A Scalable Logo Recognition Model with Deep Meta-Learning

Author
Jackie Brusch
Publication Date
5 December 2019

A Scalable Logo Recognition Model with Deep Meta-Learning

You’re at a conference, there’s tons of information to obtain – but where will you store all of this great learning? Mobiquity’s Data Scientist, Mark de Blaauw, with the support of Senior Data Scientist, Vesa Muhonen, and one other researcher from the Vrije Universiteit, Amsterdam, the Netherlands, set out to achieve the answer.

In their poster, “A scalable logo recognition model with deep meta-learning,” presented at BNAIC & Benelearn 2019 conference, Mark and his co-authors set out to create a deep learning (DL) logo classifier that is widely applicable for conferences. DL models typically require many samples to learn a classification task, whereas deep meta-learning models can learn different but related tasks with 1-10 training samples per class. 

Additionally, current techniques in the meta-learning domain focus primarily on accuracy, although for AI-driven applications, multiple factors play an important role: (i) robustness of accuracy to different tasks and (ii) robustness to in- and out-domain unknown class samples. 

In this poster, the team shows (i) that deep metric learning (DML) approaches are more robust to tasks with more classes than initialisation learning (IL) approaches and (ii) that Gaussian prototypical networks (GPN) are more robust to in- and out-domain unknown class samples than Reptile.

See how the team explains the steps they took to create a meta-learning classifier that can learn different sets of logo classes with only five training samples per class. Plus, see the demo application in the AWS cloud environment in which the team is deploying the Mahalanobis prototypical networks on a lambda function.

Download Model Poster

Leave your details below to access the Scalable Logo Recognition Model poster