An Ensemble Learning Approach towards Accurate Waste Segmentation in Cluttered Environment

Abstract

Environmental pollution is a critical global issue caused by large volumes of waste generated. Recycling is emerging as one of the most viable solutions to mitigate the impact of waste on environment. This study focuses on waste segregation which is a crucial step in recycling processes. Traditionally, human labor is employed for waste segregation at Material Recovery Facility. This poses health risks due to direct human contact with waste. To mitigate this risk, automation is increasingly being adopted there. Recent advancements in computer vision have significantly contributed to waste classification and recognition. Deep Learning models are able to generate accurate predictions with their ability to extract and organize features through multiple convolutional layers. In waste segregation, segmentation masks are essential for robots to accurately localize and pick objects from input stream. The complexity of real-world waste environments is characterized by deformed, broken, distorted items without specific patterns and overlapping objects, which hinders waste segmentation tasks. This study presents an Ensemble Learning approach to improve segmentation accuracy by combining two Deep Learning models. The dataset used closely mimics real-life waste scenarios. Preprocessing techniques were applied to enhance feature learning for these models. State-ofthe- art segmentation models were evaluated over ZeroWaste-f dataset. Based on their characteristic and high performance, U-Net and FPN are combined using an equal average ensemble method. The ensemble model, referred to as EL-4, achieved an IoU value of 0.8306, an improvement over U-Net’s 0.8065, and reduced Dice loss to 0.09019 from FPN’s 0.1183. Previous studies reported an IoU of 0.6588 on this dataset. This study could improve accuracy of cluttered waste sorting at Material Recovery Facility. This will facilitate acquisition of better raw material for recycling with minimal human intervention.