The IMPRO project aimed to automate the creation of product data for omni‑channel retailers by extracting information from product images and web sources using machine‑learning techniques. The core of the effort was the Image‑Supported Product Data Creation Processor (IMPRO), a prototype that integrates optical character recognition (OCR), image‑based object detection, natural‑language processing, and data fusion to produce complete product records for ERP or PIM systems.

For OCR, the team evaluated several libraries and services. The Google Vision API was chosen for most text extraction tasks because of its superior accuracy on the multilingual product images supplied by the associated partners Globus and Coop. Amazon Rekognition was also tested but did not meet the required precision. Extracted texts were cleaned by converting to lowercase, removing stop words, and eliminating connectors and numbers. The cleaned text was then translated into English with the DeepL API to standardise the language for downstream processing. To transform the unstructured text into machine‑learnable features, the project applied TF‑IDF and Weighted Log Odds Ratio, producing vectors that capture word frequency and importance. These vectors served as inputs for custom classification models that assign product attributes such as brand, category, and nutritional information.

In parallel, the Hochschule Trier team focused on visual feature extraction. Using object detection, the system first identified relevant regions on a product image—such as the product name, brand logo, or nutrition table. Depending on the region class, either a specialised reader (NutritionTableReader) or a generic OCR engine (Tesseract or EasyOCR) was invoked. The NutritionTableReader was specifically trained to parse tabular nutritional data, a common requirement for food retailers. The object detection and classification pipeline was evaluated on a training set of 1,034 images, yielding a label distribution that guided the prioritisation of classes for further optimisation.

The fusion stage combined the OCR‑derived text vectors, the detected visual attributes, and supplementary data from web crawls and internal ERP tables. This multi‑modal integration produced a unified product record that could be imported into the retailer’s ERP or PIM system. While the report does not provide explicit accuracy metrics, the prototype demonstrated a significant reduction in manual data entry effort and an increase in data consistency, as evidenced by the successful generation of complete product datasets for the partner retailers.

The project was structured into six work packages: project management, image‑based information extraction, web‑source data acquisition, intelligent product‑image‑based master data creation, image‑based master data verification, and evaluation and exploitation. retailsolutions GmbH served as the consortium coordinator, handling project management and the OCR evaluation. Hochschule Trier led the development of the object detection, classification, and NutritionTableReader components. The associated partners Globus, Coop, and Transgourmet supplied the product images and ERP data necessary for training and validation.

The initiative ran from 1 January 2021 to 31 December 2022, with a cost‑neutral extension until 31 May 2023. It was funded by the German Federal Ministry of Education and Research under grant number 01IS20085. The collaboration between a commercial software firm and an academic institution enabled the rapid prototyping of a system that promises to streamline product data management for omni‑channel retailers, reducing costs and improving data quality across multiple sales channels.