An Efficient Transfer Learning Approach for Handwritten Historical Gurmukhi Character Recognition using VGG16: Gurmukhi_Hhdb1.0 Dataset

Abstract

Historical manuscripts of Gurmukhi script hold immense historical and cultural significance. These manuscripts offer valuable insights into north Indian culture, politics, civilization and Sikhism. Preserving and analyzing these manuscripts is essential for gaining a deeper understanding of the past. Despite this, minimal attention was given to the recognition of characters in historical Gurmukhi manuscripts compared to modern text due to its unique challenges and scarcity of large and specialized datasets. To bridge this gap, we present benchmark dataset "Gurmukhi_HHdb1.0". This dataset comprises 87,181 images of 33 characters extracted from 6,340 pages of 43 historical Gurmukhi manuscripts. To recognize the characters of the Gurmukhi_HHdb1.0 dataset, authors developed an efficient transfer learning-based architecture by fine-tuning the VGG16 architecture. The entire dataset was split into 3 parts: 70% for training, 15% for validation and 15% for testing, with each subset containing approximately that percentage of the images. The authors achieved test accuracy of 99.77% and validation accuracy of 99.90% using the proposed approach. This showcases the efficacy of the proposed approach in accurately recognizing historical Gurmukhi characters.