Enhancing efficiency with AI solutions in the Unicatalog mobile app: evolving the image processing workflow

The Unicatalog mobile app, featuring promotional catalogues, provides users with numerous advantages. Various added functions accompany the catalogues and products within the app, features that are impossible to achieve with paper-based catalogues. For instance, users can subscribe to product types, search for products or categories to see when and where they are on sale, compile shopping lists from the most affordable or favourite items, and store all their loyalty cards within a single app.

However, loading promotional catalogues presented numerous challenges for the company, which were addressed using an AI solution specifically designed for image processing and computer vision tasks.

In the following section, we illustrate how the process of uploading promotional catalogues evolved over the app’s lifecycle.

The initial approach was entirely manual. Data entry operators originally entered each product from the catalogues one at a time through an administrative interface.

To add a product, operators would locate the relevant information and complete the appropriate fields. They were responsible for selecting the category and adding an illustration—the product image. Because different people performed the categorisation, there was no consistency or standardised approach to image selection. Some operators would cut the image directly from the catalogue on the administration page, while others would search for a similar image online. Both methods were time-consuming and often led to the selection of inappropriate photos. Frequent typos and data entry errors also occurred. The workflow was monotonous and slow, taking around 8 hours to complete a single catalogue.

With advances in AI, we reached a point where it was worth experimenting with AI tools. Using the first AI tool, we were able to automate data extraction within the catalogue upload process. ChatGPT 3.5 could roughly identify product names, descriptions, and prices with moderate accuracy. Although it was proficient at recognising prices, it often struggled to distinguish between multiple prices, such as regular, sale, or loyalty card prices. As a result, data entry operators were tasked with reviewing and correcting AI-generated data. At this stage, the AI development process did not yet support categorising products or attaching images, leaving these tasks to the operators.

While GPT-3.5 made frequent errors, the initial release of GPT-4 performed more reliably. However, it was the GPT-4 Turbo model, introduced during development, that made a substantial impact, reducing processing time by approximately 70% compared to manual data entry. Processing a 60-page catalogue, which previously took 7-8 hours, was reduced to around 2 hours (approximately 120 minutes).

In some cases, the AI’s recognition results were suboptimal, occasionally missing certain products or misinterpreting data due to excessive information on a single page. To address this, we modified the workflow by having operators frame each product and its information within the catalogue pages, submitting only the framed sections to the AI. This adjustment improved two aspects: recognition quality increased significantly, and the framed section served as the product image. With this more targeted use of AI, we further accelerated the upload process, allowing a 60-page catalogue to be processed in around 75-90 minutes.

With the introduction of framing, data recognition was largely resolved, but product categorisation remained the responsibility of data entry operators. To address this, we trained an AI-based classification model to automate this task, removing it from the operators' duties. Now, the only remaining task in the process is to verify the AI-generated categorisation.

The last manual task, framing, was successfully replaced by an AI-based tool. We chose object recognition as the appropriate AI technology. Training the AI-based object detection model required numerous sample examples, which we created locally using the Label Studio tool. For each retail chain, we manually framed 10 weeks of promotional catalogues and used this data to train our framing model on the Roboflow platform. The resulting model is deployed via an API provided by Roboflow: we send a page to the API, and the model returns the automatic framing, marking each product’s area.

Framing, data recognition, and categorisation are now fully performed by AI tools, with the only remaining human task being the verification of AI processing. Consequently, our benchmark 60-page catalogue now takes just 45 minutes to process - less than 10% of the time originally required. Instead of multiple data entry operators, we now only need a single data verifier. However, our goal is to eliminate this verification step as well, further reducing the need for human intervention by eventually automating it with AI solutions.

We aim to improve the process to the point where verification is no longer necessary, achieving fully reliable AI processing. With continued AI software development and further refinement of the workflow, we believe this goal is achievable shortly.