LOCALISATION OF TEXT IN A NATURAL SCENE IMAGE BASED ON DEVANAGARI SCRIPT RULE USING DEEP LEARNING

Text localization play a crucial role in scene images to read the text content effectively. There have been built many deep learning (DL) rooted models for classification as well as localization of manifold language text in scene images in recent years. Though, after a thorough review it is explored that there is very less research conducted on text localization based on the Devanagari script rule. However, previously developed models for text classification and localization have a few setbacks namely lower performance metrics, hardship in text localization with multi-scaling shapes, intricacy in handling irregular text etc. This research presents a novel framework which is developed for the localization of text in a natural scene image based on the Devanagari script rule using deep learning. This framework initially pre-processes the input datasets for standardization. Secondly, it integrates the enhanced YOLO-v7 for candidate component detection, improvised support vector machine (SVM) and AdaBoost as a hybrid text classifier model. Lastly, text segmentation and clustering are done using balanced iterative reducing and clustering using the hierarchies (BIRCH) algorithm and transfer learning is used in the localization model. This proposed model obtains accuracy, precision, recall and F1 score on vpdDataset as well as ICDAR 2015 datasets 98.33%, 99.01%, 99.05%, 97.76%, and 98.09%, 98.02%, 99.03%, 97.46%, respectively. Therefore, the obtained findings of this proposed model are very optimal and enhanced in comparison with previous text detection and localization models.