<![CDATA[Purpose: This study evaluates the accuracy, calibration, efficiency, and deployment trade-offs of EfficientNet-Lite0 and MobileNetV3-Small for 114-class Lontara script recognition.Methods: A public Kaggle dataset containing 11,483 images was divided into training, validation, and test subsets using a stratified 70:15:15 split. Both transfer-learning pipelines were assessed using accuracy, macro and weighted metrics, confusion matrices, Expected Calibration Error, confidence-based rejection, McNemar’s exact test, paired bootstrap confidence intervals, TensorFlow Lite conversion, post-training quantization, and preliminary CPU inference diagnostics. Findings: EfficientNet-Lite0 achieved 94.31% test accuracy and a macro F1-score of 0.9415, compared with 90.60% and 0.9023 for MobileNetV3-Small. The paired difference in correctness was statistically significant (p=1.5893 ×10^(-6)), and the bootstrapped accuracy difference was 3.71 percentage points with a 95% confidence interval of 2.21 to 5.22. EfficientNet-Lite0 showed a lower observed calibration error, although the between-model ECE difference was not conclusive. MobileNetV3-Small produced substantially smaller and faster float16 TensorFlow Lite models, while several visually similar classes remained vulnerable to systematic high-confidence errors.Research implications: Model selection should reflect whether recognition quality or computational efficiency is prioritized. Results are limited to one dataset split, one reported training run, and preliminary CPU diagnostics.Originality: The study integrates statistical comparison, calibration, selective prediction, quantization, deployment diagnostics, and class-level error analysis within one Lontara recognition experiment.]]>
Copyrights © 2026
