Handwritten Text Recognition - Recurrence Obsolete? - Comparative Study of Recurrent vs Recurrence-Free Convolutional Architectures

Sep 15, 2026

Uploaded by Ezra

Unconstrained handwritten text recognition remains a major challenge for deep neural networks. Recent recurrent approaches, especially LSTM-based models, achieve state-of-the-art results but involve many trainable parameters and limit parallel training, increasing training time and complicating exploration of architectures. Recurrence-free alternatives, including Fully Convolutional Networks with gated mechanisms, have been proposed with competitive performance. This work studies convolutional architectures and compares them to a CNN+BLSTM baseline on an offline task using the RIMES dataset and an augmented grid-background variant.

LanguageEnglish

CategoryPresentations

FormatPDF

Pages6

Reading Time2 min

AccessFree to download

Introduction

Problem setting and challenges

Prior work: HMM and hybrid neural models

Recurrence-free approaches and research question

Related Works

Models based on recurrent layers

Recurrence-free models

Have convolutions already made recurrence obsolete for unconstrained handwritten text recognition ? Denis Coquenet, Yann Soullard, Clément Chatelain, Thierry Paquet Normandie University -University of Rouen LITIS Laboratory -EA 4108 Rouen, France {denis.coquenet, yann.soullard, clement.chatelain, [thierry.paquet}@litislab.eu](thierry.paquet}@litislab.eu) arXiv :2012 .04954v 1 [ cs .CV] 9 Dec 2020 Abstract—Unconstrained handwritten text recognition remains an important challenge for deep neural networks. These last years, recurrent networks and more speciﬁcally Long Short-Term Memory networks have achieved state-of-the-art performance in this ﬁeld. Nevertheless, they are made of a large number of trainable parameters and training recurrent neural networks does not support parallelism. This has a direct inﬂuence on the training time of such architectures, with also a direct consequence on the time required to explore various architectures. Recently, recurrence-free architectures such as Fully Convolutional Networks with gated mechanisms have been proposed as one possible alternative achieving competitive results. In this paper, we explore convolutional architectures and compare them to a CNN+BLSTM baseline. We propose an experimental study regarding different architectures on an ofﬂine handwriting recognition task using the RIMES dataset, and a modiﬁed version of it that consists of augmenting the images with notebook backgrounds that are printed grids. Index Terms—Handwritten text recognition, Convolutional Networks, Recurrence-free models, BLSTM. I. INTRODUCTION Unconstrained ofﬂine handwriting recognition consists in transcribing the textual content of an image into character sequences. Several characteristics can make this task very complex: each person has his own way of writing with speciﬁc character shapes, slant, alignment and character spacing. Moreover, models are faced with potentially multiple languages and different alphabets. The background and the quality of the input images can also dramatically affect the performance of the model. For a long time, handwriting recognition has been handled by Hidden Markov Model (HMM) based systems [1], with the help of language models. The main drawback of these systems is their inability to deal with long term dependencies in sequences. Hybrid systems combining HMM with Recurrent Neural Network (RNN) [2] or Convolutional Neural Network (CNN) [3] have demonstrated their superiority over HMM. In recent years, neural networks have proven their effectiveness in many applications related to document analysis [4, 5, 6, 7] . However, even though they have been steadily enhancing, there is still room for improvement regarding unconstrained ofﬂine handwriting recognition. Recently, recurrence-free architectures based on Fully Convolutional Networks with strong improvements such as gated mechanisms, shared-weight layers, residual connections and extensive data augmentation have been proposed [8, 9] and seems to achieve state-of-the-art results for handwriting recognition tasks. This paper raises the question whether recurrence is still necessary for modeling dependencies within signals such as handwriting. In order to address this question, we propose an experimental study comparing traditional recurrent architecturesand convolutional models in terms of performance, number of trainable parameters and training time. The rest of this paper is organized as follows. Section II gives an overview of related works in the ﬁeld. Section III presents the different architectures used for the experiments. Section IV is devoted to the experimental study, including a description of the datasets, the data augmentation techniques and the results obtained. We conclude this work in Section V. II. RELATED WORKS In this section, we brieﬂy present state-of-the-art works based on recurrent models as well as recent recurrence-free networks. A. Models based on recurrent layers

FAQs

Why do recurrence-based networks slow down training for unconstrained handwriting recognition?

Recurrent models contain many trainable parameters and recurrent training does not support parallelism, which increases training time and makes architecture exploration harder.

What is the main alternative investigated in this paper?

The study focuses on recurrence-free convolutional architectures, particularly Fully Convolutional Networks with gated mechanisms, and compares them against a CNN+BLSTM baseline.

Which datasets and augmentation are used for the experiments?

Experiments use the RIMES dataset and a modified version where images are augmented with notebook background printed grids.

Sep 15, 2026

Uploaded by Ezra