An Empirical Study of Scaling Law for OCR

TL;DR

The paper presents an empirical study of scaling laws in Optical Character Recognition (OCR) by tuning model size, data volume, and compute for Transformer-based scene text recognition. It introduces the REBU-Syn dataset (6M real, 18M synthetic; plus 60M MJST+), and evaluates two architectures, TrOCR and PARSeq, across a broad size and data spectrum. The authors demonstrate smooth power-law relationships: performance scales with $N$, $D$, and $C$ under fixed conditions, and highlight that larger models use samples more efficiently, that data-source mix and task-specific pretraining substantially affect OCR performance, and that data scaling can yield large gains at lower costs. The approach yields a new state-of-the-art of 97.42% top-1 average accuracy on six public benchmarks, underscoring the practical value of scaling laws for OCR and guiding future resource allocation and dataset design.

Abstract

The laws of model size, data volume, computation and model performance have been extensively studied in the field of Natural Language Processing (NLP). However, the scaling laws in Optical Character Recognition (OCR) have not yet been investigated. To address this, we conducted comprehensive studies that involved examining the correlation between performance and the scale of models, data volume and computation in the field of text recognition.Conclusively, the study demonstrates smooth power laws between performance and model size, as well as training data volume, when other influencing factors are held constant. Additionally, we have constructed a large-scale dataset called REBU-Syn, which comprises 6 million real samples and 18 million synthetic samples. Based on our scaling law and new dataset, we have successfully trained a scene text recognition model, achieving a new state-ofthe-art on 6 common test benchmarks with a top-1 average accuracy of 97.42%. The models and dataset are publicly available at https://github.com/large-ocr-model/large-ocr-model.github.io.

Figures (7)

• Figure 1: Mean word accuracy vs Parameters on the 6 common test benchmarks. P-Ti, P-S and P-B refer to PARSeq-Ti, PARSeq-S and PARSeq-B, respectively. * indicates training with REBU-Syn.
• Figure 1: Error analysis of the Union14M benchmark. We select three representative models and show their prediction results (Text in black represents correct prediction and red text vice versa).
• Figure 2: Improvement in TrOCR model performance with increasing model size, data volume, and training computation. Model performance is measured by calculating the average word error rate on 6 common test benchmarks Left: Evaluation of model performance with changing model sizes. Center: Evaluation of model performance with varying data volumes. Right: Performance variations with different data sizes under varying computational resources. The x-axis represents the model's training time, measured in 8 GPU hours. For optimal performance, all three factors must be scaled up in tandem. Empirical performance exhibits a power-law relationship with each individual factor when it is not constrained by the other two factors.
• Figure 2: Visual anwer comparison for QWen-VL-Chat with or without OCR in natural scenes VQA.
• Figure 3: Visual anwer comparison for QWen-VL-Chat with or without OCR in Document VQA.