Infrared spectrum detection chip-oriented adjustable coding and deep learning reconstruction inversion method and system

Abstract

The invention discloses an infrared spectrum detection chip-oriented adjustable coding and deep learning reconstruction inversion method. The method comprises the steps of constructing a spectral response matrix of a target infrared spectrum detector; forming a training set by the real spectrum curve and the light current data corresponding to different voltage channels; constructing an initial model based on the spectral response matrix and a deep learning network framework, wherein the initial model comprises a data preprocessing module, a feature extraction module, a feature fusion module and a prediction module; training the initial model by using the training set to obtain a reconstruction model for reconstructing the spectrum curve; and inputting the light current data collected by the target infrared spectrum detector into the reconstruction model so as to output a spectrum curve corresponding to the light current data. The invention also provides an adjustable coding and deep learning reconstruction inversion system. According to the method provided by the invention, the resolution and the anti-noise capability of spectrum reconstruction of the infrared spectrum detector can be remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of optoelectronic information processing technology, and in particular relates to an adjustable coding and deep learning reconstruction inversion method and system for infrared spectral detection chips. Background Technology

[0002] Traditional infrared spectral reconstruction techniques face three major technical bottlenecks: First, while physical model-based inversion algorithms (such as ridge regression and compressed sensing) possess clear physical interpretability, their reconstruction resolution is insufficient when processing narrow-interval bimodal spectra, making it difficult to effectively resolve fine structures with peak intervals less than 5 nm. In some published patents or literature, although regularization strategies are employed to improve stability, their spectral resolution capabilities are still limited by the bandwidth and non-orthogonality of the response function matrix. Second, while deep learning-based data-driven methods (such as convolutional neural networks and fully connected networks) possess the ability to automatically extract spectral features and tolerate certain nonlinear noise, they typically rely heavily on large-scale, high-quality labeled training data, making them difficult to adapt to real-world scenarios with insufficient training samples or distribution drift. Furthermore, these models are often "black box structures," lacking sufficient physical interpretability and failing to align with the spectral transmission processes in actual spectral systems. Third, existing methods face unique challenges from strong noise interference in the mid-infrared band (3000–5000 nm), such as thermal noise, detector dark current fluctuations, and atmospheric absorption edge effects.

[0003] Traditional reconstruction models (such as wavelet transform + ridge regression) show a significant increase in reconstruction error, with MSE rising by more than 30%, increased spectral peak shift, and easy masking of weak peaks, especially under low-channel sampling conditions. Specifically, while the best-performing solution in the current technology improves noise resistance through wavelet denoising and ridge regression, its bimodal resolution is still limited to 10nm. While other deep learning solutions reduce data requirements, they are not optimized for the strong noise environment unique to the mid-infrared band, limiting their practical applications. Furthermore, the contradiction between high spectral resolution and limited hardware resources severely restricts the development of existing technologies. Especially in high-requirement applications that simultaneously demand 1nm single-peak resolution, better than 5nm bimodal spacing, and strong noise resistance (maintaining reconstruction stability at >60% noise levels), existing algorithms cannot achieve all of these simultaneously.

[0004] Patent document CN103207015A discloses a spectral reconstruction method, including: a filter initialization step, acquiring the transmission spectrum curve of each filter channel on the filter; a spectrum acquisition step, acquiring the initial spectral information of light after passing through the filter; and a spectral reconstruction step, reconstructing the initial spectral information using the transmission spectrum curve through a non-negative matrix full-rank decomposition method to obtain the reconstructed spectral information. This invention achieves the detection of light's spectral information by reconstructing the spectral information acquired by an image sensor.

[0005] Patent document CN120063487A discloses a visible-infrared fusion compressed spectral imaging system and method based on a single pixel. The imaging method includes the following steps: Step 1: Encode the signal of the target to be measured to obtain the encoded infrared signal and the encoded visible light signal of the target to be measured; Step 2: Input the infrared signal of the target to be measured into the infrared imaging optical path to collect the mid-wave infrared spectral intensity value, and input the visible light signal of the target to be measured into the visible light imaging optical path to collect the visible light spectral intensity value; Step 3: Compress and sample the visible light spectral intensity value and the mid-wave infrared spectral intensity value, and perform reconstruction operations to obtain the infrared hyperspectral image and the visible light hyperspectral image of the target to be measured; Step 4: Perform image fusion based on the infrared hyperspectral image and the visible light hyperspectral image to obtain a fused image with rich details. Summary of the Invention

[0006] The purpose of this invention is to provide an adjustable coding and deep learning reconstruction inversion method and system for infrared spectral detector chips, which can significantly improve the resolution and noise resistance of spectral reconstruction of infrared spectral detectors.

