<![CDATA[Street food is a significant part of Lhokseumawe City's culinary sector, yet high inter-class similarity and intra-class variation among food types make manual identification difficult, while architectures such as ResNet50 and VGG16 are computationally too complex for resource-constrained mobile devices. This study designs an Android-based street food identification application using MobileNetV3-Large through transfer learning, and analyzes how unfrozen bottleneck block depth in a Partial Fine-Tuning scenario affects classification performance across 30 street food classes. The model was trained in two phases, Feature Extraction and Partial Fine-Tuning, testing three block-unfreezing scenarios (the last 3, 5, and 7 bottleneck blocks) on 9,000 images split with a 70:15:15 ratio. The 5-block scenario achieved the best performance, with 90.30% accuracy, 90.51% precision, 90.30% recall, and a 90.24% F1-Score, outperforming the 3-block (88.81%) and 7-block (89.93%) scenarios. Unfrozen block depth was found to significantly affect trainable parameters without changing inference-time FLOPs, indicating that deeper fine-tuning does not always improve accuracy. The best model was converted to TensorFlow Lite using Dynamic Range Quantization, producing a 3.37 MB model with average inference times of 164 ms and 692 ms on two Android devices, demonstrating a practical balance between accuracy and efficiency for on-device deployment.Keywords: MobileNetV3-Large, Partial Fine-Tuning, Transfer Learning, Bottleneck Block, Image Classification, Street Food, TensorFlow Lite.]]>
