Analysis of paper types based on three dimensional fluorescence spectroscopy combined with Resnet34

Abstract

Printing paper represents one of the most prevalent forms of physical evidence in document forensics, where accurate brand and model identification provides critical investigative leads. To enable rapid, precise identification of commercial printing paper brands, we propose a novel method combining 3D fluorescence spectroscopy with an enhanced ResNet34 network. First, 3D fluorescence contour maps of diverse paper brands were acquired across excitation (280–420 nm) and emission (300–592 nm) wavelengths. These data were augmented via random flipping, scaling, and cropping to generate an expanded dataset of 6398 samples. Subsequently, the ResNet34 backbone was streamlined by removing redundant intermediate layers to improve efficiency. Feature extraction capabilities—particularly for central regions of fluorescence contour images—were strengthened by integrating the CBAM attention mechanism, with training dynamics visualized for optimization. Comparative experiments identified optimal training strategies and hyperparameters. The highest-performing model achieved 97.27% accuracy on the test set, significantly outperforming conventional methods. The proposed system demonstrates strong robustness with a per-image inference time of 0.82 seconds, confirming its practical utility for forensic paper analysis.