Robust Waste Image Classification Under Class Imbalance and Visual Variability: Weak-to-Strong Learning with Activation-Guided CutMix

Automated waste image classification remains challenging because of substantial class imbalance, large intra-class visual variability, and high inter-class visual similarity. We propose a backbone-agnostic weak-to-strong training framework that combines class-balanced focal supervision over weak, strong, and activation-guided mixed views, asymmetric Kullback–Leibler (KL) consistency, and Activation-Guided CutMix. The detached weak-view prediction distribution serves as the reference for aligning the strong-view prediction, while Activation-Guided CutMix centers the replacement region at the highest-response location derived from an intermediate feature map of the base weak view rather than selecting its location randomly. The framework was evaluated on two waste-classification datasets using seven CNN backbones. ConvNeXt-Base achieved 98.88% mean accuracy and 98.44% macro-F1 on the corrected Waste Classification dataset, while ConvNeXt-Tiny achieved 79.20% mean accuracy and 76.73% macro-F1 on WaRP-C. Ablation studies demonstrated performance gains across different backbone architectures, with component-level contributions varying by dataset and backbone. The t-SNE, confusion-matrix, and Grad-CAM analyses provided complementary qualitative evidence of more clearly organized feature distributions, reduced off-diagonal class confusion, and more object-focused activation patterns. Because all auxiliary operations are confined to training, the proposed framework preserves the original single-branch, single-forward-pass inference architecture.