DeepWasteSort-SI-SSO: A Vision Transformer-Based Waste Image Classification Framework Optimized with Self Improved Sparrow Search Optimizer

Automated waste classification is essential for improving recycling efficiency and supporting sustainable waste management systems. However, conventional convolutional neural network (CNN) approaches primarily focus on localized feature extraction, which may limit their ability to capture complex spatial relationships in heterogeneous waste materials. This study proposes DeepWasteSort-SI-SSO, a Vision Transformer (ViT)-based framework enhanced with a Self-Improved Sparrow Search Optimization (SI-SSO) strategy for hyperparameter tuning. The optimization process focuses on key training parameters, including learning rate, batch size, and dropout rate, to improve convergence stability and reduce the risk of suboptimal local minima. The framework was evaluated on a balanced four-class waste image dataset (paper, wood, food, and leaves; N = 4000) using a five-fold cross-validation protocol. Experimental results achieved an average accuracy of 95.5% (±0.007), a macro-averaged AUC-ROC of 0.975, and a Cohen’s Kappa coefficient of 0.938, indicating strong agreement between predicted and true labels. Comparative experiments against ResNet-50 and a baseline ViT configuration suggest that SI-SSO optimization improves performance stability with only a modest increase in computational cost. These findings highlight the potential of optimized Transformer-based approaches for automated waste image classification under controlled evaluation conditions.