[0007] To achieve the first objective of this invention, the following technical solution is provided: a tunable coding and deep learning reconstruction inversion method for infrared spectral detection chips, comprising the following steps: Construct the spectral response matrix of the target infrared spectrometer; The target infrared spectral detector is illuminated with different real spectral curves, and corresponding photocurrent data are collected under different bias voltage channels. The real spectral curves and the photocurrent data are paired to form a spectral dataset, which includes a training set. An initial model is constructed based on the spectral response matrix and the deep learning network framework. The initial model includes a data preprocessing module, a feature extraction module, a feature fusion module, and a prediction module. The data preprocessing module is used to convert the input photocurrent data into a photocurrent response curve, which is in the form of a spectral signal tensor that can be processed by a deep learning model.

[0008] The feature extraction module includes a local feature extractor and a global feature extractor. The local feature extractor is used to extract local features of the waveform in the photocurrent response curve. The global feature extractor is used to perform position encoding on the input photocurrent data and to extract features from the position sequence obtained by position encoding to obtain the corresponding global sequence features. The feature fusion module is used to discretize local features into global sequence features to output the corresponding fused features; The prediction module makes predictions based on the input fusion features and outputs the prediction results. The initial model is trained using the training set to obtain a reconstruction model for reconstructing the spectral curve; The photocurrent data collected by the target infrared spectrometer is input into the reconstruction model to output the spectral curve corresponding to the photocurrent data.

[0009] Specifically, the process of constructing the spectral response matrix is ​​as follows: The spectral response curves corresponding to different bias voltages were acquired using a target infrared spectrometer. All spectral response curves were stacked by wavelength and by voltage channel to construct a response matrix of voltage channel number minus wavelength sampling point number.

[0010] Specifically, the position code is generated using a fixed sine-cosine function, with the following formula: ; ; in, For wavelength position index, For feature dimensions.

[0011] Specifically, the conversion relationship between the spectral data and the photocurrent data is as follows: ; in, Indicates responsiveness. Represents wavelength, For spectral data intensity, and For infrared spectroscopy detectors to detect wavelengths The upper and lower limits, This is photocurrent data.

[0012] Specifically, the local features are automatically extracted from the photocurrent response curve by the deep learning network. These features are implicit representations and are not pre-defined as specific physical parameters.

[0013] Specifically, the global feature extractor outputs a global feature sequence that matches the length of the position sequence by performing layer-by-layer self-attention calculation and residual connection on the position sequence.

[0014] Specifically, during the training process, a multi-index function is used to update the parameters of the initial model.

[0015] Specifically, the multi-index function includes mean square error and spectral matching degree.

[0016] To achieve the second objective of this invention, the following technical solution is provided: an adjustable coding and deep learning reconstruction inversion system for performing the steps of the above-described adjustable coding and deep learning reconstruction inversion method for infrared spectral detection chips, in order to reconstruct a high-resolution spectral curve.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: By recovering or inferring the complete spectral curve of a target in the mid-infrared band from optical response data with limited channels, the reconstructed spectral resolution and noise robustness are significantly improved, effectively overcoming the key bottlenecks of existing technologies in three aspects: "missing spectral information", "limited resolution" and "strong noise interference". Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the adjustable coding and deep learning reconstruction inversion method provided in this embodiment; Figure 2 This is a schematic diagram of the model reconstruction provided in this embodiment; Figure 3 This is a schematic diagram of the reconstruction result provided in this embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, this embodiment provides a tunable coding and deep learning reconstruction inversion method for infrared spectral detection chips, which includes the following steps: The specific steps for constructing the spectral response matrix of the target infrared spectrometer in this embodiment are as follows: Applying different bias voltages to the dynamically tunable filter at the front end of the dynamically tunable infrared focal plane detector, so that it... Different spectral transmittance curves were obtained under each voltage channel.

[0021] The filter has selective transmission capability for infrared light of different wavelengths under different voltages, so that the detector corresponds to a specific set of spectral response curves for each channel.

