A Context-Preserving Vision–IoT Framework for Automated Solid Waste Segregation Using Hierarchical Token Decomposition

Objectives: This study aims at creating a smart vision-IoT system. It is to propose an automated waste classification through preserving material, geometric and spatial context clearly. To resolve ambiguity and misclassification that can be seen in current waste segregation frameworks. Method: It proposes a Context-Preserving Vision-IoT (CPV-IoT) framework, which includes a Hierarchy Tokens Decomposition (HTD) mechanism in the encoding of surface, geometric and contextual semantics before inference. An attention-based transformer is used to carry out contextual reasoning and then lightweight classification is done. The model is tested with the help of simulation-based experiments based on a publicly available waste image dataset, which simulates the work of IoT-enabled smart bins. Findings: The simulation evidence shows that the proposed framework has accuracy of 96.87% and F1-score of 96.41%. The results are higher than classical machine learning models and more recent deep learning baselines. Ablation analysis establishes that both hierarchical tokenization and contextual reasoning lead to performance improvement. Novelty: The originality of this work is that hierarchical semantic tokenization of waste images is presented, which makes it possible to represent the visual representation in a structured way and then reason about it. This method will contribute to the existing study on waste classification by showing that context-preserving abstraction can dramatically increase the reliability of classification in simulated IoT-based waste management systems. Keywords: Waste segregation, Vision transformer, Hierarchical tokenization, IoT smart bins, Sustainable waste management