The cursive nature, Nastaliq writing style and a large number of different ligatures make ligature recognition very difficult in Urdu. In this paper, we present a segmentation-free approach to holistically recognize Urdu ligatures. We first generate a rich dataset which contains 17,010 ligatures with different orientation and different degrees of noise. Secondly, the ligatures are clustered (categorized) in order to reduce the search space and make the learning robust. Finally, we employ a deep neural network with dropout regularization to classify ligatures. The detailed experiments show that a deep neural network with dropout regularization and clustering of ligatures significantly enhances the classification accuracy.

中文翻译：

---

使用深度神经网络的基于结扎分类的Nastaliq Urdu识别

草书的性质，Nastaliq的写作风格以及大量不同的连字使乌尔都语很难识别连字。在本文中，我们提出了一种无分割的方法来全面识别乌尔都语连字。我们首先生成一个丰富的数据集，其中包含17010个具有不同方向和不同噪声程度的连字。其次，对连字进行聚类（分类）以减少搜索空间并提高学习的鲁棒性。最后，我们采用带有辍学正则化的深度神经网络对连字进行分类。详细的实验表明，具有辍学正则化和连字聚类的深度神经网络显着提高了分类准确性。