Automated Image Labelling using Active Learning and Transfer Learning Technique

Author Name : Ankan Mazumdar, Sai Krishna Rohith K, Harish C R, Ananya Redhu, Dhruv Dixit, Thadagath Bollabathini, Naveen Kumar Madhan

ABSTRACT

Data labeling can be a very time- and/or money-consuming procedure. For this, a domain specialist is occasionally required. Active Learning is an approach which uses fewer training data to achieve better optimization by iteratively training a model. We can use a classification model to perform the majority of the labeling using active learning, requiring us to label samples only when absolutely necessary. Active learning seeks to address this issue by asking to annotate only the most informative data from the unlabeled set. If given the freedom to choose which data to label, the active learning methodology has the potential to significantly reduce the amount of labeled data needed to train a model with improved accuracy. The data that the model is most unsure of is prioritized through active learning, and labels are only requested for those. As a result, the model picks up new information more quickly.

This is a very important problem ,the reason being, in real world situations, we might want to go ahead to obtain manual labels for new data, but we have constraints on how much labeled data we can actually obtain within a certain time & limited budget.