[0022] By arranging the responsivity curves of all channels according to the wavelength dimension and stacking them according to the voltage channel dimension, a spectrum of size can be constructed. ( Number of voltage channels The response matrix (where the number of wavelength sampling points is the number of sampling points) The response matrix describes the spectral modulation characteristics of the dynamically tunable infrared focal plane.

[0023] The target infrared spectrometer is illuminated with different real spectral curves, and corresponding photocurrent data are collected under different bias voltage channels. The real spectral curves and the photocurrent data are paired to form a spectral dataset, which includes a training set. In this embodiment, the conversion relationship between the spectral curves and the photocurrent data is as follows: ; in, Indicates responsiveness. Represents wavelength, For spectral data intensity, and For infrared spectroscopy detectors to detect wavelengths The upper and lower limits, This is photocurrent data.

[0024] Since the data in the real-world environment is discrete, the integral form above can be transformed into... : ;

[0025] because Where U is voltage, I is current, and r is resistance, it can also be converted to Right now Because some detectors output photovoltage data (instead of photocurrent data) after receiving light signals and processing them through the readout circuit, this deep learning algorithm can also use photovoltage data and the corresponding original spectra under different voltage channels as training sets to train the model, and use the actually measured photovoltage data to reconstruct the spectrum.

[0026] The training set in this embodiment mainly comes from three sources: (1) spectral data and corresponding light response data measured by actual spectrometers; (2) spectral data and corresponding light response data simulated by computer through models; and (3) spectral data in existing public datasets.

[0027] like Figure 2 The diagram illustrates the model reconstruction provided in this embodiment. First, the dual-feature fusion-driven deep learning spectral reconstruction algorithm reads training and testing data from the spectral dataset for model training. This deep learning spectral reconstruction process combines the local feature extraction capability of the residual network with the global dependency modeling capability of the Transformer. The processing of the photoresponse data and spectral data is mainly divided into two paths: one uses the residual network to process local features, and the other uses the Transformer module to process global features, resulting in a hybrid fusion of the dual-path feature output.

[0028] The deep learning spectral reconstruction algorithm driven by dual feature fusion has the following technical advantages: Dual-modal feature fusion: simultaneously captures local features (convolutional path) and global dependencies (Transformer path); Position awareness capability: Wavelength position information is preserved through sine / cosine encoding; Deep optimization: Residual connections ensure stable training of networks with 20+ layers; Computationally efficient: Adaptive pooling replaces fully connected layers, reducing parameters by 80%; Strong generalization: Dropout and weight decay effectively prevent overfitting; Based on the Python and Qt5 frameworks, a graphical user interface application integrating deep learning spectral reconstruction algorithms was developed. The system adopts a modular design concept, consisting of a model training module and a spectral reconstruction module, enabling parameter configuration of the deep learning model, visualization of the training process, and rapid reconstruction of high-resolution spectral data.

[0029] The model includes a data preprocessing module, a feature extraction module, a feature fusion module, and a prediction module. The feature extraction module includes a local feature extractor and a global feature extractor. The local feature extractor consists of four residual blocks. The first residual block does not change the number of channels, the second residual block increases the number of channels from 64 to 128, the third residual block does not change the number of channels, and the fourth residual block further increases the number of channels from 128 to 256. Each of these residual blocks consists of two one-dimensional convolutional layers and a skip connection, which can effectively capture the local features of the input spectrum. The global feature extractor contains self-attention and feedforward neural networks. This feature extractor is composed of six stacked encoder blocks. Each encoder block contains eight self-attention mechanisms and a feedforward neural network. The expansion coefficient of the feedforward neural network is 4. Subsequently, a multi-head self-attention mechanism is used to globally model the input sequence to learn the long-range dependencies between different spectral bands. The global features are implicit representations, including but not limited to the relative relationships between spectral peaks and the overall spectral profile.

[0030] The initial model constructed above is trained using the constructed training set to obtain a reconstruction model for reconstructing the spectral curve.

[0031] In this embodiment, mean squared error (MSE) and spectral matching degree (normalized cross-correlation NCC) are used, where the formula for calculating MSE is: .

[0032] The formula for calculating spectral matching degree (normalized cross-correlation NCC) is as follows: .

