<![CDATA[Automatic recognition of regional food specialties using embedded devices still faces challenges, particularly prediction imbalance (bias) due to the very small dataset size. This study developed a recognition system for eight types of Yogyakarta specialties (gudeg, sate klatak, bakpia, jogja tahu, pecel gudeg, ayam geprek, es dawet, and klepon) based on ESP32‑CAM integrated with a Flask server and a TensorFlow model (MobileNetV2). The training dataset consisted of only 51 images (average 6–7 images per class). The main problem encountered was extreme bias: the initial model predicted the "klepon" class in 85% of the tests. To address this, a post-prediction anti-bias algorithm was proposed that combined multiple randomization techniques, HSV-based color feature analysis, and weighted probability distributions. Tests were conducted on 15 scenarios with varying lighting and shooting angles. The results show that the anti-bias algorithm successfully reduced the dominance of klepon predictions from 85% to 15%, and achieved a more balanced distribution of predictions between classes. The system's average response time was 1.2 seconds with a 94% success rate for ESP32-server communication. The recognition accuracy under optimal conditions reached 78% (72% average). This system demonstrates that a hybrid edge-cloud approach with a post-prediction anti-bias algorithm can be an effective solution for object recognition on very small datasets, especially for IoT-based culinary applications.]]>
Copyrights © 2026
