Over the years, research in the field of cultural heritage preservation and document analysis has exponentially grown. In this study, we propose an advanced approach for non-destructive estimation of paper fibers using macro images. Expanding on studies that implemented EfficientNet-B0, we explore the effectiveness of six other deep learning networks, including DenseNet-201, DarkNet-53, Inception-v3, Xception, Inception-ResNet-v2, and NASNet-Large, in conjunction with enlarged patch sizes. We experimentally classified three types of paper fibers, namely, kozo, mitsumata, and gampi. During the experiments, patch sizes of 500, 750, and 1000 pixels were evaluated and their impact on classification accuracy was analyzed. The experiments demonstrated that Inception-ResNet-v2 with 1000-pixel patches achieved the highest patch classification accuracy of 82.7%, whereas Xception with 750-pixel patches exhibited the best macro-image-based fiber estimation performance at 84.9%. Additionally, we assessed the efficacy of the method for images containing text, observing consistent improvements in the case of larger patch sizes. However, limitations exist in background patch availability for text-heavy images. This comprehensive evaluation of network architectures and patch sizes can significantly advance the field of non-destructive paper analysis, offering valuable insights into future developments in historical document examination and conservation science.

A few unintended typos were introduced after the proofreading stage, and the authors hence wish to make the following corrections to this paper

The structural analysis of paper fibers is vital for the noninvasive classification and conservation of traditional handmade paper in cultural heritage. Although digital still cameras (DSCs) offer a low-cost and noninvasive imaging solution, their inferior image quality compared to white-light confocal microscopy (WCM) limits their effectiveness in fiber classification. To address this modality gap, we propose an unpaired image-to-image translation approach using cycle-consistent adversarial networks (CycleGANs). Our study targets a multifiber setting involving kozo, mitsumata, and gampi, using publicly available domain-specific datasets. Generated WCM-style images were quantitatively evaluated using peak signal-to-noise ratio, structural similarity index measure, mean absolute error, and Fréchet inception distance, achieving 8.24 dB, 0.28, 172.50, and 197.39, respectively. Classification performance was tested using EfficientNet-B0 and Inception-ResNet-v2, with F1-scores reaching 94.66% and 98.61%, respectively, approaching the performance of real WCM images (99.50% and 98.86%) and surpassing previous results obtained directly from DSC inputs (80.76% and 84.19%). Furthermore, Grad-CAM visualization confirmed that the translated images retained class-discriminative features aligned with those of the actual WCM inputs. Thus, the proposed CycleGAN-based image conversion effectively bridges the modality gap, enabling DSC images to approximate WCM characteristics and support high-accuracy paper fiber classification, which is a practical alternative for noninvasive material analysis.

Accurate paper fiber identification is crucial for cultural heritage conservation. To address the destructive nature of traditional staining and the “black-box” limitations of macroscopic AI models, this study explores the feasibility of a non-destructive testing paradigm using micro-hyperspectral imaging (Micro-HSI). Three traditional Japanese pure bast fibers (Kozo, Mitsumata, and Gampi) were analyzed as standard samples. Raw relative reflectance spectra from microscopic regions of the fibers were extracted via Micro-HSI. Dynamic normalization and Savitzky–Golay first-derivative filtering were applied to suppress scattering and baseline drift. Principal component analysis (PCA) and linear discriminant analysis (LDA) were subsequently employed for dimensionality reduction and supervised classification. The results showed that while unsupervised PCA suffered from inter-class overlap due to shared cellulose-dominated structures, supervised LDA amplified weak chemical fingerprint differences, achieving complete class separation of the highly similar fibers. Analysis of the feature loadings confirmed that the classification relies on the visible-range reflectance baseline, lignin π→π∗ transition absorption (400–450 nm), and near-infrared O-H and C-H overtone vibrations (~835 nm). This proof-of-concept study demonstrates that combining Micro-HSI with chemometrics enables high-precision, non-destructive fiber separation while retaining rigorous physicochemical interpretability, thereby providing an optical reference baseline for future historical paper analysis.

Accurate paper fiber identification is essential for cultural heritage conservation. Traditional staining methods are destructive, while macroscopic AI models often lack physicochemical interpretability. This study explores the feasibility of a non-destructive analytical approach using micro-hyperspectral imaging (Micro-HSI) to overcome both limitations. Three traditional Japanese bast fibers, Kozo, Mitsumata, and Gampi, were analyzed as standard reference samples. Relative reflectance spectra were extracted from microscopic fiber regions using Micro-HSI. Dynamic normalization and Savitzky–Golay first-derivative filtering were applied to suppress scattering effects and baseline drift. Principal component analysis (PCA) and linear discriminant analysis (LDA) were applied in parallel for dimensionality reduction and supervised classification, respectively. The results indicated that unsupervised PCA exhibited substantial inter-class overlap because of the shared cellulose matrix among the fiber types. In contrast, supervised LDA amplified subtle chemical differences and achieved clear separation among the three fibers. Feature-loading analysis indicated that the classification was mainly associated with visible range reflectance characteristics, lignin π→π* absorption bands in the 400–450 nm region, and near-infrared O−H and C−H overtone vibrations near 835 nm. Leave-One-Specimen-Out Cross-Validation yielded an overall accuracy of 77.8%, with error-free classification of Kozo (F1 = 1.00) and misclassification limited to the chemically similar Gampi and Mitsumata pair. This proof-of-concept study demonstrates that combining Micro-HSI with chemometric analysis enables non-destructive fiber discrimination while retaining physicochemically interpretable spectral features. The findings also establish a microscopic spectral reference framework for future non-destructive analysis of historical paper materials.