EffOCR: An Extensible, Open-Source Package for Efficiently Digitizing World Knowledge

Abstract

Billions of public domain documents remain trapped in hard copy or lack an accurate digitization. Modern natural language processing methods cannot be used to index, retrieve, and summarize their texts; conduct computational textual analyses; or extract information for statistical analyses, and these texts cannot be incorporated into language model training. Given the diversity and sheer quantity of public domain texts, liberating them at scale requires optical character recognition (OCR) that is accurate, extremely cheap to deploy, and sample-efficient to customize to novel collections, languages, and character sets. Existing OCR engines, largely designed for small-scale commercial applications in high resource languages, often fall short of these requirements. EffOCR (EfficientOCR), a novel open-source OCR package, meets both the computational and sample efficiency requirements for liberating texts at scale by abandoning the sequence-to-sequence architecture typically used for OCR, which takes representations from a learned vision model as inputs to a learned language model. Instead, EffOCR models OCR as a character or word-level image retrieval problem. EffOCR is cheap and sample efficient to train, as the model only needs to learn characters’ visual appearance and not how they are used in sequence to form language. Models in the EffOCR model zoo can be deployed off-the-shelf with only a few lines of code and include lightweight models designed for mobile phones that are extremely cheap to deploy. Importantly, EffOCR also allows for easy, sample efficient customization with a simple model training interface and minimal labeling requirements due to its sample efficiency. We illustrate the utility of EffOCR by cheaply and accurately digitizing 20 million historical U.S. newspaper scans, evaluating zero-shot performance on randomly selected documents from the U.S. National Archives, and accurately digitizing a Japanese document collection for which all other OCR solutions failed.

1 Introduction

Vast document collections remain trapped in hard copy or lack accurately digitized texts. For example, the U.S. National Archives holds approximately 13.28 billion pages of textual records, most of which are in the public domain. These documents are preserved because they are central to the workings of the U.S. government, have long-term research value, or provide valuable information for the public, but working with most of them is costly and time-consuming. The U.S. National Archives are not unique: many other countries have national archives with public domain collections numbering in the billions of pages, not to mention state and local archives and libraries. Without accurate machine-readable data, modern natural language processing (NLP) tools cannot be used to index, retrieve, and summarize materials; conduct computational textual analyses; or extract information for statistical investigations. Public domain texts, if accurately digitized, could also provide massive scale information for training large language models, with no risks of copyright infringement.

Using optical character recognition (OCR) to digitize public domain collections on a large scale entails several challenges.

Cost: First, the OCR solution must be cheap to deploy, given document collections whose size numbers in the millions or even billions of pages. Commercial engines - as well as large open-source OCR models - fall well short of this requirement. Using them to digitize large-scale collections would require astronomical budgets.

Accuracy: Second, digitized texts need to be sufficiently accurate for end users’ objectives, which are highly diverse. Accuracy can be particularly central for quantitative applications, for which small errors can create major statistical outliers. Models for lower resource languages, if they exist, tend to perform much worse than models for high resource settings like English.

Sample efficient, easy training: Documents are highly heterogeneous in terms of their fonts or handwritings, languages, scripts, backgrounds, and artifacts from scanning and aging. There are a diversity of documents for which no existing OCR solution works zero-shot, particularly in low resource languages. Yet stakeholders who would like to digitize these documents rarely have familiarity with deep learning frameworks. Bringing high quality OCR to low resource settings requires a simple API for training and a sample efficient architecture, with an accessible compute and annotation burden.

A diversity of pre-trained and tuneable models: Users have diverse accuracy needs, scaling requirements, and budgets. A comprehensive OCR solution would make it easy to compare the accuracy and deployment costs of models of varying sizes so that users can choose the one that best suits their needs for a particular application.

To meet these objectives, we developed EffOCR, an open-source OCR package designed for researchers, libraries, and archives seeking a computationally and sample efficient OCR solution for digitizing diverse document collections. EffOCR has two key ingredients: 1) a novel OCR architecture and 2) a carefully designed interface to facilitate off-the-shelf OCR usage, customization via model training if necessary, and easy sharing of OCR models.

