Data augmentation: A comprehensive survey of modern approaches

Table of Contents

Title: Data augmentation: A comprehensive survey of modern approaches

Authors: Alhassan Mumuni, Fuseini Mumuni

Published: Dec 2022

Link: https://doi.org/10.1016/j.array.2022.100258

Summary (Generated by Microsoft Copilot):

Introduction:

  • The paper reviews modern data augmentation techniques in computer vision to improve machine learning model performance.

Challenges:

  • Obtaining large, quality annotated data is time-consuming and resource-intensive.
  • Small and poorly representative datasets lead to poor model performance.

Methods:

  • Data transformation and data synthesis techniques.
  • Meta-learning to enhance data augmentation processes.

Novelties:

  • New taxonomy categorizing methods into data transformation and data synthesis.
  • Extensive coverage of modern techniques like neural rendering and generative adversarial networks (GANs).

Results:

  • Quantitative performance comparisons of various methods on benchmark datasets like CIFAR-10, CIFAR-100, and ImageNet.

Performances:

  • Significant performance improvements with advanced data augmentation methods over baseline models.

Limitations:

  • Some methods may introduce artificial, out-of-distribution samples.
  • Geometric transformations may harm performance in controlled settings.

Discussion:

  • Emphasizes the importance of task-relevant augmentations.
  • Highlights future research directions and challenges in data augmentation.