Deep Learning-Based Waste Classification with Transfer Learning Using EfficientNet-B0 Model

Recycling of waste is a significant challenge in modern waste management. Conventional techniques that use inductive and capacitive proximity sensors exhibit limitations in accuracy and flexibility for the detection of various types of waste. Indonesia generates approximately 175,000 tons of waste per day, highlighting the urgent need for efficient waste management solutions. This study develops a waste classification system based on deep learning, leveraging the powerful EfficientNet-B0 model through transfer learning. EfficientNet-B0 is designed with a compound scaling method, which uniformly scales network depth, width, and resolution, providing an optimal balance between accuracy and computational efficiency. The model was trained on a dataset containing six classes of waste—glass, cardboard, paper, metal, plastic, and residue—totalling 7014 images. The model was trained using data augmentation and fine-tuning techniques. The training results show a test accuracy of 91.94%, a precision of 92.10%, and a recall of 91.94%, resulting in an F1-score of 91.96%. Visualization of predictions demonstrates that the model effectively classifies waste in new test data. Implementing this model in the industry can automate the waste sorting process more efficiently and accurately than methods based on inductive and capacitive proximity sensors. This study underscores the significant potential of deep learning models, particularly EfficientNet-B0, in industrial waste classification applications and opens opportunities for further integration with sensor and robotic systems for more advanced waste management solutions.