The novel EffOCR model architecture is treated in detail in Carlson et al. (2023), where we compare accuracy, sample efficiency, and deployment costs to a range of popular OCR engines. In short, OCR predominantly models text recognition as a sequence-to-sequence (seq2seq) problem, in which learned representations from a vision model are taken as inputs to a learned language model. Learning how vision embeddings are used in sequence to form language requires large amounts of data. For example, the predominant transformer sequence-to-sequence OCR package Li et al. (2021b) was trained on 684 million text lines using 32 32GB V100 GPU cards. State-of-the-art seq2seq OCR is sample-inefficient to tune and infeasible for users to extend to low resource languages, which may not even have a transformer large language model (LLM) that can be used to initialize the model, as language modeling advances are concentrated in less than two dozen languages Joshi et al. (2020). The typical stakeholder working with low resource documents has a minimal budget for training and limited experience with deep learning frameworks, underscoring the need for a much more sample efficient framework with an easy-to-use API.

Additionally, seq2seq OCR requires autoregressive decoding, which makes inference slower than it would be, all else equal, with parallel decoding.

EffOCR abandons the seq2seq OCR model that predominates in the literature, instead modeling OCR as a word or character level image retrieval problem. EffOCR first localizes words using highly accurate, scalable object detection methods Ultalytics (2023); Chen et al. (2019); Wu et al. (2019). Recognition is then modeled as a contrastively trained image retrieval problem, where image embeddings of the same character or word have similar representations, regardless of their style. EffOCR is trained primarily on digital fonts, combined with a modest number of character and word crops from real-world documents. At inference time, characters/words are recognized by computing their nearest neighbor in an offline dictionary of exemplar embeddings created with a digital font. Carlson et al. (2023) show, using English, Japanese, and Polytonic Greek benchmarks, that the EffOCR architecture is accurate, highly sample efficient, cheap to train, and extremely fast to deploy when using backbones designed for mobile phones.

To meet the challenges of digitizing large-scale and low-resource document collections, the EffOCR package contains the following components:

1. An off-the-shelf toolkit for applying OCR models with just a few lines of code

2. A repository of pre-trained OCR models that underlies off-the-shelf usage

3. ONNX runtime support for fast deployment

4. Comprehensive tools for efficient model tuning

5. Supports models from popular backends Chen et al. (2019); Ultalytics (2023) for initializing localization and any timm-supported model for initializing recognition

6. Easy sharing of models, to promote reusability, reproducibility, and extensibility

EffOCR has been extensively tested. For example, we have used it to cheaply digitize 20 million pages of historical public domain U.S. newspaper scans that are extremely heterogeneous, posting the massive-scale output to Hugging Face. Creating this dataset within our modest budget while meeting accuracy requirements would have been impossible without EffOCR. We have also examined performance in settings where no existing OCR solutions provide usable output, and tested zero-shot performance on a random selection of U.S. National Archive documents, with a model that did not see any similar content during training. Tutorials are available at https://effocr.github.io/.

EffOCR has a GNU General Public License. It is being actively maintained and crowd-sourcing of annotations to expand the pre-trained model zoo to other languages and settings, including handwriting, is underway.

The rest of this paper is organized as follows. Section 2 briefly compares EffOCR to existing, popular OCR solutions. Section 3 describes the key features of the OCR package, and Section 4 examines several use cases: using EffOCR to digitize 20 million historical newspaper scans, using EffOCR zero-shot on randomly selected collections from the U.S. National Archives, and using EffOCR to digitize a historical Japanese publication for which all existing OCR solutions fail. Finally, Section 5 discusses the limitations of the EffOCR package.

2 Comparisons to Other OCR Engines

There is a vast literature on OCR. Of primary interest here are widely used OCR softwares, which are the most plausible alternatives to EffOCR.

EffOCR- as the name suggests - is tailored towards applications requiring computational or sample efficiency. Carlson et al. (2023) conduct detailed experiments comparing the EffOCR architecture to other widely used solutions, considering accuracy, sample efficiency, and computational efficiency. We refer the interested reader to that paper for details, summarizing the two key themes that emerge here.

