Exploring Data Augmentation for Trash Classification

Spring 2023 Final Project for ECE 285: Visual Learning

Abstract: This project investigates the impact of data augmentation techniques on the performance of deep learning models trained on the small TrashNet dataset. Two pre-trained models, ResNet50 and VGG16, were used to classify the trash images. Data augmentation was performed through traditional image transforms and a more experimental method using a Deep Convolutional Generative Adversarial Network (DCGAN) to generate additional training data. The results indicate that while augmentation techniques hold promise for improving performance on small datasets, the quality of the generated data significantly influences model performance. The findings, including that the original dataset-trained model performs best, provide important insights into the challenges and potential solutions when working with small datasets in the field of image classification. Overall, it is found that data augmentation for the TrashNet dataset is not a better approach to potentially be used to improve trash sorting and recycling efforts