Integrated framework utilizing scene text detection and recognition techniques for enhancing point of interest...

This paper focused on enhancing text recognition, script identification, language classification, and point of interest (POI) extraction from images captured by Mobile Mapping Systems (MMS). The initiative was undertaken to improve the existing Computer vision-based artificial intelligence modules. The advancements made is to be contributed to the system's implementation, bringing improved functionality and accuracy to the system. The current system, called Text Detection and Recognition (TDR), consists of several neural modules operating in sequential stages. The first module identifies areas of interest within the MMS images, focusing on shop signboards, traffic signs, and directional boards. The second stage involves detecting text words within these areas and cropping the relevant pixels from the image. These cropped images are then processed in the third stage, where the language script is detected, identifying one of ten Indian scripts. In the fourth stage, specific character recognizers corresponding to the identified script are used to recognize the text. The outputs from these stages are aggregated into a correlated JSON output. Additionally, a parallel fifth stage detects various fields within the MMS images, such as name, address, pin, icon, phone and GSTIN number, ultimately extracting a comprehensive human-readable address for any POI from the MMS image. The primary focus includes investigating novel text recognition algorithms to improve accuracy and efficiency, exploring various script identification algorithms to enhance language classification capabilities, implementing a dictionary-based approach for more accurate word detection, and developing methods for correcting the words that the CRNN model predicts to reduce errors. This work is novel because it combines word correction, OCR, detection, classification, and POI field extraction into a single pipeline designed specifically for Indic scripts. By obtaining 96.17% script recognition accuracy, 92.5% word accuracy, and 33% average precision in POI detection, the suggested framework outperforms previous benchmarks like IndicText (93.6%) and transformer-based OCR (88.5%).