Waste Classification Using Pre-Trained Vision Transformer and CNN Models

Abstract

Efficient waste classification is crucial for sustainable waste management, particularly in rapidly developing areas.
In this paper, the performance of the Vision Transformer (ViT) and Convolutional Neural Network (CNN) models
(VGG19, ResNet50V2, InceptionResNetV2, and DenseNet121) was evaluated on a publicly available dataset comprising a total of ten waste categories. All models were trained using the same preprocessing and data augmentation techniques, and their accuracy, precision, recall, F1-score, and loss values were tested. Overall, ViT outperformed all CNN baselines with a test accuracy of 96.24%, an F1 Score of 0.96, a recall of 95% and a precision of 96%. VGG and ResNet50 remained competitive among the CNN models, but ViT demonstrated a better generalisation on the test set, which opens a potential use: automated waste classification. However, there is a need for more testing on real-world and domain-specific datasets to improve recycling accuracy and to provide cleaner and more efficient waste management.