Customization is highly relevant: As the preponderance of researchers still using data entry firms suggests, sometimes no existing OCR solution provides acceptable accuracy. For typewritten Japanese documents from the mid-20th century, that are of considerable relevance to studying Japan’s remarkable 20th century growth performance, Carlson et al. (2023) show that the best performing engine (Baidu, the leading commercial OCR for Asian languages) gets over half of characters wrong. The widespread failure of OCR to provide acceptable results is also evidenced by a large post-OCR error correction literature (e.g., Lyu et al. (2021); Nguyen et al. (2021); van Strien. et al. (2020)).

EffOCR is significantly more sample efficient than leading open-source OCR engines: EasyOCR JaidedAI (2021), TrOCR Li et al. (2021b), and PaddleOCR Du et al. (2022), as shown in the supplementary materials. Learning to recognize the visual features of individual characters is a highly parsimonious problem, making EffOCR cheap to tune or train from scratch. Because EffOCR does not need to understand language, it is straightforward to extend to new languages and scripts, including those that lack a transformer large language model to initialize state-of-the-art seq2seq. The convolutional models in the EffOCR model zoo can be trained on a Google Colab account, whereas training TrOCR on 684 million text lines required 32 32GB V100 cards.

A central aim of EffOCR is to democratize access to OCR to low resource languages and settings that are difficult to study because existing solutions are not suitable to these use cases. While we do not have the resources to train OCR models for all these settings, our simple APIs for training models and uploading them to the EffOCR model hub can encourage the crowdsourcing of this effort.

The most accurate OCR engines in high resource settings (e.g., English) are costly to deploy at scale: TrOCR (Base) is a highly accurate state-of-the-art English OCR. With 334 million parameters, it is nearly 50 times slower to deploy than our pre-trained lightweight EffOCR English word recognition model, while offering only relatively modest gains on the evaluation tasks in Carlson et al. (2023). For English, Google Cloud Vision (GCV) - a proprietary commercial product - dominated all open-source solutions (including EffOCR), but would have been orders of magnitude more costly to deploy. In our experience, it is frequently outside academic budgets for larger projects.

Lightweight EffOCR models are also faster than Tesseract and PaddleOCR - with the comparison to EasyOCR depending on the hardware used for deployment. This is despite having around 8x more parameters than Tesseract and around 4x more than EasyOCR (parameter count is similar to PaddleOCR). This is achieved through parallel rather than sequential decoding and ONNX integration. EffOCR is also significantly more accurate on tasks like digitizing the 20 million U.S. historical newspaper scans.

Users for whom neither computational nor sample efficiency is of concern - because they are working in a well-resourced context and don’t face cost constraints for the scale of their problem - are not our target audience and may well find an existing OCR engine like Google Cloud Vision better meets their needs. In practice, academic or large-scale archival digitization of document collections often involves low-resource languages or settings, tight budget constraints, or both.

3 The EffOCR Library

3.1 Off-the-shelf Usage

At the core of EffOCR is an off-the-shelf toolkit. EffOCR is a modular framework, that first localizes lines, characters, and (for some models) words using object detection, and then recognizes characters and words by embedding their crops and retrieving their nearest neighbor from an offline index of exemplar embeddings created from a digital font.

Localization: EffOCR supports two widely used backends for localization inference: MMDetection Chen et al. (2019), which includes state-of-the-art object detection models, and Yolo Ultalytics (2023), which includes fast, efficient object detection models. Users can deploy line, word, and character models from the pre-trained model zoo, that use Yolo v8 Ultalytics (2023) (optimized for efficiency), Yolo v5 Jocher (2020) (fewer dependencies) or Cascade R-CNN Cai and Vasconcelos (2018) (optimized for accuracy). Pre-trained localization models are available for alphabetic English/Latin, Polytonic Greek, and CJK characters (which vary significantly in their aspect ratios and groupings).

Recognition: EffOCR recognizes word and character crops using contrastively trained image retrieval models. The EffOCR model zoo currently contains 30 pre-trained models, covering English, Polytonic Greek, and horizontally and vertically-written Japanese. We chose these languages to examine the utility of EffOCR in a high resource setting, in a setting where existing solutions fail, and in an intermediate case.

