Lightweight Script Classification for Multilingual Scene Text Recognition Using MobileNetV2

In multilingual scene text recognition, accurate identification of the script used in each text region is essential before applying language-specific OCR.This paper proposes a lightweight script classification module based on MobileNetV2 [1], integrated into a broader Telugu scene text recognition pipeline.The system first detects word-level text regions using an enhanced EAST detector and then classifies each region into one of three script classes Telugu, English, or Hindi.The proposed classifier leverages transfer learning, efficient preprocessing, and a balanced dataset augmented to address class imbalance.Experimental results show that the classifier achieves a high overall accuracy of 94.81%, with minimal inter-script confusion, even in visually cluttered scenes.Qualitative examples and a detailed confusion matrix validate the model's robustness and generalizability.This approach demonstrates how lightweight deep learning models can be effectively used in real-world OCR systems, particularly for Indian languages.Future directions include expanding script coverage, enabling handwritten text recognition, and integrating the module into an end-to-end OCR pipeline.