<![CDATA[Waste management, particularly in sorting plastic bottle and non-bottle waste, remains a challenge in supporting effective recycling systems. This study aims to implement the MobileNetV2 model on the ESP32-CAM device using the Edge Impulse platform and to evaluate the model’s performance in real-time object classification. The dataset consists of 200 images containing bottle and non-bottle objects with variations in lighting conditions, shooting angles, and backgrounds. The model was trained using Edge Impulse and then converted into TensorFlow Lite format for deployment on the ESP32-CAM device.The training results show that the model achieved high performance with an accuracy of 92.50%, supported by an AUC of 0.97, precision of 0.98, recall of 0.97, and F1-score of 0.97. Based on the simplified confusion matrix with visual verification, the model achieved 100% accuracy in detecting bottle objects and 95% accuracy for non-bottle objects, with a 5% misclassification rate. However, during real-world implementation on the ESP32-CAM device, the model’s performance decreased to 65.7% accuracy due to differences between training and real-world conditions as well as hardware limitations.Despite this, the system successfully performed real-time image classification on an embedded device. This study demonstrates that the edge artificial intelligence approach using MobileNetV2 and Edge Impulse can be effectively applied to resource-constrained devices, although improvements are still needed in terms of model generalization and dataset diversity.]]>
