Garbage Image Classifier using Modified ResNet-50

Abstract

This research proposes a deep learning model pretrained with ResNet-50 to classify 12 types of garbage. The model uses a modified ResNet-50 architecture with the Adamax and Adadelta optimizers and varying learning rates (0.1, 0.01, and 0.001). Six experiments were conducted to determine the most optimal training parameter configuration for the proposed model. Results show that the model performed best with the Adadelta optimizer and a learning rate of 0.1, achieving a validation accuracy of 93.85%. In comparison, the Adamax optimizer with a learning rate of 0.001 yielded a validation accuracy of 93.44%. Despite these results, there is a tendency for misclassification in the metal, plastic, and white-glass classes. Future work should focus on addressing these misclassification issues by expanding the dataset for these problematic classes. This can be achieved either by collecting additional images specific to these classes or by employing advanced data augmentation techniques to enhance the existing dataset and improve the model's accuracy.