Hi, Morning@Lohikans,

We were happy to see you at Recognition in the Wild Morning@Lohika. Thanks to our speakers Oles Dobosevych (Deputy Dean, Faculty of Applied Science @ Ukrainian Catholic University) & Kostiantyn Liepieshov (Research Intern, Machine Learning Lab @ Ukrainian Catholic University, co-founder of Footty and 3rd-year student at Faculty of Applied Sciences @ Ukrainian Catholic University) from Lviv.

 

As usual, the event took place at Lohika Lviv office

Please, share you feedback filling in this feedback form 

Oles Dobosevych

 Oles Dobosevych is Deputy Dean, Faculty of Applied Science @ Ukrainian Catholic University. Oles has more than seven years of experience in IT industry that varies from web development to personal analytics. Currently, Oles is Head of Data Science at Geniusee, completing Ph.D. thesis in Functional Analysis and teaching at Ukrainian Catholic University. His professional interests are machine learning and functional analysis.

Link to web profile: https://www.linkedin.com/in/dobosevych/

Kostiantyn Liepieshov

Kostiantyn Liepieshov is Research Intern, Machine Learning Lab @ Ukrainian Catholic University, co-founder of Footty (the startup that at the moment is accelerating at Rockstar incubator at Amsterdam) and 3rd-year student at Faculty of Applied Sciences @ Ukrainian Catholic University. His professional interests are machine learning and computer vision.

Link to web profile: https://www.linkedin.com/in/inkognita/

 

On recognition of Cyrillic Text in the Wild

We was  taking about the process of developing of the FCN-UA – the first published network for scene-text recognition of Cyrillic text trained in multilanguage setup. The model solves the problem of recognition multiple languages in one setup.

The goal of the text recognition is to map the input image to a sequence of characters. There are many different existing approaches to this task. One of them was to create single character recognizers and then do the post-processing after finding all letters in the image. It was time-consuming and needed sophisticated approaches for letter merging. The others were based on the multiword classification mapping the input image to the big dictionary (90k dictionary of English words). To approach this problem we use trained a fully convolutional neural network that maps preprocessed image of the word to the sequence of characters finding out the language of the text simultaneously.

Please, share you feedback filling in this feedback form 

Here you will find the presentation of our speakers: https://bit.ly/2FUUHPI