An efficient deep learning approach for automated FTIR spectral interpretation and simultaneous molecular bond localization

Abstract

The accurate identification of functional groups via Fourier Transform Infrared spectroscopy constitutes a fundamental pillar of chemical analysis, yet conventional manual interpretation is constrained by its reliance on expert proficiency and susceptibility to subjective errors. Addressing these limitations, this study proposes an image-based framework utilizing the YOLOv8n object detection architecture for automated localization and classification of eleven representative molecular-bond absorption classes in FTIR spectral plots. By reformulating spectral peak interpretation as a computer vision task, the developed model achieved a mean Average Precision (mAP@0.5) of 0.962 and an overall classification accuracy of 94.6% on the generated validation dataset. Quantitative evaluations further showed strong agreement with the ground-truth generated labels, with Cohen's Kappa and Matthews Correlation Coefficient values of 0.962 and 0.963, respectively. To align the output with standard spectroscopic interpretation, the study integrated a post-processing mechanism that converts bounding box predictions into vertical spectral markers. The revised study additionally reports overlap-focused dataset statistics and a preliminary qualitative application on a real FTIR spectrum of a wound dressing containing Liquidambar orientalis resin, while emphasizing that broader experimental validation is required before routine analytical deployment.