This month, the Innovation Promotion Foundation selected SearchBooster, a subsidiary project of MediaNation, a smart search service for online stores, as the winner of a grant to develop product search by image as part of the Digital Economy national program. The project utilizes machine learning, neural networks, and computer vision algorithms. It will be available for implementation on websites by May next year. Information on the outcome of project selection and the winners was published on the foundation’s website. 

This new technology will be useful for online stores with a range of over 500 products. It will help website visitors pinpoint the products they need faster than through browsing an extensive catalog.

The SearchBooster project has Western competitors such as Multisearch, Algolia, and other smart search developers. However, due to recent events, they have stepped away from the market. Most of these systems only work with text search, excluding the possibility of selecting goods by visuals. There is only one similar Western patent, the Google Vision API. However, companies that use it will have to set up indexing and implement the feature on their web site by their own. SearchBooster, conversely, will offer a comprehensive, ready-to-install software product.

Other solutions are local tech used by certain online stores. In particular, Ebay implemented a similar feature on their website. It enables customers to use images to search for products among the 1.2 billion items in the catalog. In a year, all Russian e-commerce projects will be getting the opportunity to enjoy this feature.

Aiming to find a certain product, the user will be prompted to take a picture of the product or upload an existing one using the search bar. After that, ML and AI-based algorithms will analyze the uploaded photo and select the best fitting products in real time. The user will be able to take a picture of something on the street or in any other place, immediately find it in your store and place an order. The main object in the photo will be determined with at least a 95% accuracy, and the search speed will not exceed 10 seconds.

The project is planned for implementation in two stages:

Stage 1 (May — November 2022). Over the first half of the year, SearchBooster experts will teach smart algorithms to remove the background and identify the main object in the photo. Engineering solutions for clearing the image background and highlighting the main object will be implemented and tested. The best-fitting neural network architecture will also be selected, followed by training based on existing datasets. Then, the modules for determining the main object and removing the background will be integrated.

Stage 2 (December 2022 — May 2023). In the second half of the year, neural networks will learn to search for similar images based on vector distance. The agency’s experts will develop a search module that will receive a digital image and search for images of indexed goods closest to it. Then, they will hook up the resulting technology to the main searchbooster.io service and conduct A/B testing on real major clients. 

After trial runs with online stores, the service will enter the refinement stage; feedback from users will be collected, and the search by image feature will be improved.