Detection and estimation of 3D shape of waste based on neural networks

Abstract

Abstract This article considers the possibility of using neural networks to detect and recognize different types of waste and determine its shape. The developed algorithm can be used to automate the waste sorting process for further sorting, recycling or disposal. This article deals with the detection of 28 classes of objects made of iron, plastic, cardboard and glass with varying degrees of deformation and damage. The detection is based on the YOLO8 neural network. The quality of the YOLOv8 model was evaluated on test data using the following metrics: box_loss, cls_loss, dfl_loss, mAP50, mAP50-95, precision, recall. After training the neural network, the YOLOv8 metrics on the test data are: box_loss=1.14, cls_loss=2.21, dfl_loss=1.21, mAP50=0.48, mAP50-95=0.38, precision=0.60, recall=0.44. The 3d shape detection is done after analyzing the image by YOLOv8 model based on GLPN neural network.