Abstract

Handwritten word recognition, a classical pattern recognition problem, converts a word image into its machine editable form. Mainly two basic approaches are followed to solve this problem, one is segmentation-based and the other is holistic. A number of research attempts have shown that the holistic approach performs better than its counterpart when the lexicon is predefined, fixed and small in size. Relying on this, initial benchmark recognition accuracy on CMATERdb2.1.2, a publicly available database consists of handwritten city names in Bangla, was reported following a holistic word recognition protocol. In the present work, we have followed the same trend to recognize the word samples of the said database and set a new benchmark recognition accuracy. A sparse convolutional neural network (CNN)-based model which is a low-cost trainable model has been developed for this. We have relied on a recently proposed hypothesis, known as lottery ticket hypothesis for pruning the layers of CNN model methodically, and derived a low-resource model having much less number of training parameters. This model competently surpasses the previously reported recognition accuracy on the said database by a significant margin with an axed training cost.

中文翻译：

---

使用彩票假设基于修剪的CNN模型的手写单词识别：CMATERdb2.1.2的新基准

摘要

手写单词识别是一种经典的模式识别问题，它将单词图像转换为机器可编辑的形式。解决这一问题主要采用两种基本方法，一种是基于细分的方法，另一种是整体的方法。许多研究尝试表明，当词典是预定义的，固定的和较小的时，整体方法的性能要优于同类方法。据此，根据整体单词识别协议，报告了CMATERdb2.1.2上的初始基准识别精度，CMATERdb2.1.2是一个由Bangla手写的城市名称组成的公共数据库。在当前的工作中，我们遵循相同的趋势来识别所述数据库的单词样本并设置新的基准识别精度。为此，已经开发了一种基于稀疏卷积神经网络（CNN）的低成本可训练模型。我们依靠一种最近提出的假设（称为彩票假设）来有条不紊地修剪CNN模型的各层，并得出了训练参数数量少得多的低资源模型。该模型大大地超过了先前在所述数据库上报告的识别精度，同时削减了培训成本。