<![CDATA[Instant verification of halal status for food ingredients remains a significant challenge for Muslim consumers, particularly when facing complex compositions and foreign-language labels without internet access. This study aims to develop an offline, real-time mobile application for halal ingredient classification by implementing lightweight deep learning models on edge devices. The proposed system integrates Google ML Kit for text extraction and evaluates five neural network architectures: FastText, 1D-CNN, LSTM, GRU, and Multi-CNN trained on a balanced, bilingual (Indonesian–English) dataset comprising 123,449 ingredient samples. Experimental results demonstrate that the 1D-CNN model was the optimal choice, achieving an average accuracy of 99.72%, an average inference time of 60.99 ms, and a compact model size of 719.2 KB. Although the GRU model achieved marginally higher accuracy, the 1D-CNN provided a superior efficiency trade-off with a 2.6× faster training time and lower computational complexity. The primary scientific contribution of this research lies in demonstrating that lightweight 1D-CNN architectures can deliver high-precision, fully offline text-based classification on mobile edge devices, offering a robust alternative to latency-constrained cloud-based verification systems.]]>
Copyrights © 2026
