Recognition of Historical Gurmukhi Numerals Using Fine-Tuned Convolution Neural Network

Historical Gurmukhi manuscripts hold immense significance as invaluable treasures of heritage and knowledge. These ancient documents are replete with handwritten numerals. These numerals serve as a critical marker for understanding chronology and historical contexts. While many attempts have been made in the past to recognize handwritten numerals in the Gurmukhi script, no one has endeavored to recognize numerals in historical Gurmukhi manuscripts. Recognition of historical numerals presents unique challenges, including diverse line thickness, degraded paper quality, fragmented characters, noise, stains, and the use of multicolored ink, which are not typically found in modern numerals. This research paper addresses the challenge of numeral recognition in historical Gurmukhi manuscripts. The primary contributions of the paper are (1) first attempt for creation of the dataset of numerals of historical Gurmukhi manuscript (2) development of the deep learning model for recognition of numerals of historical Gurmukhi manuscript. The collected dataset includes 7450 samples of 10 numerals extracted from 1744 pages of 36 different historical Gurmukhi manuscripts. It presents a rich diversity of challenges. To tackle the challenges, authors adopt a transfer learning approach and fine tune a pre-trained VGG16 model. Additionally, experiments were conducted with the different optimizers and batch sizes. The experiments were carried out on Google Colab pro using Keras and TensorFlow framework. These experiments culminated in the remarkable achievement of a 0.9945 recall score, underscoring the effectiveness of the proposed approach in recognition of the numbers of historical Gurmukhi manuscripts.