One-dimensional convolutional neural networks for acoustic waste sorting

Abstract

This paper presents a successful application of one-dimensional convolutional neural networks (1D CNNs) for waste sorting at source based on acoustic data. Most of the existing methods use images for waste classification, which causes a high computational complexity and requires a huge amount of training data. Acoustic data have also been employed due to its high correlation to the waste material type. The traditional approaches usually use fixed and handcrafted features extracted from acoustic data. Finding efficient features is usually time consuming and could be difficult in complex situations. In this paper, the 1D CNN method is proposed to automatically extract features from acoustic data and achieve high classification accuracy. Orthogonal experiment method is applied to optimize three key hyper-parameters of the 1D CNN structure. The experiment shows that the proposed method can achieve 92.40% classification accuracy within a short time. A traditional method using handcrafted features with a shallow classifier is taken as a benchmark and the attained classification accuracy is only 81.92%. In addition, a classification accuracy rate of 68.06% is achieved when using a shallow classifier with raw acoustic data as input. Therefore, the proposed method is promising to be practically applied in the trash bins for automatic waste source separation.