Synthetic Randomized Image Augmentation (SRIA) to Address Class Imbalance Problem

With emerging deep learning technologies, the issue of class imbalance has become more relevant specially for object detection problem as performance of modern object detectors is more likely to suffer from class imbalance than traditional machine learning methods. Foreground-foreground class imbalance particularly affects deep learning models when trained from dataset following real world data distribution. To bridge this gap, this work overviews the limitations and capabilities of existing techniques to balance classes and proposes synthetic randomized image augmentation (SRIA) technique to address the class imbalance problem. The proposed technique comprises of two steps. First, objects from minority class training image are extracted using weak learning and subsequently new minority class images are synthesized by leveraging background substitution and data augmentation. We evaluate the proposed technique on state-of-the-art object detector i.e Yolov5 which shows performance improvement by 5% when compared with vanilla model trained on imbalanced data.