoding structure based on the pytorch library of python is built, network architecture and loss function are optimized for one-dimensional spectral data, can recover from undersampling spectral data into super-resolution spectrum, can reduce 80% time required for traditional super-resolution spectrum reconstruction, the model can be directly embedded in existing various super-resolution optical systems, while guaranteeing reconstruction accuracy and reliable spectral details, the algorithm structure is compact and efficient, and the calculation performance consumption is small, can be quickly run on any computer.

Description

Technical Field

[0001] This invention relates to a super-resolution spectral reconstruction method based on a neural network model, belonging to the field of optoelectronic technology. Background Technology

[0002] In optics, spectroscopy is a crucial means of characterizing matter. Spectral detection is limited by the resolution of the spectrometer, unable to overcome the diffraction limit, thus restricting improvements in spectral resolution. Researchers have proposed using random lasers as light sources, where the wavelengths of spectral peaks vary randomly over time, satisfying the sparsity and ergodicity of sparse sampling. Super-resolution spectral reconstruction can be achieved through sparse sampling in the frequency domain. However, reconstructing a single spectrum using this method typically requires more than ten minutes, making it unsuitable for fields such as biosensing and cell detection, which demand the capture of transient processes. Furthermore, prolonged light exposure of biological samples can cause significant phototoxicity, severely limiting the practical application and development of super-resolution spectroscopy.

[0003] Neural network models have demonstrated significant advantages in single-molecule localization microscopy, greatly improving the temporal resolution and reconstruction efficiency of super-resolution imaging. Previous studies have shown that convolutional neural networks used in two-dimensional image processing can be transferred to one-dimensional frequency domain signal processing while maintaining excellent performance. Therefore, it is necessary to explore the application of neural network models in super-resolution spectral imaging to improve reconstruction efficiency and robustness. Summary of the Invention

[0004] This invention proposes a super-resolution spectral reconstruction model based on a neural network model: the Sparse Net-SR (Sparse Network Super-Resolution Spectral Reconstruction Model). This model employs a convolutional neural network architecture, which can recover super-resolution spectra from undersampled spectral data, significantly reducing the number of sparse spectral frames required for super-resolution reconstruction, thereby greatly shortening the reconstruction time.

[0005] The super-resolution spectral reconstruction scheme based on Sparse Net-SR in this invention is as follows:

[0006] A super-resolution spectral reconstruction model based on a neural network model (Sparse Net-SR) is implemented by a convolutional neural network with an encoder-decoder structure, including two downsampling stages in the encoder, two upsampling stages in the decoder (1.3), and 500,000 trainable parameters.

[0007] The model takes a small amount of undersampled spectrum as input (1.2). The encoder compresses the sparse input signal into a feature vector containing global feature information by downsampling through layer-by-layer convolution. The decoder gradually recovers the high-resolution spectral output through transposed convolution (1.4).

[0008] The model is optimized based on the characteristics of one-dimensional spectral data, and the skip connection layer of the traditional super-resolution image prediction U-Net model is eliminated, reducing the computational complexity without affecting the reconstruction of high-resolution signals.

[0009] During training, the neural network model is optimized using a composite loss function. This loss function consists of two parts: the mean absolute error loss (L1) is used to calculate the pixel-level absolute difference between the reconstructed high-resolution spectrum and the true high-resolution spectrum. L1 is robust to outliers, promoting sparse reconstruction and thus accurately recovering sharp peaks in the spectrum while avoiding over-smoothing or blurring; the multi-scale structural similarity index (SSIM) loss term is used to quantify the structural similarity between the reconstructed spectrum and the true spectrum at multiple scales, constraining the model to preserve structural features such as wavelength correlation and peak-valley distribution.

[0010] The training dataset for the model is obtained by pairing undersampled spectra and high-resolution spectra. Different levels of Poisson noise are applied to the high-resolution spectra as the model output. Undersampled spectra are obtained as input by randomly selecting a portion of the data points.

[0011] Compared to traditional super-resolution spectral reconstruction techniques, the number of sparse spectra required for acquisition is reduced by 80%, the acquisition time is reduced by 80%, and the efficiency of super-resolution spectral reconstruction is significantly improved.

