Detection of recyclable solid waste using convolutional neural networks and PyTorch

Abstract

Waste management in the recycling business is a time-consuming and labor-intensive process. In this context, the need to improve accuracy and reduce the time associated with this process is highlighted. In order to improve the classification of recyclable solid waste and streamline the waste management process, the creation of a convolutional neural network (CNN) model using PyTorch was proposed. The MLOps methodology was implemented in the development of the proposed model. In the first phase, an interview was conducted to analyze the waste sorting process in the company GIRA. In the second phase, the Taco Trash Dataset was reclassified, a CNN architecture based on RetinaNet was designed and the model was trained with hyper parameters based on related works. The third phase, the model was evaluated by testing and A/B testing. The model demonstrated high accuracy in waste detection and classification. It successfully identified materials such as paper, cardboard, PET bottles, hard plastic containers, flexible plastics, cans, glass, Tetra Pak containers, Flex foam and PET bottle caps. The loss was minimal, reaching 0.02120%, equivalent to 97% accuracy, and 80% accuracy in a real environment based on the Technology Acceptance Model (TAM). It is concluded that the implementation of a sorting and waste detection model optimizes the time and improves the accuracy of the sorting process.