Persian Handwritten Digits and Symbols Dataset (PHDD): An Open Benchmark for Persian OCR and Pattern Recognition
Abstract
This paper introduces the Persian Handwritten Digits and Symbols Dataset (PHDD), an open benchmark designed to facilitate research in Persian optical character recognition (OCR) and pattern recognition. The dataset comprises 40,214 grayscale images of handwritten Persian numerals and mathematical symbols, collected from about 200 volunteer students at the University of Zanjan. The dataset was processed through a fully automated pipeline developed in Python using OpenCV. The pipeline performs character detection, automatic cropping, denoising, normalization, and resizing to 28 × 28 pixels, yielding uniform, high-quality images stored in lossless PNG format. All preprocessing scripts and the full data-generation workflow are publicly available. Beyond dataset compilation, a convolutional neural network (CNN) was trained for feature extraction and multi-class classification, producing robust descriptor embeddings. A two-dimensional visualization of these learned descriptors using the t-SNE technique reveals compact, well-separated clusters, demonstrating the discriminative power of the extracted features. PHDD provides a valuable foundation for applications in deep learning, computer vision, image classification, and OCR, enabling reproducible studies in handwritten digit and symbol recognition.