The EffOCR pre-trained models use a variety of backbones: two lightweight convolutional backbones that are very efficient to deploy Howard et al. (2019); Maaz et al. (2022), a state-of-the-art CNN encoder Liu et al. (2022), and three vision transformers Ali et al. (2021); Li et al. (2022); Liu et al. (2021). For English, there is a word level model that defaults to character recognition when the word is below a default (tuneable) cosine similarity threshold, as well as a character-only model.

The documentation provides more guidance on model selection. A description of the training dataset is provided alongside with the trained models such that users can quickly identify the most suitable models for their tasks.

EffOCR can be used off-the-shelf with just a few lines of code:

⬇

1import effocr

2engine = effocr.EffOCR(

3 line_detector = "./line_model",

4 localizer = "./localizer",

5 char_recognizer = "./char_recognizer",

6 word_recognizer = "./word_recognizer"

7)

8results = engine.infer(’image.jpg’)

ONNX ONNX (2021) integration is an important component, as it allows for efficient CPU deployment and interoperability between deep learning frameworks. All EffOCR stages can optionally employ ONNX-format models and ONNX-runtime inference, and models can be converted to ONNX format within the package. ONNX-runtime increases CPU throughput by up to four times Jocher (2020) for YOLO models used in EffOCR, which allows for cost-effective cloud deployment for processing large document sets. ONNX compatibility allows additional model speedups through graph optimizations, quantization, and pruning.

3.2 Customized Model Training

Many low-resource settings are poorly served by existing OCR engines, and a central aim of EffOCR is to democratize OCR for these settings by providing a simple interface for custom model training that can be used by researchers and others who have limited experience with deep learning frameworks. Custom training can be initialized using a Yolo object detection model for localization and any timm image encoder model for recognition. In the near future, support for training localization models with MMDetection will be added. This futureproofs EffOCR, as new models are developed.

EffOCR supports logging of a training run on Weights and Biases Biewald (2020). It takes industry standard coco json labels as inputs, and hence is compatible with the outputs of a range of both open-source and proprietary labeling softwares. It also exports output in the same format, so that users can easily correct model predictions if desired to speed up labeling.

Model training with EffOCR is highly efficient, e.g., the convolutional backbones can be trained on Google Colab. We trained all models on either a single Nvidia RTX 3090 or A6000 card.

3.3 Visualization, Storage and Export

EffOCR comes with a tool to visualize the OCR, side-by-side with the original image, as well as to visualize the line, word, and character predictions. These greatly facilitate quality checking the output and troubleshooting potential problems.

EffOCR offers users different options for data export. The default outputs of EffOCR include line coordinates, word coordinates, character coordinates, and the text associated with each of these annotations. The text for the full image is also assembled in the correct order. Users may choose to export only the assembled text, only text annotations associated with a given level of bounding box (line, word, or character), or all of the above.

3.4 User Contributions

By making OCR sample efficient and easy to train, EffOCR aims to promote the reusability and reproducibility of OCR pipelines. This is particularly important for low resource settings and languages, where there is little commercial incentive for product development and few alternatives to crowd-sourcing models. EffOCR users can upload their self-trained models to the EffOCR Hugging Face hub. Whenever a model is saved, a model card is automatically generated that follows best practices outlined in Hugging Face’s Model Card Guidebook. Moreover, the automatically generated card contains instructions on how to use the model in the context of EffOCR and model-specific architecture and training details in the interest of reproducibility.

3.5 Integration with Layout Parser

OCR engines typically detect lines, versus detecting and classifying different layout objects in a document. Many documents have complex layouts - e.g., newspapers have headlines, articles, captions, ads, and headers arranged in complex multicolumn layouts, and tables likewise have different types of information arranged in oftentimes complex layouts. These structures necessitate applying object detection models for document layout analysis, which have been trained to detect the coordinates of each layout object and classify its type (e.g., headline, articles, etc).

To facilitate combining EffOCR with deep learning-based document layout models, wrappers will be integrated into a popular open-source layout detection package, Layout Parser Shen et al. (2020), that will allow Layout Parser users to call any EffOCR model. Layout Parser also has wrappers to call GCV and Tesseract, which will allow users to easily compare EffOCR output to these other packages to decide what best meets their accuracy and cost objectives. Layout Parser and EffOCR were designed by the same lab, facilitating long-run coordination between the packages.

