A Model For Classification Of Trash Based On Deep Learning And Image Processing Techniques

Abstract

Environment-related problems have become increasingly hazardous towards humans and pose a risk to the ecosystem, one of which is the production of abundant trash. Despite many solutions, trash recycling still meets several difficulties. This thesis introduces a solution of using two convolutional neural network models to help classify five types of household trash (cardboard, paper, glass, fabric, metal) through images and employing a YOLOv8 model for object segmentation. The main predictor model is MobileNetV3Large, chosen for its fast performance and good overall accuracy. The support model is EfficientNetV2B0, used for boosting the system’s overall prediction confidence. The training dataset is combined from several open-source online datasets with custom augmentation techniques applied. Grad-CAM and t-SNE algorithms are employed to assess the effect of custom augmentation. The model achieves 94% accuracy with fast inference time and good stability, enabling faster trash sorting for better recycling efficiency.