Identification for Drinking Spout of Plastic Bottle Using Multi Eff-UNet

Abstract

Currently, the increase of general waste is a problem in Japan. It is garbage and human waste. One solution to this problem is to reuse the bodies of used plastic bottles to reduce the amount of plastic waste and the increase in general waste. For this purpose, some municipalities provide bags for disposing of plastic bottle bodies. However, in many cases, the bottles are not only discarded in these bags, but also with their caps attached. In such cases, the caps must be found and removed by hand in the sorting and processing facility at the waste treatment plant. This means that the same work is performed by person for a long time, causing problems such as overlooking caps due to fatigue. To solve these problems, we develop a method for automatic identification of plastic bottle bodies and caps using deep learning technique. In this paper, we propose a model that combines multiple Eff-UNets. Specifically, we combine EfficientNetB4 for local segmentation and EfficientNetB5 for global segmentation. By using our method, we conducted an experiment on images of plastic bottles collected from the internet and other sources. We obtained segmentation results of 97.9% for plastic bottle bodies and 86.0% for plastic bottle caps.