Implementing Medicinal Leaf Species Classification Using Hybrid Preprocessing, CNN Benchmarking, Squeeze-and-Excitation Attention

Abstract

Accurate medicinal-plant identification is important for botanical research and herbal quality assurance, yet visual similarity among species makes manual identification difficult. This study evaluates an auditable medicinal-leaf classification workflow rather than proposing a new transfer-learning or attention architecture. The workflow combines file-integrity auditing, fixed stratified partitions, controlled augmentation, three model configurations, held-out evaluation, persistent artifacts, and qualitative Grad-CAM inspection. The controlled dataset contains 1,009 decodable images from nine species, partitioned into 706 training, 151 validation, and 152 held-out test images. SHA-256 comparison found no byte-identical duplicates; perceptual or specimen-level near-duplicate screening was not performed and is therefore a limitation. Under the same three-epoch CPU pilot protocol, ImageNet-initialized MobileNetV3-Small achieved 98.68% test accuracy and 98.68% weighted F1-score, compared with 71.71% for a randomly initialized baseline CNN and 62.50% for a segmentation-plus-CLAHE SE-CNN. Exact paired analysis showed that MobileNetV3-Small outperformed the baseline (McNemar p < 0.001); the advantage over the hybrid model also remained significant under every error-overlap pattern compatible with the reported confusion matrices. Class-wise metrics exposed severe hybrid-model failures for Coriander and Betel. Because the component changes were confounded, training was limited to three epochs, Grad-CAM was inspected only qualitatively, and no external dataset was used, the results support MobileNetV3-Small only as the strongest model for this controlled split and do not establish field generalization or superiority of the hybrid branch.