<![CDATA[Indonesian spice identification is challenging due to high inter-class visual similarity among rhizome species, yet the systematic impact of image preprocessing on classification performance remains underexplored. This study evaluates three preprocessing techniques (CLAHE, Background Removal, and Data Augmentation) arranged in eight factorial schemes across three architectures (VGG16, MobileNetV3-Large, and YOLOv11-cls), comprising 72 training runs on an 856-image four-class dataset, with Macro-F1 on stratified 80/10/10 splits as the primary metric. Data Augmentation alone (scheme S4) produced the most consistent gain, with VGG16 reaching 93.80% Macro-F1. The full pipeline degraded performance across all architectures because CLAHE and Background Removal interact antagonistically, refuting the assumption that preprocessing effects are additive. Externally defined augmentation was also found to be silently bypassed by the Ultralytics framework. All five primary findings replicated on an independent seven-class dataset using 5-fold cross-validation, and Background Removal proved beneficial only for large, high-contrast objects. The study establishes an empirical preprocessing benchmark and practical deployment guidance for spice recognition.]]>
Copyrights © 2026
