From teacher to student: achieving high performance with compact models for waste classification

Plastic waste presents a significant environmental challenge due to its persistence and rising accumulation. Accurate classification of plastic types, including seven classes (PET, HDPE, PVC, LDPE, PP, PS, and others), is essential for efficient real-time recycling. To enable the deployment of large deep learning (DL) architectures on edge devices under real-time processing constraints, knowledge distillation is essential. Specifically, we employ TinyMobileViT-XXS, a vision transformer, as a teacher model to compress into a student network. This distilled network’s performance is then compared against state-of-the-art solutions. A major challenge in accurate classification stems from the inherent variability in waste images, including deformations, degradation, and changes in color and texture, necessitating high-precision models. Current public datasets are often fragmented, imbalanced, and limited in scope. To address this, we constructed a comprehensive, highly variable dataset by merging seven diverse public datasets. A comparative study utilizing this unified dataset provides a crucial reference point for future research in automated plastic waste classification, emphasizing the importance of robust and unified datasets for advancing this field.