Data Augmentation: A Combined Inductive-Deductive Approach Featuring Answer Set Programming

Abstract

The availability of large amounts of data is often taken for granted; however, there are significant scenarios where this is not the case. For instance, in the biomedical and healthcare domains, certain applications require extensive datasets of high-quality images. However, acquiring such images is often challenging due to various factors, including accessibility, costs, and pathology-related variability. As a result, datasets tend to be limited and typically imbalanced. To address this challenge, synthesizing photo-realistic images through advanced data augmentation techniques is essential. In this paper, we propose a hybrid inductive-deductive approach to this problem. Specifically, starting from a limited set of real labeled images, our framework leverages logic programs to declaratively specify the structure of new images. This ensures compliance with both domain-specific constraints and desired properties. The generated labeled images then undergo a deep learning-based process to create photo-realistic images that accurately adhere to the generated labels.