 |
Botond EgriVP of AI and Data Science
|
 |
Aditya KurudeDeep Learning Computer Vision Team Lead
|
Driving Logo Search with User Interactions: A Deep Learning Approach
Corsearch is a leading provider of trademark solutions, offering search and watch capabilities to its customers. Among the most valuable properties of a company are its registered figurative trademarks, or logos. With over 100 million logos to search through, finding the most similar ones remains a significant technical challenge. While manual annotations and industry-wide standardized encoding hierarchies are available to aid the search process, they still rely heavily on human efforts. One of the major obstacles to full automation is the ability to guide the search towards the most important parts, concepts, features, or properties of a logo. To tackle this issue, Corsearch AI team has researched different approaches to incorporating user interactions into the search process. In this case study, we will discuss the existing challenges in the logo search industry. Then we’ll introduce our in-house trained deep learning model, paying particular attention to its architectural improvements. Finally, we will showcase various visual tools, we have implemented and demonstrate that the trained model can utilize user inputs to guide the search in desired ways.
BIO
Botond Egri has 20+ years of experience in data science in different industries and on different levels.
He is a machine learning enthusiast: believe that creative, up-to-date and hands-on knowledge of data science technology is a game changer in the business. He is a Kaggle fan, a Kaggle Competition Expert and the host of the Kaggle Days Meetups in Hungary.
Aditya is AI practitioner with particular love for visual domain, he finds it fun to optimize different models to get real world problems solved. He has worked in deep learning domain for almost 6 years.