[0012] 1. Construct a simulated dataset containing undersampled and super-resolution spectra for model training and validation.

[0013] 2. Design a symmetric encoder-decoder convolutional neural network structure. The encoder performs layer-by-layer convolution and downsampling on the input sparse spectral signal, and finally maps it into a feature vector containing global feature information. The decoder gradually recovers the high-resolution spectral output through transpose convolution.

[0014] 3. Sparse Net-SR is trained using simulated data, with random undersampled spectra as input and corresponding standard spectra as targets, to optimize network parameters. The L1 norm combined with the multi-scale structural similarity index (L1-SSIM) is used as the loss function to replace the traditional mean squared error (training flowchart shown below). Figure 1 ).

[0015] 4. By inputting the sparse spectra collected in actual experiments into the trained Sparse Net-SR model, fast and high-quality super-resolution spectral reconstruction can be achieved (reconstruction results are as follows). Figure 3 ).

[0016] Advantages and features of this invention:

[0017] 1. It can reduce the time required for traditional super-resolution spectral reconstruction by 80%, significantly improve data processing and imaging speed, and greatly enhance reconstruction efficiency.

[0018] 2. It can be directly embedded into various existing super-resolution optical systems without additional modifications to the optical path, offering strong compatibility, easy deployment, and convenient system integration.

[0019] 3. The network structure has been specifically optimized for spectral super-resolution tasks, significantly reducing the number of parameters and lowering the requirements for computing hardware, making it suitable for conventionally configured computers.

[0020] 4. The loss function is optimized for the characteristics of one-dimensional spectral data, which effectively improves the restoration accuracy and reliability of spectral details while ensuring reconstruction accuracy. Attached Figure Description

[0021] Figure 1 This is a flowchart of the training process for the Sparse Net-SR convolutional neural network.

[0022] Wherein, 1.1: Standard high-resolution spectrum, 1.2: Undersampled spectrum, 1.3: Sparse Net-SR model, 1.4: Predicted spectrum, 1.5: Loss function;

[0023] Figure 2 These are the original spectra output under different sampling ratios, and their corresponding super-resolution spectra output after passing through SparseNet-SR.

[0024] The top row shows the undersampled original spectrum, and the bottom row shows the super-resolution spectrum predicted by the neural network; where 2.1: sampling ratio 2%, 2.2: sampling ratio 8%, and 2.3: sampling ratio 20%;

[0025] Figure 3 This is a graph showing the relationship between the reconstructed image quality and the sampling ratio for two different processing methods;

[0026] Among them, 3.1: the SSIM values ​​of the original and standard spectra of sparse sampling change with the sampling ratio, and 3.2: the SSIM values ​​of the predicted and standard spectra based on Sparse Net-SR change with the sampling ratio. Detailed Implementation

[0027] Example 1: A super-resolution spectral reconstruction technique based on a neural network model, comprising the following steps:

[0028] 1. A high-resolution reference spectrum is reconstructed by acquiring low-resolution spectral sequences. To enhance model robustness, different levels of Poisson noise are applied to this high-resolution spectrum, generating 500 sets of noisy standard spectra as training targets. Undersampled spectra are constructed by randomly selecting a portion of spectral data points (sampling ratios set to 10%, 20%, ..., 90%, and 100%, respectively), and used as network input to build a simulated dataset for training and testing.

[0029] 2. Based on a one-dimensional convolutional neural network, a symmetrical encoder-decoder structure is used to construct Sparse Net-SR, with a total of approximately 500,000 parameters. The specific components are as follows:

[0030] Encoder (downsampling path):

[0031] First layer: kernel size 3, input channel 1, output channel 64, stride 1, activation function is ReLU; max pooling layer: kernel size 2, stride 2;

[0032] Second layer: kernel size 3, input channels 64, output channels 128, stride 1, activation function is ReLU; max pooling layer: kernel size 2, stride 2.