4 Applications

Figure 1: This figure shows a diversity of examples processed with EffOCR, with predicted transcriptions on the right.

Scalability: We have tested the utility of EffOCR with various real-world applications. In the first application, we cheaply and accurately digitized 20 million newspaper page scans from Library of Congress’s Chronicling America collection Library of Congress (2022). The resulting dataset, American Stories, is available for download on Hugging Face.. Figure 1 illustrates why this is a challenging task: newspapers are extremely heterogeneous in their fonts and image quality. Dell et al. (2023) provide a detailed analysis of the quality of the resulting text dataset.

We first trained character EffOCR using synthetic data plus a labeled set of 291 newspaper lines Carlson et al. (2023), created in a couple of hours. We then bootsrapped word level annotations by creating them with the character level EffOCR model, filtering out lines with a high non-word rate.

With EffOCR, combined with layout analysis using Layout Parser, we could digitize the dataset with a $60K USD cloud compute budget (plus pipeline development costs). GCV makes significant layout errors when fed full newspaper scans and achieves best performance when fed individual lines. At current prices, digitizing the collection at the line level, since GCV charges per image, would have cost over $23 million USD. TrOCR Base, the most accurate open-source OCR, would have exceeded our budget by a factor of nearly 50.

Zero-Shot Performance: Second, we show that our English lightweight word-level model has strong zero-shot performance on randomly selected document collections from the U.S. National Archives. This model saw only newspapers in training, to test true zero shot performance. We selected a single textline from each of 300 random documents from separate National Archive record groups. EffOCR achieved a 11.2% CER on the diverse collection, compared with a 11.8% CER from Tesseract (Best), a 12.1% CER from EasyOCR, and a 51% CER from TrOCR (Small), which appeared to struggle with blurry and partially obscured text. We suspect the results could be significantly improved by including a random sample of documents from the National Archives in training, to broaden the set of real world documents that the model is exposed to.

All open-source models performed significantly worse than GCV (1.2% CER), but as discussed earlier cost concerns presently preclude its use at scale. Despite being engineered for low-resource, few-shot contexts, EffOCR remains competitive in high-resource, zero-shot situations.

Low Resource Settings: Finally, we use EffOCR to digitize historical Japanese firm level records for vertically written Japanese documents Teikoku Koshinjo (1957), where the best available solution (from Baidu OCR) mispredicts over half of characters. We use the evaluation set in Carlson et al. (2023), which consists of randomly selected segments that were double labeled.

Using a training set of 898 labeled table cells, we achieve a CER of 0.7%, 80 times more accurate than the best existing solution. As a result, we are able to study a variety of questions about Japan’s remarkable growth performance that would have been impossible to examine without EffOCR.

To further examine the limits of sample efficiency, we calculate the character classification error when the (character) model only sees one (or up to 5) labeled character(s) for each of the characters that appear in the training set, which comprise 77% of the characters in the test set. This results in character classification errors of 13.4% and 2.0% respectively. While the model does clearly benefit from seeing multiple crops of characters that appear frequently, this illustrates viable few shot performance.

5 Limitations

If large portions of a document are illegible, vision-only OCR will not be suitable and language understanding may be helpful for inferring content. For high resource languages such as English when cost is not a concern, users may get the best mileage from a leading commercial product such as GCV.

Currently, the EffOCR model zoo has pre-trained models supporting typewritten English, Japanese, and Polytonic Greek. Over the coming months, we will be crowd-sourcing annotations (including handwriting) from package users and colleagues. We will use them, along with digital fonts, to pre-train additional models. In addition, users are encouraged to contribute their models.

EffOCR does not currently support handwriting. We started with typewritten documents because there are billions of public domain typeface documents that are of considerable interest to researchers and the general public. We are planning to expand the model zoo to include handwriting and users have already offered to contribute annotations. Synthetic handwriting generators, e.g. Bhunia et al. (2021), can provide extensive data for pre-training for scripts that they support, analogous to the use of digital fonts for typeface documents. We will make synthetic handwriting datasets available so that package users can also use them for training their own custom models.

