Identifisering av trykte objekter for sortering av retur

Machine learning for fast sorting of printed objects Recycling and sorting of returns is an import environmental issue. Solutions for sorting and handling of returns need to be simple, fast and cost efficient. In this project HS News Systems, Tomra and Norwegian Computing Center have joined forces to develop solutions for image-based sorting of returns through identification of printed matter such as magazines and newspapers and printed container labels. For a sorting station for magazine returns speed is an important aspect as a magazine needs to be recognized as one of many thousand categories in just one second. An additional challenge is that the magazines often have a plastic wrapping that causes reflections and which may also contain additional products such as books, CDs, toys etc. For beverage containers, there is a potential for exploiting the labels to a larger extent than today. Challenges here are also related to processing speed, in addition to specific needs for these objects, such as curved surfaces where only parts of the object are visible at a time, and where one class of objects can be represented by several labels. Traditional image-based techniques have fallen short for these problems where we have a large number of classes, high speed requirements and partly occluded objects. Through the project methods based on novel techniques have been studied, and solutions for reliable and efficient image-based sorting have been developed. For sorting of printed matter this has led to a much improved product for the Kongsberg based company HS News Systems, where they are now able to offer a sorting station that is able to handle twice as many products per unit time as their competitors. HS News Systems is an only Norwegian vendor on this international market, and has by this achieved a competitive advantage. The new sorting station is now running in production at Bladcentralen, Norway?s largest distributor of magazines and periodicals. A typical situation here is that incoming magazines need to be classified into one of 10-15,000 classes. With the novel solutions, the sorting station is able to handle 4000 objects per hour with a recognition rate surpassing 98%, which is considered as very good. The new methods can handle recognition under challenges such as partial occlusion due to reflections from plastic wrapping or products contained in this wrapping. This is important due to an increase in the amount of plastic wrapped magazines coming with auxiliary products in the cover. The solution includes automatic and robust classification to minimize the number of errors and notify the operator when the system is in doubt. In addition, new intelligent methods enabling automatic selection of images for the reference database have made the training of the system less dependent on the operator. This simplifies the process and makes the performance more stable over time. In the laboratory the methods have been tested for very large databases, with up to 80,000 images of magazines and books. For this large database identification is still made in less than one second, with a correct recognition rate of 99%. Based on similar techniques and methods, algorithms and models designed for labels on beverage containers have also been developed. Tests indicate that these approaches have a potential also for these objects.

Prosjektet fokuserer på effektive løsninger for automatisk retursortering gjennom identifikasjon av trykte objekter for to ulike anvendelser; sortering av trykksaker og sortering av drikkevareemballasje basert på etiketter. Begge problemstillingene stille r strenge krav til robusthet, nøyaktighet og effektivitet. Bildebasert gjenkjenning basert på lokale deskriptorer representerer et paradigmeskifte metodemessig som gir nye muligheter for å finne løsninger på disse sorteringsproblemene. Med utgangspunkt i disse metodene er hensikten å utvikle et system for sortering av trykksaker. I tillegg skal de samme metodenes potensiale evalueres også for sortering av drikkevareemballasje samt juksutfordringer knyttet til dette. Innovasjonen består i å ta frem løsni nger som gjør at systemene er mindre avhengige av strekkode. Det gir nye muligheter for håndtering av objekter der strekkoden mangler eller er ødelagt, men på sikt kan det også åpne for nye muligheter. For sortering av trykksaker kan det forenkle systemen e gjennom at det blir tilstrekkelig med avbildning fra en side, og det kan også øke funksjonaliteten ved å gjøre det mulig å håndtere bunker av like tidsskrifter. For drikkevareemballasje kan det gi nye muligheter for håndtering av fremtidens ikke-rotasjo nssymmetriske objekter. Forskningsutfordringene består i å komme fram til løsninger som kan fungere i sann tid med hensyn både til beregning av robuste deskriptorer og matching av disse for problemstillinger der antallet klasser kan komme opp i flere tus en.