<![CDATA[Fine-grained recognition of traditional Indonesian cuisine poses a significant challenge for computer vision systems. The primary problem stems from high inter-class texture and color similarities among regional dishes, coupled with severe model performance degradation when training is constrained by limited dataset size. To address these limitations, this study implements a structured machine learning pipeline comprising digital image pre-processing (resizing to 224 x 224, normalization, and geometric data augmentation), a two-stage transfer learning protocol, and an automated recipe retrieval system. Three distinct Convolutional Neural Network (CNN) architectures—ResNet-50 (deep residual learning), MobileNetV3-Large (lightweight attention-guided depthwise convolutions), and EfficientNet-B0 (compound scaling)—were evaluated to classify five traditional Indonesian dishes (Ayam Betutu, Coto Makassar, Gudeg, Papeda, and Rendang). A two-stage transfer learning protocol was implemented across two experimental data partitioning schemes: Scenario A (80:10:10) and Scenario B (60:20:20). Experimental results show that MobileNetV3-Large performed best under optimal data availability (Scenario A), achieving a 95.65% test accuracy. However, under a 25% training data reduction (Scenario B), ResNet-50 demonstrated superior structural resilience, surging to a top test accuracy of 97.83%, whereas MobileNetV3-Large and EfficientNet-B0 dropped to 91.30% and 89.13%, respectively. Confusion matrix analysis revealed that primary misclassifications stemmed from shared liquid broth backgrounds and shredded protein textures between Gudeg and Ayam Betutu. Finally, real-world inference testing on unobserved images confirmed practical deployment viability, successfully mapping predicted class labels and confidence scores (ranging from 30.41% to 69.57%) to corresponding recipe database entries.]]>
Copyrights © 2026
