An end‐to‐end deep learning approach for Raman spectroscopy classification

Raman spectroscopy has numerous advantages as a means of analyzing materials and is widely used in petrochemical, material, food, biological, medical, and other fields. Its analysis process is fast, nondestructive, and requires no prepreparation. Meanwhile, the research on applying machine learning methods in Raman spectral recognition is becoming increasingly popular. In this study, an end‐to‐end deep learning method called deep residual shrinkage‐VGG (DRS‐VGG) is proposed, which is able to match Raman spectral features with model structure and reduces the reliance on feature engineering. The addition of identity shortcut and soft thresholding in the model eliminates redundant signals to achieve end‐to‐end spectral identification. The effectiveness of the proposed model is verified in three subsets of the RRUFF Raman database and bacterial Raman dataset from different perspectives without data augmentation, and the recognition accuracy is 97.84%, 92.81%, and 95.08%, respectively. Compared with other methods, the proposed DRS‐VGG model achieved a significant improvement in speed or accuracy. The model's understanding of the spectra is visualized by the gradient‐weighted class activation mapping (Grad‐CAM), which explains the excellent classification performance. Additionally, the weight pruning technique is used to achieve model compression and improve recognition accuracy by shrinking the weights and fine‐tuning the biases. In this paper, an end‐to‐end deep learning method called deep residual shrinkage‐VGG (DRS‐VGG) is proposed, which is able to achieve intelligent identification of Raman spectra without preprocessing. The classification results in the publicly available RRUFF Raman and bacteria Raman spectra dataset show that the proposed method is superior in speed and recognition accuracy. Additionally, the gradient‐weighted class activation mapping (Grad‐CAM) and the weight pruning technique are used to improve model interpretability and utility.