EXPLORATION OF DEEP GENERATIVE MODELS FOR ADDRESSING DATA IMBALANCE IN COMPUTER VISION (Conference Paper)

Abstract

Computer vision tasks primarily depend on image data collection, preprocessing of the images, and the feature extraction steps to accomplish the task. When the collected images count is highly imbalanced within a category or between multiple categories involved, the anticipated tasks cannot be achieved. Data imbalance is a communal tricky problem that occurs in acquiring image data collection in specific complex real-world scenarios such as medical diagnosis, fraud detection, manufacturing defect detection, object detection, biological research etc., are inevitable. Several traditional methods for class imbalance are available but more advanced data generation models show promising results. The effectiveness of the computer vision algorithms has an inverse relationship to the quality and quantity of the dataset. In recent years, data generation models like Generative Adversarial Network, Variational Autoencoders, have gained enormous attention from researchers across domains, due to their potential for learning to generate images and their effectiveness in restoring balance in imbalanced datasets. They not only enhance the dataset volume but also contribute to the generalization and robustness of the machine learning algorithms. Moreover, data generation models play a vital role in improving classification accuracy, object detection and semantic segmentation tasks in deep learning. The survey is done with an intention to deliver a comprehensive study of data generation models, their learning representations from data, and their efficiency to address various challenges related to dataset quality, size and diversity.