WASTE CLASSIFICATION USING VISION TRANSFORMERS

Effective identification of recyclable waste is a major challenge in resource management and environmental protection. The present study explores the integration of transformer-based architectures for the accurate classification of recyclable waste, including plastic, glass, metal, and paper. A dataset consisting of digital images of different types of waste was used to train and evaluate the proposed architectures. To improve the generalization of the model a division of the data set was pursued for training, validation, and testing areas, as well as the implementation of data augmentation and transfer-learning techniques. Compared to traditional methods and different convolutional neural network architectures, transformer-based architectures have demonstrated superior performance both in terms of accuracy and computational efficiency. Analyzing the experimental results, the proposed models demonstrated accuracy values of over 95%. The study finally notes that the use of transformer-based architectures for the classification of waste from digital images presents a major potential in the development of efficient waste management practices and for reducing the impact of waste on the environment.