AdamR-GRUs: Adaptive momentum-based Regularized GRU for HMER problems

Abstract

Handwritten Mathematical Expression Recognition (HMER) is essential to online education and scientific research. However, discerning the length and characters of handwritten mathematical expressions remains a formidable challenge. In this study, we propose an encoder–decoder architecture incorporating a novel Adaptive momentum-based Regularized Gated Recurrent Unit (AdamR-GRU) to solve this problem. It incorporates a novel gate within the GRU framework to maintain input histories, enabling extended memory retention and faster convergence. To assess the performance of AdamR-GRU, we conducted a series of comprehensive experiments on the CROHME 2014, 2016, and 2019 datasets. The results indicate that the AdamR-GRU model exhibits competitive performance with existing state-of-the-art methods regarding Expression Recognition Rate (exprate). Moreover, our model achieves state-of-the-art accuracy concerning 1, 2, and 3 symbol-level errors (top 1, 2, 3 error accuracy) on the CROHME 2014 and 2016 datasets. For the CROHME 2019 dataset, the model attains state-of-the-art performance in top 2 and 3 error accuracy. AdamR-GRU demonstrates reduced computational complexity and memory usage compared to alternative contemporary models.