Contrastive Learning-Based Standard-Free Calibration Transfer for Near-Infrared Spectroscopy

This paper proposes a standard-free calibration transfer method for near-infrared (NIR) spectroscopy based on contrastive learning, aiming to address the challenge of calibration transfer caused by distribution shifts in spectral data across different instruments. We adapt and modify the original Contrastive Unpaired Translation (CUT) framework, initially designed for image domains, to accommodate one-dimensional spectral data, leveraging its strengths in one-way transfer and unsupervised learning. Experiments are conducted on a cross-device tobacco NIR dataset collected from two instruments. Quantitative results demonstrate that the proposed method achieves superior performance across multiple evaluation metrics, including Pearson correlation coefficient (PCC), cosine similarity (CS), mean absolute percentage error (MAPE) and root mean square error (RMSE). The proposed method outperforms existing standard-free methods such as Multiplicative Signal Correction (MSC) and Finite Impulse Response (FIR) filtering. Moreover, its performance, when compared to these methods, is closer to that of the existing standard calibration transfer method, Direct Standardization (DS). The method requires only standard-free data, significantly reducing the cost and complexity of calibration transfer in industrial applications where standard samples are scarce. These results highlight the potential of contrastive learning as a scalable and effective solution for real-world calibration transfer tasks.