<![CDATA[Automatic sorting of solid waste is difficult because discarded items are often badly deformed and some material classes, such as transparent plastic and glass, look almost identical. Conventional Convolutional Neural Networks (CNNs) mainly learn local patterns; as a result, they tend to overlook global context and to overfit when only a small and unevenly distributed dataset is available. This study develops a hybrid deep learning model that pairs the dense local feature learning of DenseNet201 with the global context modeling of a Vision Transformer (ViT) in order to make waste classification more robust. An ImageNet-pretrained DenseNet201 serves as the backbone, and its high-level feature maps are converted into a sequence of patch tokens that is processed by a custom ViT encoder. Experiments were conducted on a public garbage image dataset with six unevenly represented classes, namely cardboard, glass, metal, paper, plastic, and trash. On-the-fly data augmentation and early stopping were employed during training to limit overfitting. The hybrid model reached an overall accuracy of 87% and a macro F1-score of 0.85, exceeding the CNN baselines trained under identical settings. It was especially accurate on fibrous materials, with precision values of 0.95 for cardboard and 0.89 for paper. These results indicate that coupling self-attention with densely connected convolutional layers alleviates the main weaknesses of pure CNNs and provides an accurate, interpretable, and computationally efficient basis for practical automated waste sorting systems.]]>
Copyrights © 2026
