A method for pre-sorting mixed mail using convolutional neural networks and transfer learning

Research output: Chapter in Book/Report/Conference proceedingConference paperpeer-review


Sorting mixed mail has become increasingly important for postal service
providers. Low-value consignments, that have unfavorable material
properties, such as odd shape or flexibility, have increased in terms of
volume in recent years and cause substantial problems in sorting
processes. In order to automate the sorting of mixed-mail items the so-
called ‘Free-Fall-Sorter’ was developed at the Institute of Logistics
Engineering at Graz University of Technology. In this paper, an image
classification system is proposed to pre-sort mixed mail for this new
machine. The computer vision system used for the experiments applies
transfer learning to a publicly available convolutional neural network
called ‘MobileNet V2’. The classification system is designed to be
composed of low-cost hardware components, yet capable of providing
real-time inference. The results show a high accuracy of the re-trained
Translated title of the contributionEin Verfahren zur Vorsortierung von Postsendungen mittels faltenden neuronalen Netzen und Transferlernen
Original languageEnglish
Title of host publicationMHCL 2022
Subtitle of host publicationXXIV International Conference on "Material Handling, Constructions and Logistics" September 21st - 23rd, 2022, Belgrade
Place of PublicationBelgrade
PublisherFaculty of Mechanical Engineering, Belgrade University
Number of pages10
ISBN (Print)978-86-6060-134-8
Publication statusPublished - Sept 2022
Event24th International Conference on Material Handling, Contructions and Logistics: MHCL 2022 - Faculty of mechanical Engineering - University Belgrade, Belgrad, Serbia
Duration: 21 Sept 202223 Sept 2022


Conference24th International Conference on Material Handling, Contructions and Logistics
OtherXXIV International Conference on Material Handling, Contructions and Logistics
Internet address

Fields of Expertise

  • Mobility & Production

Cite this