Skip to content
Open access

A Comprehensive Evaluation of Data Augmentation Strategies for Offline Handwritten Text Recognition Across Deep Learning Architectures

2026 · International journal on emerging technologies · 0 citations

Abstract

Offline handwritten text recognition (HTR) remains one of the most challenging problems in document image analysis because of the enormous variability of individual handwriting styles, inconsistent stroke widths, irregular slant and the scarcity of large annotated corpora. Data augmentation has become the de-facto strategy for mitigating data scarcity, yet the community lacks a systematic, controlled comparison of augmentation techniques across the diverse family of modern HTR architectures. In this work we present a comprehensive benchmark that evaluates ten widely used image-space augmentation techniques - rotation, scaling, translation, shearing, Gaussian noise, blur, elastic distortion, random erasing, perspective transformation and brightness/contrast adjustment - as well as several combined strategies, on five representative recognition models that span the evolution of the field: CRNN, Puigcerver, HTR-Flor, the Vertical Attention Network (VAN) and the Transformer-based TrOCR. All models are trained and tested under identical protocols on the IAM handwriting database, with the CVL database used for cross-dataset validation, so that the observed differences can be attributed to the augmentation rather than to inconsistent experimental settings. We report character error rate (CER) and word error rate (WER) for the baseline and every augmentation configuration, complement the accuracy analysis with a detailed computational study of training time, inference latency, memory footprint and parameter count, and confirm the reliability of the improvements with paired significance testing (paired t-test and Wilcoxon signed-rank test). The results show that elastic distortion is consistently the single most effective augmentation, reducing CER by up to 1.5 percentage points, that attention-based and transformer-based models benefit less from geometric augmentation than convolutional-recurrent models, and that aggressive blur and large random-erasing masks can degrade performance. Based on these findings we provide practical, architecture-aware recommendations for selecting augmentation strategies in offline HTR.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.