TeluguScriptify: A Custom Deep Learning Model for Handwritten Telugu Text Recognition and Tool Development

This paper proposes a new method for transforming handwritten Telugu text into editable electronic documents. Handwritten documents in the Telugu language are of very high historical value. There is, however, very high difficulty in recognizing the various types of style of handwriting since there are no very successful Telugu recognitions to compare with efficiency in comparison with Chinese and Arabic. To solve this problem, the recognition for Telugu script was designed with a deep learning custom model, along with a novel cosine similarity-based pre-processing technique to further enhance the model’s performance. The model architecture contains 3 convolutional blocks, 3 max-pooling layers, 4 dropout layers, 2 skip layers, and 2 dense layers. The experimental results show that the custom deep learning model is the frontrunner, achieving 97.3% testing accuracy with a corresponding validation loss of 0.116, the model beats well-known models such as Dense-Net and Alex-Net. The models were trained and validated on 52 Telugu characters from ‘ (Figure presented.) ’ (aa) to ‘ (Figure presented.) ’ (bandira). The state-of-the-art object detection model, YOLO v8n (8.2 nano) model, is used to detect the text regions in handwritten documents. © The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2025.