A survey on generative adversarial networks for imbalance problems in computer vision tasks

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Imbalanced image datasets are a significant challenge in computer vision, negatively impacting algorithm performance.
• Real-world applications like anomaly detection, medical imaging, and defect detection frequently encounter inevitable data imbalances.
• Existing solutions for data imbalance often fall short in addressing complex, real-world image data.

Purpose of the Study:

• To survey recent advancements in Generative Adversarial Network (GAN)-based techniques for mitigating imbalance problems in image data.
• To provide a comprehensive overview of synthetic image generation using GANs for addressing data scarcity and imbalance.
• To categorize and analyze GANs' applications in tackling image, object, and pixel-level imbalances within computer vision tasks.

Main Methods:

• Reviewing deep generative image models and the core concepts of Generative Adversarial Networks (GANs).
• Proposing a taxonomy to classify GANs-based techniques into three categories: image-level (classification), object-level (detection), and pixel-level (segmentation) imbalances.
• Analyzing specific GANs-based solutions tailored to each imbalance category.

Main Results:

• GANs demonstrate significant potential in generating synthetic images to balance datasets.
• Adversarial learning within GANs effectively addresses data imbalance issues across various computer vision tasks.
• The proposed taxonomy provides a structured understanding of GANs' role in resolving different levels of image data imbalance.

Conclusions:

• GANs are a promising approach to overcome data imbalance challenges in computer vision.
• By generating synthetic data and leveraging adversarial learning, GANs can substantially improve the performance of computer vision algorithms.
• This survey highlights the versatility and effectiveness of GANs in addressing critical issues in image classification, object detection, and segmentation.