Analysis of Convolutional Neural Networks for Document Image Classification

TL;DR

This paper empirically investigates how CNN architectures and parameters affect document image classification performance, surpassing state-of-the-art results by tailored data augmentation and input size adjustments, and analyzing learned features.

Contribution

It provides a comprehensive empirical study on CNN design choices for document images, highlighting effective augmentations and architectural adaptations.

Findings

Shear transform data augmentation improves accuracy.

Larger input images enhance classification performance.

CNNs learn region-specific layout features.

Abstract

Convolutional Neural Networks (CNNs) are state-of-the-art models for document image classification tasks. However, many of these approaches rely on parameters and architectures designed for classifying natural images, which differ from document images. We question whether this is appropriate and conduct a large empirical study to find what aspects of CNNs most affect performance on document images. Among other results, we exceed the state-of-the-art on the RVL-CDIP dataset by using shear transform data augmentation and an architecture designed for a larger input image. Additionally, we analyze the learned features and find evidence that CNNs trained on RVL-CDIP learn region-specific layout features.