[0033] The photocurrent data collected by the dynamically adjustable infrared focal plane detector is input into the reconstruction model to output the spectral curve corresponding to the photocurrent data. In this embodiment, the reconstruction model first projects the m-dimensional input optical response data onto a high-dimensional embedding space of embed_size=256 through a linear mapping, and adds positional encoding based on a sine function to introduce sequence information in the spectral dimension. Subsequently, the Transformer path performs layer-by-layer self-attention calculation and residual connections on the embedded sequence, outputting a global feature sequence matching the sequence length; its overall representation is extracted by average pooling of this sequence. Simultaneously, the original input data is also fed into the residual convolution path, and after four residual blocks and global average pooling, a local feature vector of length 256 is obtained.

[0034] In the feature fusion module, the global features output by the Transformer branch and the local features output by the residual convolution branch are concatenated and fused (forming a fusion vector of length 512), and then the final spectral reconstruction output is completed through two fully connected layers.

[0035] The entire model architecture employs residual connections, layer normalization, and multi-head attention mechanisms to improve training stability and expressive power, constructing a global-local feature complementary fusion infrared spectral reconstruction framework. Compared to traditional models, this structure improves reconstruction accuracy while exhibiting stronger robustness and generalization ability, making it particularly suitable for high-resolution, high-noise reconstruction scenarios in the mid-infrared band.

[0036] This embodiment also provides an adjustable coding and deep learning reconstruction inversion system for performing the steps of the adjustable coding and deep learning reconstruction inversion method for infrared spectral detection chips provided in the above embodiment.

[0037] like Figure 3 As shown, the system provided in the above embodiments is used to perform spectral reconstruction tests on the optical response data of wide-band single-peak, narrow-band single-peak, double-peak, and triple-peak.

[0038] The left image shows the single-peak reconstruction effect, and the right image shows the double-peak reconstruction effect. The full width at half maximum (FWHM) of both single-peak and double-peak spectral reconstructions is 50 nm. In single-peak and double-peak spectral reconstruction scenarios, the algorithm can achieve an MSE of less than 0.001 and a spectral matching degree of over 0.96, demonstrating excellent spectral reconstruction accuracy. For three-peak and complex multi-peak spectral reconstruction tasks, the algorithm still maintains excellent performance with an MSE of <0.002 and a spectral matching degree of over 0.93, demonstrating outstanding adaptability to multi-peak spectral reconstruction.

[0039] Experimental results show that the algorithm not only exhibits high-precision characterization characteristics in single / double-peak spectral reconstruction, but also demonstrates excellent generalization reconstruction performance for three-peak and complex multi-peak systems, effectively verifying its robustness and applicability in different spectral complexity scenarios.

[0040] Spectral reconstruction experiments were conducted on single-peak spectral photoresponse data with a peak height of 1.0 and a full width at half maximum (FWHM) of 100 nm, focusing on overcoming the challenge of resolving single peaks with a center wavelength spacing of only 1 nm. Experimental data show that the algorithm achieves MSE < 0.001 in both single-peak and double-peak reconstruction scenarios, and the reconstructed spectrum closely matches the original spectrum. Its excellent reconstruction accuracy is directly verified, fully demonstrating the algorithm's strong resolving power and high-precision spectral reconstruction performance under ultra-narrow wavelength spacing.

[0041] Spectral reconstruction experiments were conducted on bimodal spectral response data with a peak height of 1.0, a full width at half maximum (FWHM) of 3 nm, and a peak spacing of 5 nm, successfully overcoming the spectral resolution bottleneck of narrow-spacing bimodal peaks. The experiments achieved high-precision reconstruction with a mean squared error (MSE) of 0.000987 and a spectral matching degree of 0.937834. These quantitative results demonstrate that even in the extreme scenario of bimodal peaks with narrow spacing (5 nm) and narrow linewidth (FWHM 3 nm), the algorithm still exhibits excellent spectral reconstruction accuracy and spectral shape reproduction capability, effectively verifying its robustness in high-resolution spectral analysis tasks.

