Smart Trash Can: Image Classification and Segmentation for Waste Quality Management in the Home

Abstract

Recycling organic waste through composting or fermentation is a simple but, so far, underrated way to capture CO2. Its effective and qualitative production requires the strict separation of recyclable organic waste from residual waste, but the persistent lack of public awareness still results in a considerable amount of contamination. Impurities from just a single container inevitably reduce the quality and can result in the expensive and less sustainable energy recovery of entire truckloads. To avoid the incineration of valuable organic waste, the problem must be tackled at the producer’s site by preventing the addition of unsuitable waste. We developed a Smart Trash Can that allows to unobtrusively take photos of real waste in the home. Over six months, a total of 450 photos were collected, which were then manually labeled and segmented according to the captured waste types. Based on the collected dataset, two machine-learning approaches for computer vision have been implemented. While the first one binarily classifies the images as either being pure organic or containing impurities with a mean accuracy of 90.35%, the second one detects and segments impurities in the images with a mean accuracy of 98.24% and a mean intersection over union value of 96.43%. With the presented system, not only can contaminated waste be separated from pure organic waste, but also could feedback, incentives, or nudges encourage the producer to separate their waste more carefully in the future.