Deep residual and hybrid CNN models for confidence-aware real-world waste classification for sustainable waste management

Efficient waste classification is crucial for promoting recycling and achieving sustainable waste management. Real-world waste streams, however, often include mixed, deformed, and contaminated items, making manual sorting inefficient and error prone. A deep learning-based system for multi-class classification of heterogeneous waste using the RealWaste dataset is presented in this paper, which reflects actual disposal conditions such as cluttered backgrounds and overlapping materials. We fine-tune and evaluate several convolutional neural networks (CNNs), including InceptionV3, ResNet101, DenseNet, VGG, EfficientNet, and MobileNet. Among these, ResNet101 demonstrated the best performance, achieving a validation accuracy of 98.86%, loss of 0.0379, and 0.99 as F1 score. We also introduce hybrid models (e.g., ResNet101 + InceptionV3), which improved precision in complex categories such as textiles and miscellaneous trash. Furthermore, a confidence score evaluation strategy is proposed to assess model reliability, revealing high confidence (≥ 0.95) for visually distinct classes like vegetation, plastic, and food organics. Our findings establish a robust and scalable benchmark for deploying intelligent waste classification systems in real-world, sustainability-driven environments.