Robotic Sorting of Used Button Cell Batteries: Utilizing Deep Learning

Abstract

In this study, a technique has been developed to enable the automated sorting and processing of used button cell batteries. The objective of this system is to automatically classify button cell batteries into their chemistries based on the markings on the surfaces. These markings can potentially include their item code, manufacturer, and/or chemistry. Due to the large input image size (16 mega pixels) traditional object detection networks could not be trained with the equipment available. To combat this, 3 different deep learning techniques have been examined; strict convolutional, image splitting, and deep scaling networks. Each of the network types come with their own strengths and weaknesses, and can run near or at real-time speeds, with accuracy rates of 80% or above. The promising results are currently being integrated with high speed robotics to increase the capacity and profitability for our industry partner; Raw Materials Company (RMC).