[0042] For a double-peak limiting spectrum with a peak height of 1.0, a full width at half maximum (FWHM) of 3 nm, and a peak spacing of 5 nm, 2%, 5%, 10%, 20%, 40%, and 60% Gaussian noise was injected. Noise-resistant reconstruction experiments were conducted to investigate the algorithm's robustness in resolving the limiting double peak under noisy environments. Data shows that the algorithm can accurately reproduce the spectral characteristics of the limiting double peak under different noise levels, demonstrating excellent stable reconstruction performance. Especially when facing high noise interference levels such as 40% and 60%, which exceed the threshold of typical application scenarios, the algorithm can still effectively extract the double-peak features. These results fully validate its engineering applicability and anti-interference potential in extreme noise environments, providing technical support for spectral analysis in complex real-world scenarios.

[0043] The above description (including the accompanying drawings and specific embodiments) is only for illustrating the technical concept and preferred embodiments of the present invention, and its description should not be used to limit the scope of protection of the present invention. Any equivalent substitutions, modifications or variations made by those skilled in the art based on the technical concept of the present invention, as long as they do not depart from the functional and structural principles of the present invention, should fall within the scope of protection defined by the claims of the present invention.

Claims

1. An adjustable coding and deep learning reconstruction inversion method for an infrared spectral detection chip, characterized in that, The method comprises the following steps: constructing a spectral response matrix of a target infrared spectral detector; irradiating the target infrared spectral detector with different real spectral curves, collecting corresponding photocurrent data under different bias voltage channels, and pairing the real spectral curves with the photocurrent data to form a spectral data set; constructing an initial model based on the spectral response matrix and a deep learning network framework, the initial model comprising a data preprocessing module, a feature extraction module, a feature fusion module, and a prediction module; the data preprocessing module is configured to convert the input photocurrent data into photocurrent response curves; the feature extraction module is configured to extract local features of the photocurrent response curves and global feature sequences of the photocurrent data; the feature fusion module is configured to discretize the local features into global sequence features to output corresponding fusion features; the prediction module is configured to perform prediction according to the input fusion features to output prediction results; training the initial model using the spectral data set to obtain a reconstruction model for reconstructing spectral curves; inputting the photocurrent data collected by the target infrared spectral detector into the reconstruction model to output spectral curves corresponding to the photocurrent data.

2. The adjustable coding and deep learning reconstruction inversion method for the infrared spectrum detection chip according to claim 1, characterized in that, The construction process of the spectral response matrix is as follows: collecting corresponding spectral responsivity curves under different bias voltages by the target infrared spectral detector; stacking all the spectral responsivity curves in the order of wavelength arrangement and voltage channel to construct a response matrix with the number of voltage channels and the number of wavelength sampling points.

3. The adjustable coding and deep learning reconstruction inversion method for the infrared spectrum detection chip according to claim 1, characterized in that, The feature extraction module comprises a local feature extractor and a global feature extractor, the local feature extractor is configured to extract local features of waveforms in the photocurrent response curves, and the global feature extractor is configured to perform position encoding on the input photocurrent data and extract features from a position sequence obtained by position encoding to obtain corresponding global sequence features.

4. The adjustable coding and deep learning reconstruction inversion method for the infrared spectrum detection chip according to claim 1, characterized in that, The position encoding is generated by a fixed sine-cosine function, and the formula is as follows: ； ； wherein, is a wavelength position index, is a feature dimension.

5. The adjustable coding and deep learning reconstruction inversion method for the infrared spectrum detection chip according to claim 4, characterized in that, The conversion relationship between the spectral data and the photocurrent data is as follows: ； wherein, represents responsivity, represents wavelength, is the spectral data intensity, and are the upper and lower limits of the wavelength that can be detected by the infrared spectrum detector, is the photocurrent data.

6. The adjustable coding and deep learning reconstruction inversion method for infrared spectral detection chips according to claim 1, characterized in that, The global feature extractor outputs a global feature sequence matching the length of the position sequence by performing layer-by-layer self-attention calculation and residual connection on the position sequence.

7. The adjustable coding and deep learning reconstruction inversion method for infrared spectral detection chips according to claim 1, characterized in that, In the training process, a multi-index function is used to update the parameters of the initial model.

8. The adjustable coding and deep learning reconstruction inversion method for infrared spectral detection chips according to claim 1, characterized in that, The multi-index function comprises a mean square error and a spectral matching degree.

9. An adjustable coded and deep learning reconstruction inversion system, characterized in that, The method is used to perform the steps of the adjustable coding and deep learning reconstruction inversion method for an infrared spectral detection chip as claimed in any one of claims 1-8 to reconstruct high-resolution spectral curves.

Images