Non-Targeted Electronic-Nose Fingerprinting and Microscopic Imaging for Leather Classification via Transformer-Based Multimodal Fusion

The rapid, non-destructive screening of leather authenticity requires analytical measurements that capture complementary volatile-response and microstructural information. This study developed RLFormer, a multimodal workflow combining cross-responsive metal-oxide-semiconductor electronic-nose (E-nose) fingerprints with microscopic surface images. Four classes—synthetic, pig, sheep, and cattle leather—were represented by 100 specimens each and measured during a standardized 10-day campaign. The relative standard deviations of the category-level daily mean E-nose responses ranged from 1.4 to 13.3% (global mean, 6.18%), and a variance-component analysis revealed a progressive between-day response attenuation common to all categories. A ResNet50 visual encoder and a two-layer long short-term memory (LSTM) encoder generated modality-specific representations that were integrated by a transformer encoder. With days 1–7 used for training and days 8–10 for testing, RLFormer achieved an accuracy of 97.67±2.16% and a Cohen’s kappa of 0.9689±0.0288 over five random seeds, compared with an 87.50±2.50% accuracy for direct ResNet50–LSTM fusion. The results demonstrate the value of combining non-targeted volatile-response patterns with microscopic texture for rapid leather classification. The workflow is intended for screening rather than compound-specific identification and requires independent validation before deployment.