Multi-Source Domain Adaptation for Quality Control in Retail Food Packaging

<p>Retail food packaging contains information whichinforms choice and can be vital to consumer health, includingproduct name, ingredients list, nutritional information, allergens,preparation guidelines, pack weight, storage and shelflife information (use-by / best before dates). The presence andaccuracy of such information is critical to ensure a detailedunderstanding of the product and to reduce the potential forhealth risks. Consequently, erroneous or illegible labeling has thepotential to be highly detrimental to consumers and many otherstakeholders in the supply chain. In this paper, a multi-sourcedeep learning-based domain adaptation system is proposed andtested to identify and verify the presence and legibility ofuse-by date information from food packaging photos takenas part of the validation process as the products pass alongthe food production line. This was achieved by improving thegeneralization of the techniques via making use of multi-sourcedatasets in order to extract domain-invariant representations forall domains and aligning distribution of all pairs of source andtarget domains in a common feature space, along with the classboundaries. The proposed system performed very well in theconducted experiments, for automating the verification processand reducing labeling errors that could otherwise threaten publichealth and contravene legal requirements for food packaginginformation and accuracy. Comprehensive experiments on ourfood packaging datasets demonstrate that the proposed multisourcedeep domain adaptation method significantly improvesthe classification accuracy and therefore has great potential forapplication and beneficial impact in food manufacturing controlsystems.</p>