Groupe Dubreuil: Develops an application for real-time product identification using VisionAI
About Groupe Dubreuil
Groupe Dubreuil is a family-owned group based in the North West of France that has been in business for over 100 years diversified into 7 areas.
For 3 generations, they have been entrepreneurs, curious and determined, with a very clear vision of the Group that they want to shape.
Tell us your challenge. We're here to help.
Contact usGroupe Dubreuil breaks down silos between business units by creating a powerful, reusable application for real-time product identification to meet its customer-centric strategy.
Google Cloud results
- 300,000 products in the catalog
- 500+ products recognized per month on the Agrizone mobile application (before we widely communicated about it)
- Product (or similar) found 90% of the time
- Average response time of 2.13 sec
Optimizes shopping experience using Vision AI
Introduction to Groupe Dubreuil's context and challenges
Groupe Dubreuil is a family-owned group based in the North West of France that has been in business for over 100 years. It is diversified into 7 areas, including aviation, automobiles, agriculture, construction, energy, trucks, and hospitality. The Group has a consolidated revenue of over 2.7 billion euros and employs 5,800 people.
Through its business activities, Groupe Dubreuil sells numerous parts and products through its subsidiaries. It can sometimes be challenging for our customers to find the product they are looking for among references listed in the company's catalogs.
To address this issue, Groupe Dubreuil has set the goal of creating an artificial intelligence application that can, from any photo of a product, find and suggest to users the corresponding product. The end goal is to facilitate the process of searching and purchasing a product to enhance user experience for our customers.
Project presentation: The product catalog and its users
The project to create a product search application for Agrizone was born out of Groupe Dubreuil's desire to meet the needs of its customers in the agricultural field.
As Agrizone is one of the leaders in e-commerce in the sale of agricultural parts, it was essential to allow its customers to quickly and easily find the parts they need.
Agrizone wants to be the partner of farmers in their daily life. The site offers the essentials of agricultural parts with more than 1,200 brands and over 750,000 references. It is a time saver for farmers, cereal growers, wine growers, horticulturists and foresters.
Some key figures on Agrizone:
- 35 people
- 227,000 customers
- Customer ratings on Ekomi: 4.4/5 (34,950 answers)
The challenge was important: to allow our customers to find a part among a variety of more than 300,000 references on the Agrizone website. To meet this requirement, our application had to present the corresponding product in less than 3 seconds after taking a picture.
First iteration for product identification: Architecture overview
How to successfully build intelligence that is capable of identifying a product from a photo? This question challenged data-scientists at Groupe Dubreuil.
To identify our product, it's necessary to match information such as manufacturer reference numbers, technical indications, and brand names from our product catalog with the photo taken by our customers. However, this information is often engraved on the parts or inscribed, so we have this information in photos, but it needs to be transcribed digitally. This is where Vision AI OCR solution has excelled.
Usage of Vision AI for OCR perfectly met our expectations. Image processing rapidity for OCR is unbeatable, allowing us to meet our time constraints. In addition, OCR capabilities of Vision AI were impressive in an agricultural context where parts are sometimes dirty and damaged.
"Vision AI's OCR capabilities and speed exceeded our expectations, delivering impressive results even in challenging conditions such as an agricultural context where characters to detect are often in parts, sometimes dirty and damaged."
—GOULPEAU Etienne, Data Scientist, Groupe Dubreuil
Using state-of-the-art computer vision models such as EfficientNet, a vector representation of each product image is created in a database of embedding vectors.
The embedding vectors allow for a finer characterization of the visual properties of the images, which facilitates the search for similar products.
Then, a distance metric must be defined, and the embedding vector of the image to be searched can be calculated using the EfficientNet model.
The closest embeddings to the searched image can be selected using a k-nearest neighbor search. For large datasets, an index data structure such as ANNOY can be created to allow for a quick search of items.
Solution optimization (normalization and standardization)
In our agricultural product detection project, we faced challenges related to reading the text on the captured photos, especially due to the engraving of the information on the parts that often made it difficult for OCR solutions to read. In our product detection case, one misidentified character may lead to misidentification of products and compromise the accuracy of our solution.
However, in collaboration with the Google Cloud team, we were able to improve our Vision AI solution by using specific image pre-processing to improve image quality. This optimization significantly increased our product detection rate from 63% to 92% on a sample of ground-truth images with manufacturer reference numbers.
"Leveraging Google Cloud's powerful infrastructure, we efficiently developed our compact yet comprehensive Image Scan solution."
—LY Tidiane, Data Scientist, Groupe DubreuilTo further improve the relevance of our search results, we combined embedding search with the use of OCR to extract text from images. We also reworked the quality of the images using a denoising technique to remove unwanted artifacts and noise in the images. This combination of methods resulted in more accurate and relevant results for the users of our application. By using denoised and white-framed images, along with the extracted text, we were able to refine our embedding search and improve the relevance of the search results. Thanks to these optimizations, we were able to offer a more reliable and accurate solution for the detection of agricultural products with a matching rate between 72% and 83% accuracy, thus meeting the specific needs of our users.
Final architecture and productization
Today, we are able to identify a product with a photo in 2.13 seconds on average. Our product detection solution is all powered with Google Cloud Compute Engine. It allowed us to easily put our solution into production.
We are very satisfied to have successfully designed the product detection solution from the initial reflection to its production release. Today, the scan is used more than 500 times per month on the Agrizone mobile application before we widely communicate about it. These technologies for product search with a photo are new and innovative, and we believe that they align with the changing behavior of our customers, who are increasingly connected.
"With the solution in place, we're expecting a substantial decrease in call center dependency and quicker customer transactions, which will fuel operational efficiency and offer a smoother, improved shopping experience."
—BILLION Olivier, Marketing Director, Groupe DubreuilNow, Groupe Dubreuil is considering expanding its product detection technologies from a photo to other subsidiaries. Skills acquired through the use of Google Cloud technologies, such as Vision AI or Compute Engine, allow us to envision the creation of a new solution for identifying parts in the automotive industry.
Tell us your challenge. We're here to help.
Contact usAbout Groupe Dubreuil
Groupe Dubreuil is a family-owned group based in the North West of France that has been in business for over 100 years diversified into 7 areas.
For 3 generations, they have been entrepreneurs, curious and determined, with a very clear vision of the Group that they want to shape.