Supplementary Materials

S-1 Model Architecture and Model Zoo

Figure S-1 shows the EffOCR model architecture, and Table S-1 summarizes the models in the EffOCR model zoo. Readers seeking technical details for the EffOCR models contained in the pre-trained model zoo are referred to the detailed supplementary materials in Carlson et al. (2023).

S-2 Sample Efficiency

To examine how efficiently EffOCR learns in comparison to leading open source architectures, we train different OCR models from scratch using varying amounts of annotated data. EffOCR-C (Base) is compared to SVTR (implemented via PaddleOCR) Du et al. (2022), CRNN (implemented via EasyOCR) Shi et al. (2016), and TrOCR Li et al. (2021b). All architectures are pre-trained from scratch on 8,000 synthetic text lines, starting from pre-trained checkpoints not customized for OCR when supported by the framework. They are then fine-tuned on the study’s benchmark datasets, with varying train-test-validation splits: 70%-15%-15%, 50%-25%-25%, 20%-40%-40%, 5%-47.5%-47.5%, and 0%-50%-50% (i.e., zero-shot). These exercises are performed for the English newspaper character level models and horizontal Japanese, as vertical Japanese is not supported by the comparison architectures.

Figure S-2 plots the percentage of the benchmark dataset used in training on the x-axis and the CER on the y-axis. On just 99 labeled table cells for Japanese and 21 labeled rows for LoCCA (the 5% train split), EffOCR’s CER is only 5% (Japanese) and 7% (English), showing viable few shot performance. The other architectures remain unusable. EffOCR performs nearly as well using 20% or training data as using 70%, where it continues to outperform all other alternatives. This illustrates that its parsimonious architecture learns efficiently.

S-3 Training Config Details

The EffOCR package exposes a wide variety of training options and hyperparameters to users. A few key elements are described here, readers looking for more details are directed to the package documentation.

Recognizer Training Options:

• timm_model_name Model name from timm Wightman (2019) package used as a base encoder for the recognizer.

• render_dict Folder to store crop renders and gold training data locally.

• font_dir_path Local path to draw tff (font) files from, which are used to create character/word renders.

• hns_txt_path Local file path to draw hard negative samples from. Hard negative text files are created by default at the end of recognizer training. Most recognizer training applications use two stages, an initial run and a hard negative sampling run.

• latin_suggested_args Uses default arguments for alphabetic writing systems such as Latin, Greek, and Cyrillic.

In addition to these options, a wide variety of standard model training parameters are exposed, including learning rate, optimizer options, weight decay, batch size, device selection, and number of training epochs.

Localizer Training Options:

• vertical Whether model should expect characters aligned horizontally (as in English and many Latin scripts) or vertically (as in many character-based scripts).

• no_words Detect only characters, not words. Recommended for languages without word groupings.

• iou_thresh Training and validation IOU threshold for character/word detection.

• conf_thresh Training and validation confidence threshold for character/word detection.

As with the recognizer, other standard training parameters are exposed. In particular, adjusting the image input shape may be valuable for particularly long or short lines.

Hyperparameters and training procedure used to generate models listed in the Model Zoo (Table S-1) are listed in Carlson et al. (2023).

S-4 Visualization

Figure S-3 shows the EffOCR visualization interface.

S-5 American Stories

Figure S-4 plots the number of articles in the American Stories dataset, created with EffOCR, across time.

S-6 The Public Domain

Table S-2 provides detailed information about the requirements for information published in the United States to be in the public domain, in order to give readers a better sense of these collections.

S-7 Inference Speed

EffOCR implements two features designed to increase computational efficiency. First, both localization and recognition inference is run in a multithreaded fashion, ensuring that compute resources are fully utilized. Second, EffOCR provides support for ONNX runtime and ONNX-format models, which provide up to a 3x speedup on a CPU compared to native PyTorch runtime ONNX (2021). GPUs are typically cost prohibitive for digitization at scale.

