Learning Semantic Concepts from Weakly Labeled Data

Hicsonmez S., Rezazadeh I., Unal D., Yaruktepe D., DUYGULU ŞAHİN P.

27th Signal Processing and Communications Applications Conference (SIU), Sivas, Turkey, 24 - 26 April 2019 identifier identifier

  • Publication Type: Conference Paper / Full Text
  • Volume:
  • Doi Number: 10.1109/siu.2019.8806477
  • City: Sivas
  • Country: Turkey
  • Hacettepe University Affiliated: Yes


As in many areas with deep learning methods, great success is achieved in object and scene recognition, but for a good result a large number of labeled data is needed. Our aim in this study is to eliminate the need for tagged data collection that requires too much human labor, and to use the abundant images as data set, paired with the object name on the Internet and social media. However, this data is not as clean as manually labeled data, and it does not work well when used directly. In this study, Association with Model Evolution (AME) method is adapted to eliminate noisy data. Data that were automatically collected and cleaned with AME were then used as the experimental set for Convolutional Neural Networks (CNN). It is observed that the performance is increased by 4% with using the AME cleaned data.