Vision Transformer Based Classification of Bulky Waste

This paper uses the ViT (Vision Transformer) model to perform the classification of large waste images. A total of 41,347 large waste images collected through crowdsourcing are used for experimentation. The primary goal of this paper is to adjust and compare various optimizers to achieve the highest possible waste classification accuracy. Experimental results show that when using AdamW (Adam Weight Decay) as the optimizer, the accuracy was 83.89%, which is higher than that of SGD (Stochastic Gradient Descent) at 81.22% or Adam (Adaptive Moment Estimation) at 83.21%. This demonstrates the ability to automatically adjust the learning rate when training large waste data and effectively train the model using AdamW, which performs well in multi-class classification tasks.