Position identification of stacked flue-cured tobacco leaves driven by adaxial-abaxial feature decoupling

This is the authors' abstract. We don't add key points for this paper.

Accurate identification of stalk position is a prerequisite for quality grading in flue-cured tobacco. However, the random interleaving of adaxial and abaxial surfaces in stacked scenarios inevitably leads to "feature neutralization," where conflicting optical characteristics severely compromise data separability. To address this bottleneck, this study proposes an unsupervised segmentation strategy based on variance-weighted channel fusion to resolve complex stacked images into biologically distinct regions. Without relying on large-scale annotated datasets, the proposed framework achieved high segmentation robustness (Intersection over Union > 0.85) across diverse lighting conditions and production origins. Furthermore, a rigorous comparative analysis was conducted to quantify the contribution of surface-specific features to classification performance, revealing that texture feature serves as the dominant discriminator. Consequently, the Random Forest model constructed using decoupled adaxial-abaxial features achieved a classification accuracy of 89.5 %, significantly outperforming the conventional whole-leaf mixed approach (72.9 %). These findings confirm that adaxial-abaxial feature decoupling is a critical methodological paradigm for enhancing the precision and objectivity of intelligent grading systems in complex industrial environments.