Design Of Signal Processing Algorithm For Micro-spectrometer Based On Deep Learning

With the enhancement of computer processing power and the decrease of microprocessor cost,the computational micro-spectrometer has attracted more and more attention from researchers due to its advantages,such as its small size,low production cost,and high real-time efficiency.Computational micro-spectrometers can directly read encoding information from detectors and perform calculations to obtain the spectral information of substances.Although this type of spectrometer has many advantages,there are still some problems,mainly including: 1)In wide-band and complex spectral reconstruction tasks,traditional spectral reconstruction algorithms need more accuracy.This is because the data collected by detectors are limited in quantity and relatively sparse,which can lead to the loss of some important spectral features in the reconstructed material spectra by traditional spectral reconstruction algorithms,thus affecting spectral reconstruction accuracy;2)The spectral reconstruction performance is unstable.Interference factors such as manufacturing errors and random noise may exist in the production and use of spectrometers,which can lead to a decrease in the spectral reconstruction performance and affect the application effect;3)The insufficient high-quality training samples lead to a decrease in the accuracy of reconstructed spectra.Collecting high-quality training data is costly,and reducing the number of training samples can result in a drastic drop in spectral reconstruction performance,thus affecting the accuracy of reconstructed spectra.To address the above issues,this paper focuses on high-precision spectral reconstruction tasks research.The main work and innovations are as follows:(1)This paper simulated a mid-infrared filter detector array covering the wavelength range of2.58)to 58),and implemented spectrum reconstruction using a deep learning network model.This approach effectively solves the problem of traditional spectral reconstruction algorithms losing spectral feature information in wide-band spectrum reconstruction tasks and exhibits good performance in terms of spectral reconstruction accuracy.(2)To improve the accuracy and stability of reconstructed spectra,this paper introduces the super-resolution theory for the first time and designs a super-resolution spectral reconstruction model.The model adopts residual network modules to extract and learn spectral features and uses linear interpolation algorithm to augment the obtained spectral feature data.This model not only significantly improves the accuracy of the spectral reconstruction,but also maintains good spectral reconstruction performance in the face of possible manufacturing errors and noise in practical engineering,demonstrating a certain degree of stability.(3)To further improve the performance of spectral reconstruction in practical engineering under manufacturing errors and noise,this paper introduces the meta-learning framework based on the super-resolution spectral reconstruction model.The model uses a hyper-network structure in the meta-learning framework and designs a weight prediction network,which predicts non-shared weight parameters for the spectral reconstruction tasks of different substances,extracts more abundant feature information,and further improves the accuracy and noise resistance of the spectral reconstruction model.In addition,this paper uses a small amount of high-quality training data to continue optimizing model parameters to enhance the model’s ability to deal with hardware manufacturing errors.(4)To address the issue of reconstruction artifacts in the spectral reconstruction model under high noise,this paper introduces a spectral optimization model based on the diffusion model on the spectral reconstruction model.The model consists of diffusion and inverse diffusion processes.In the diffusion process,the real spectrum of the material is gradually noised to simulate the reconstructed spectrum under noise interference.In the reverse diffusion process,a deep learning-based reverse diffusion model is used to generate each small step of the diffusion process in reverse until the spectrum is recovered.Experimental results show that the model can remove reconstruction artifacts while preserving spectral features and has good spectral optimization capabilities.(5)To verify the practicality of the spectral reconstruction model and the spectral optimization model,this paper deployed the models on the Jetson Xavier NX device based on the ONNX Runtime inference framework.Through real-time analysis and spectral reconstruction results,it was confirmed that the models have high feasibility on embedded devices.In summary,the spectral reconstruction model and spectral optimization model designed in this paper can effectively address the problems of computational micro-spectrometers,and accurately reconstruct material spectra.Especially,the meta-learning-based super-resolution spectral reconstruction model performs well not only in the reconstruction of three types of substances in this experiment,but also in a wider range of material spectral reconstruction,demonstrating its good generalization ability.It also exhibits better robustness than other models in the face of noise interference,manufacturing errors,limited training samples,and reduced filter detector units.The spectral reconstruction model combined with the spectral optimization model based on the diffusion model,it can handle different levels of noise and more perfectly complete the spectral reconstruction tasks.Moreover,this paper verifies the feasibility and real-time performance of the spectral reconstruction models on real hardware,laying a foundation for its practical engineering applications.