Evaluation of deep convolutional nets for document image classification and retrieval

By Adam W. Harley, Alex Ufkes, Konstantinos G. Derpanis. IEEE International Conference on Document Analysis and Recognition, 2015

468 citations recorded by the source corpus.

This paper presents a new state-of-the-art for document image classification and retrieval, using features learned by deep convolutional neural networks (CNNs). In object and scene analysis, deep neural nets are capable of learning a hierarchical chain of abstraction from pixel inputs to concise and descriptive representations. The current work explores this capacity in the realm of document analysis, and confirms that this representation strategy is superior to a variety of popular handcrafted alternatives. Extensive experiments show that (i) features extracted from CNNs are robust to compression, (ii) CNNs trained on non-document images transfer well to document analysis tasks, and (iii) enforcing region-specific feature-learning is unnecessary given sufficient training data. This work also makes available a new labelled subset of the IIT-CDIP collection, containing 400,000 document images across 16 categories.

Subfields: Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/bd86b4b551b9d3fb498f62008b037e7599365018

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

1. Furu Wei
2. Chenliang Li
3. Lei Cui
4. Songfang Huang
5. Ming Yan
6. Bin Bi
7. Sangdoo Yun
8. Wei Wang

Citations in the corpus

• LayoutLM: Pre-training of Text and Layout for Document Image Understanding (2019, methodology): “The third is the RVL-CDIP dataset5 [8] for document image classification, which consists of 400,000 grayscale images in 16 classes.”
• LLaVAR: Enhanced Visual Instruction Tuning for Text-Rich Image Understanding (2023, methodology): “Subsequently, we trained an image classifier using a DiT [35]-base backbone, which was fine-tuned on the RVL-CDIP dataset [36].”
• FUNSD: A Dataset for Form Understanding in Noisy Scanned Documents (2019, methodology): “To ensure that real data are used, featuring highly varying form structures and realistic noise, we used a subset of the RVL-CDIP dataset5 [10].”
• LayoutLMv2: Multi-modal Pre-training for Visually-rich Document Understanding (2020, methodology): “RVL-CDIP RVL-CDIP (Harley et al., 2015) consists of 400,000 grayscale images, with 8:1:1 for the training set, validation set, and test set.”
• LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking (2022, methodology): “It is a subset of the IIT-CDIP collection labeled with 16 categories [15].”
• WebSRC: A Dataset for Web-Based Structural Reading Comprehension (2021, background): “, 2006) and RVL-CDIP (Harley et al., 2015) are two datasets collected for document classification.”
• OCR-Free Document Understanding Transformer (2021, methodology): “Earlier VDU attempts have been done with OCR-independent visual back-bones [27,1,15,12,31], but the performances are limited.”
• StructuralLM: Structural Pre-training for Form Understanding (2021, methodology): “Finally, we evaluate the document image classification task using the RVL-CDIP dataset (Harley et al., 2015).”