Reproduce “A Data Driven Approach for Compound Figure Separation Using Convolutional Neural Networks” on OpenResearch

About this paper

Satoshi Tsutsui, David Crandall

Researchers at Indiana University developed the first data-driven, Convolutional Neural Network-based system for separating subfigures within compound scientific figures, achieving 85.9% accuracy on the ImageCLEF 2016 dataset and 91.7% on ImageCLEF 2015, surpassing traditional rule-based methods. The system also offers a publicly released, fast Python library for wider research use.

Key insights

• Transfer learning from a large natural image dataset (ImageNet) significantly improved performance on the specialized domain of scientific figures, boosting mean Average Precision (mAP) from 0.698 to 0.773.
• Generating realistic, grid-structured synthetic compound figures effectively augmented the training data and further enhanced the detector's accuracy, demonstrating the importance of domain-aware data synthesis.
• CNN-based object detection provides a robust and efficient solution for compound figure separation, eliminating the need for brittle manual feature engineering and achieving processing speeds of 0.12 seconds per figure on a GPU.

Summary by alphaXiv.