Table S-3 provides a comparison between EffOCR and other commonly used OCR frameworks’ python implementations. It is important to note that these numbers - across softwares - can vary significantly dependending on the hardware resources available. All comparisons are made on four 2200 MHz CPU cores, selected to represent a plausible and relatively affordable research compute setup. EffOCR performance is competitive with other widely used frameworks, with EffOCR (Small) having the fastest performance. Tesseract Ooms (2023) testing used the pytesseract package with default settings. EasyOCR JaidedAI (2021) testing used the easyocr package with default English settings. PaddleOCR PaddlePaddle (2022) testing used the paddleocr package with use_angle_cls option and default English settings. TrOCR Li et al. (2021a) testing used the transformers package implementation, with trocr-base-printed and trocr-small-printed models for Base and Small tests, respectively. EffOCR testing used default settings with pretrained ONNX English newspaper models from the model zoo.

Figure S-1: EffOCR and Seq2Seq Model Architectures. This figure represents the EffOCR architecture, as compared to a typical sequence-to-sequence OCR architecture.

Figure S-2: Sample Efficiency. This figure plots the percentage of the benchmark dataset used in training against the character error rate, for different OCR model architectures: CRNN (EasyOCR), SVTR (PaddleOCR), TrOCR (Transformer OCR), and EffOCR small and base convolutional models.

Figure S-3: Visualization. This figure shows the EffOCR visualization interface.

Figure S-4: American Stories. This figure plots the number of articles in the American Stories dataset, created with EffOCR, across time.

Training Set	Line Detection	Localizer		Word Recognition		Character Recognition
	YOLO	YOLO	MaskRCNN	MobileNetV3	EdgeNeXt	MobileNetV3	EdgeNeXt	ViT	ConvNeXt	XCiT
English Newspapers	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓
English Mixed Archival	✓	-	✓	✓	-	✓	-	-	-	-
Japanese Vertical	✓	✓	✓	N/A	N/A	✓	✓	✓	✓	✓
Japanese Horizontal	✓	✓	✓	N/A	N/A	✓	✓	✓	✓	✓
Polytonic Greek	✓	✓	-	N/A	N/A	✓	-	-	✓	-

Table S-1: Models Currently Available in the EffOCR Model Zoo. Note Japanese models do not use word-level recognition.

Date of Publication	Conditions	Copyright Term
Public Domain
Anytime	Works prepared by an officer/employee of the	None
	U.S. Government as part of their official duties
Before 1928	None	None. Copyright expired.
1928 through 1977	Published without a copyright notice	None. Failure to comply with required formalities
1978 to 1 March 1989	Published without notice and	None. Failure to comply with required formalities
	without subsequent registration within 5 years
1928 through 1963	Published with notice	None. Copyright expired
	but copyright was not renewed
Copyrighted
1978 to 1 March 1989	Published without notice, but with	70 (95) years after the death of author (corporate author)
	subsequent registration within 5 years
1928 through 1963	Published with notice	95 years after publication
	and the copyright was renewed
1964 through 1977	Published with notice	95 years after publication
1978 to 1 March 1989	Created after 1977 and published with notice	70 (95) years after the death of author (corporate author)
		or 120 years after creation, if earlier
1978 to 1 March 1989	Created before 1978 and first published	The greater of the term specified in the previous entry
	with notice in the specified period	or 31 December 2047
From 1 March 1989 through 2002	Created after 1977	70 (95) years after the death of author (corporate author)
		or 120 years after creation, if earlier
From 1 March 1989 through 2002	Created before 1978 and	The greater of the term specified in the previous entry
	first published in this period	or 31 December 2047
After 2002	None	70 (95) years after the death of author (corporate author)
		or 120 years after creation, if earlier

Table S-2: This table summarizes U.S. copyright law, based on a similar table produced by the Cornell libraries. For concision, we focus on works initially published in the United States. A variety of other cases are also covered at https://guides.library.cornell.edu/copyright.

Model	Textline/s	Article/s
EffOCR Base	0.46	0.02
EffOCR Small	21.07	1.08
Tesseract	4.47	0.21
EasyOCR	19.80	1.03
PaddleOCR	13.56	0.61
TrOCR (Base)	0.43	0.02
TrOCR (Small)	0.97	0.05

Table S-3: Comparison of EffOCR speeds with other popular OCR frameworks in CPU environment. Tests included both Textline (single lines of text) and Article (5-40 lines of text) examples.