ABSTRACT Handwritten digit classification is a well-known and important problem in the field of optical character recognition (OCR). The primary challenge is correctly classifying digits which are highly varied in their visual characteristics primarily due to the writing styles of different individuals. In this paper, we propose the use of Convolutional Neural Networks (CNN) for the purpose of classifying handwritten Bangla and Hindi numerals. The major advantage that we face by using a CNN-based classifier is that no prior hand-crafted feature needs to be extracted from the images for efficient and accurate classification. An added benefit of a CNN classifier is that it provides translational invariance and a certain extent of rotational invariance during recognition. Applications can be found in real-time OCR systems where input images are often not perfectly oriented along a vertical axis. In this work, we use modified versions of the well-known LeNet CNN architecture. Extensive experiments have revealed a best-case classification accuracy of 98.2% for Bangla and 98.8% for Hindi numerals outperforming competitive models in the literature.

中文翻译：

---

使用深度学习在孟加拉语和印地语中进行手写数字分类

摘要 手写数字分类是光学字符识别（OCR）领域众所周知的重要问题。主要的挑战是正确分类数字，这些数字的视觉特征差异很大，这主要是由于不同个人的书写风格造成的。在本文中，我们建议使用卷积神经网络 (CNN) 来对手写的孟加拉语和印地语数字进行分类。我们使用基于 CNN 的分类器面临的主要优势是不需要从图像中提取事先手工制作的特征来进行有效和准确的分类。CNN 分类器的另一个好处是它在识别过程中提供平移不变性和一定程度的旋转不变性。可以在实时 OCR 系统中找到应用程序，其中输入图像通常不能完美地沿垂直轴定向。在这项工作中，我们使用了著名的 LeNet CNN 架构的修改版本。大量实验表明，孟加拉语和印地语数字的最佳分类准确率为 98.2%，印地语数字为 98.8%，优于文献中的竞争模型。