Infrared Spectral Deconvolution Algorithm Based on Masked Pre-training Transformer

Nowadays, infrared spectral analysis technology has been widely used in various fields in real life, however, infrared spectral signals are susceptible to degradation and contamination due to the aging of infrared spectrometer equipment, bandwidth overlap, and random noise. In order to improve the quality and reliability of IR spectral signals, denoising and reconstruction are important preprocessing steps. To address these challenges, a masked pre-trained Transformer model is proposed in this paper. To train the model's ability to suppress noise in IR spectra, the encoder randomly masks the IR spectral sequences during pre-training, and the decoder reconstructs the masked inputs, which can greatly enhance the downstream spectral predictor with the learnt hidden representations. The masking strategy makes the learned representation more robust by capturing the dependencies between IR spectral sequences for feature extraction. The experiments are compared with the current state-of-the-art deconvolution technique, and the experimental results show that the method has excellent deconvolution performance and is able to effectively recover the texture details of the infrared spectrum.

Keywords: Infrared Spectroscopy; spectral deconvolution; Masking strategy, Transformer.