Analysis of the Robustness of a Deep Learning Model for Non-Organic Waste Classification Against Variations in Lighting and Image Noise

Abstract

The increasing amount of non-organic waste presents challenges in effective waste management and sorting processes. Deep learning technology has been widely used for image-based waste classification; however, most previous studies mainly focused on improving model accuracy under ideal conditions without considering model robustness against visual disturbances in real-world environments. This study aims to analyze the robustness of deep learning models for non-organic waste classification under illumination variations and image noise using the public TACO (Trash Annotations in Context) dataset. Three deep learning models were employed, namely EfficientNet-B0, ResNet50, and MobileNetV2. The experiments were conducted by applying multiple brightness levels and Gaussian noise disturbances. The results showed that EfficientNet-B0 achieved the best performance with an accuracy of 87.84%, precision of 87.17%, recall of 85.35%, and F1-score of 85.39%. Furthermore, EfficientNet-B0 obtained the highest robustness score of 0.823 compared to ResNet50 and MobileNetV2. The findings indicate that illumination variations and image noise significantly affect model performance, especially under severe visual disturbances.