[0033] Decoder (upsampling path):

[0034] First layer: transposed convolution kernel size 3, input channels 128, output channels 64, stride 2, activation function ReLU;

[0035] Second layer: transposed convolution kernel size 3, input channels 64, output channels 1, stride 2.

[0036] This network omits skip connections from the traditional U-Net image super-resolution neural network, effectively reducing computational complexity while maintaining reconstruction quality.

[0037] 3. Sparse Net-SR is trained using 500 sets of undersampled spectra as input and the corresponding standard spectra as targets. The loss function combines L1 loss and structural similarity (SSIM) loss, specifically defined as:

[0038]

[0039] in, Mean absolute error:

[0040]

[0041] here Represents the standard spectrum. To predict the spectrum for the network, The number of spectral data points. and These are the intensity values ​​for the corresponding locations.

[0042] Based on structural similarity index:

[0043]

[0044] The calculation formula is:

[0045]

[0046] In the formula , They are respectively and The local mean, , For variance, Covariance. Constant. , , The dynamic range of the input data (the difference between the maximum and minimum values). This multi-scale similarity measure effectively maintains the overall consistency of the spectral structure.

[0047] 4. Load the trained Sparse Net-SR model into the computer, input the sparse, undersampled spectral data into the one-dimensional convolutional network model, and obtain the corresponding high-resolution spectrum through the model's pre-prediction output (spectral reconstruction effect as shown). Figure 2 (High-precision reconstruction is achieved when the sampling ratio is 20%), and the entire process can be completed within 1 second.

Claims

1. A super-resolution spectral reconstruction method based on a neural network model, characterized in that, Includes the following steps: Construct a simulation dataset that contains undersampled spectra and their corresponding high-resolution spectra; A one-dimensional convolutional neural network model is constructed, which adopts an encoder-decoder structure to reconstruct a high-resolution spectrum from the input undersampled spectrum; The neural network model is trained using the simulated dataset to optimize the network parameters; The undersampled spectra collected are input into the trained neural network model, and the reconstructed super-resolution spectra are output.

2. The super-resolution spectral reconstruction method according to claim 1, characterized in that, The encoder includes two downsampling stages, each consisting of a convolutional layer and a pooling layer, for compressing the input sparse spectral signal into a feature vector; the decoder includes two upsampling stages, each consisting of a transposed convolutional layer, for recovering the feature vector into a high-resolution spectrum.

3. The super-resolution spectral reconstruction method according to claim 1, characterized in that, The neural network model does not contain skip connection structures to reduce computational complexity while maintaining reconstruction accuracy.

4. The super-resolution spectral reconstruction method according to claim 1, characterized in that, The loss function used in the training process is a weighted combination of L1 loss and multi-scale structural similarity index loss, which is used to simultaneously optimize the pixel-level accuracy and structural consistency of the spectrum.

5. The super-resolution spectral reconstruction method according to claim 4, characterized in that, The L1 loss is used to calculate the absolute error between the reconstructed spectrum and the standard spectrum, and the multi-scale structural similarity loss is used to evaluate the structural similarity between the reconstructed spectrum and the standard spectrum at multiple scales.

6. The super-resolution spectral reconstruction method according to claim 1, characterized in that, The method for constructing the simulated dataset includes: applying Poisson noise of different intensities to the high-resolution spectrum to generate multiple sets of standard spectra; and generating corresponding undersampled spectra through random sampling to form training sample pairs.

7. The super-resolution spectral reconstruction method according to claim 1, characterized in that, The neural network model has approximately 500,000 parameters and is suitable for operation on conventional computing devices.

8. The super-resolution spectral reconstruction method according to claim 1, characterized in that, The method can reduce the number of sparse spectral acquisitions required for super-resolution spectral reconstruction by 80% and shorten the reconstruction time to less than 1 second.

9. A super-resolution spectral reconstruction system based on a neural network model, characterized in that, include: The data acquisition module is used to acquire undersampled spectral data; The data processing module loads the neural network model trained by the method described in any one of claims 1 to 8; The output module is used to display or store the reconstructed high-resolution spectrum.

10. A computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the super-resolution spectral reconstruction method as described in any one of claims 1 to 8.

Images