Plastic Subtype Recognition System Using a TinyML Based Lightweight CNN on ESP32 Smart Bin

Abstract

Plastic waste segregation is important for optimizing the recycling process and minimizing the negative effects associated with improper disposal of plastic waste. Nonetheless, most of the smart waste management systems are designed mainly for general waste classification, fill level determination, or cloud computing, neglecting real time identification of different types of plastics at the disposal site. In this research, a plastic type recognition model based on edges is developed, embedded in a smart bin using the Light Weight Convolutional Neural Network and Tiny ML. The developed recognition model is able to classify plastic waste into five different types, namely PET, HDPE, LDPE, PP, and others/mixed. Hence, this result can be used to automatically sort the plastic waste to its correct compartment. The model was evaluated using a custom dataset created by combining images from multiple publicly available plastic waste image datasets. The lightweight MobileNetV2 architecture based CNN was trained with images of size 160 x 160 pixels (RGB format) and tested against the validation set. The trained model was then quantized to produce an INT8 TensorFlow Lite model to guarantee a smaller computation footprint for deploying the model in ESP32 hardware. While the trained Keras model gave an accuracy score of 92.08% on the validation data set, the quantized INT8 model produced an accuracy of 90.68% which translates to an approximate drop in accuracy of about 1.40 percent point. The implication is that with the help of lightweight deep learning models and INT8 quantization, one can still achieve an accurate classification of plastic subtypes. The conclusions drawn from this study is that TinyML based edge inference can be a practical way for plastic subtype classification in order to automate the process of smart waste segregation via classification without relying on cloud based classification continuously.