Authors - Sakshi Tiwari, Snigdha Bisht, Kanchan Sharma Abstract - Effective waste management is critical to achieving sustainability in urban regions like Delhi-NCR, where heterogeneous waste streams pose a classification challenge. In this research, we propose WasteIQNet, an intelligent deep hybrid model designed for precise waste classification across 18 categories under a well-defined hierarchy: Wet (Compostable, Special_Disposal) and Dry (Recycle, Reduce, Reuse). Leveraging the WEDR dataset, we first standardized over 1.75 lakh images via JPEG conversion, 256×256 resizing, and RGB formatting. SMOTE+ENN was applied to balance class distributions to 20,000 images each. Feature extraction was achieved through simulated DASC-like global vector embeddings using MobileNetV3. Our baseline hybrid model integrated MobileNetV3Large and GraphSAGE, achieving an initial accuracy of 80.56%. After optimizing the model for multi-label learning through sigmoid activation, threshold-based decoding, and hierarchical label interpretation, we conducted extensive enhancements. Hyperparameter tuning with Optuna, Feature-wise Attention (FWA), and Top-K Mixture of Experts (TopK-MoE) improved accuracy to 83.33%. Subsequent normalization and activation function experiments (Mish, Swish, GELU) led to a peak accuracy of 94.44% using GELU. We further introduced Dynamic Sparse Training (DST) and Model-Agnostic Meta-Learning (MAML), raising accuracy to 95.04%. Final enhancements included label smoothing and early stopping, culminating in a best-in-class accuracy of 97.87%. WasteIQNet demonstrates a scalable, interpretable, and high-performance solution for automated waste classification, supporting smart city initiatives and responsible environmental management.
Open Zoom
