En | Ru

Neural network search for similar trademarks «Gardium»

Computer vision

A comprehensive image analysis system based on computer vision technology:

Solution

Task

Search for similar images in our own database of registered trademarks by class according to the International Classification of Goods and Services (ICGS).

Client

«Gardium» — Patent and Law Office

A neural network for searching for similar images

Image filter for allocation to the classes of the ICU

An algorithm for analyzing and marking up images in the client base

The service helps avoid litigation over intellectual property rights, as well as fight plagiarism and patent trolls.

How it works

01/

A company asks Gardium to register a trademark: Sends a photo or PDF file

A neural network searches for similar images among those selected by the algorithm

"Gardium" assesses the uniqueness of the trademark, begins registration

The algorithm selects images in the database based on the specified class

"Gardium" loads the image into the search, specifies the class of the ICU

Details

02/

When searching for similar images, the neural network takes into account colors, shapes, secondary elements (e.g., background), and other attributes. The top of the list displays the images with the greatest number of matching features.

Project challenges

03/

Problem

Solution

Lack of marked-up data for training and testing the neural network.

An algorithm that classifies images independently. Using its markup, we trained a neural network.

Stages of development

04/

Data exploration (marking images in the client's database for future searches)

Step 01

Testing hypotheses about the search, selection of effective computer models

Step 02

Training a neural network on data marked up by an algorithm

Step 03

Implementing a classifier to assign images to classes

Step 04

Testing the service inside EORA and on the client side

Step 05

Development and service integration with the customer base via API

Step 06

Examples of system operation

05/

Project Team

06/

Project Manager

Alexey Guchko

Data scientist

Vyacheslav Schultz

Teamlead

Vladislav Vinogradov

Ivan Izmailov

Backend Developer

Quote

07/

«We didn't have pre-labeled data to help the neural network understand which choices are correct and which are not. To train the neural network, we used the unsupervised learning method. The result exceeded our expectations. The search accuracy was very high, and the client was satisfied.»