<![CDATA[Tempe is a traditional Indonesian fermented soybean product whose quality is highly influenced by environmental conditions. Manual quality assessment remains subjective and inconsistent, creating the need for an automated and objective real-time detection system. This study proposes a quantization-based optimization of the YOLOv8m model using TensorFlow Lite (TFLite) for edge-based tempe quality recognition deployed on a Raspberry Pi 4 device. A dataset of 1,000 images categorized into good and defective tempe was used for model training and evaluation. The methodological novelty of this work lies in the explicit benchmarking of FP16 and INT8 post-training quantization strategies under identical hardware conditions, analyzing their impact on inference time, model size, memory usage, and detection accuracy. Experimental results show that compared to the baseline YOLOv8 FP32 model, the INT8-quantized model reduced inference time from 420 ms/frame to 120 ms/frame (71.4% faster), decreased model size from 85.6 MB to 24.9 MB (70.9% smaller), and reduced RAM usage from 780 MB to 410 MB (47.4% lower), while maintaining high detection performance with mAP@0.5 of 0.935 (1.89% accuracy drop). The FP16 model achieved intermediate efficiency gains with minimal accuracy degradation. The optimized models were integrated into a Flask-based web monitoring system, achieving stable real-time visualization with consistent frame updates under edge deployment constraints. These findings demonstrate that INT8 quantization provides the best efficiency–accuracy trade-off for embedded vision systems. This study contributes to Informatics and Computer Science by providing a systematic edge AI deployment framework for efficient model optimization and benchmarking in resource-constrained environments.]]>
Copyrights © 2026
