Neural Network Based OCR for Keg Identification

Abstract

A keg asset management system that can reduce the annual rate of keg attrition by 5% to 20% can deliver significant savings to breweries with large fleets of kegs. A typically large brewery can have at least tens of thousands of kegs, a sizable investment given an initial cost of around USD100 per keg. A key element in a keg tracking system is on-line keg identification. This research explores the feasibility of an intelligent machine vision approach to identifying the unique serial number embossed on the dome of each keg at manufacture. The demonstration system developed auto-locates candidate serial numbers and applies optical character recognition (OCR) techniques. The neural network based OCR
achieved the best performance over template matching achieving an overall recognition rate of 92% and no missed digits. If non-permanent serial number occlusions can be removed by caustic washing prior to the image capture stage in a production line implementation, the
recognition rate approaches 97%.