Real-Time Waste Classification Using EfficientNet-B3 with Advanced Data Augmentation and Transfer Learning

Abstract

Given the billions of tons of solid waste produced worldwide each year, waste recycling is critically important for the global environment and economy. Classifying recyclable materials is essential to overcome the inefficiencies and high error rates associated with manual sorting. Deep learning models offer a powerful solution for this purpose. In this study, a real-time artificial intelligence-based waste classification system was developed to provide an efficient and accurate approach. The EfficientNet-B3 architecture was applied to the “Trash Type Classification Combination Dataset,” which consists of eight different waste classes. Extensive data augmentation techniques including spatial transformations, color space adjustments, random erasing, and mixup were employed to address class imbalance and prevent overfitting. Experimental results show that the EfficientNet-B3 model, utilizing transfer learning and fine-tuning, achieved a state-of-the-art test accuracy of 98.80% and an F1 score of 0.9855. Furthermore, the developed model was successfully deployed as a real-time web application using the Streamlit framework. Overall, the model demonstrates high performance and produces reliable outputs for waste classification tasks.