DEVELOPMENT OF AN OCR MODULE FOR READING KK AND A SYNTHETIC DATA GENERATION MODULE FOR KTP, SIM, AND KK

• Main Author: Riemenn, Louis
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/78168
• Institution: Institut Teknologi Bandung

Data digitalization has become a crucial topic, especially in the current modern era. Manual document processing requires more time and effort compared to automated processing utilizing digital data. In the context of document digitalization, such as ID cards (KTP), driver's licenses (SIM), and family cards (KK), optical character recognition is employed to transform text in images into a digital format. This digital format can then be further processed digitally. In this final project, modules for generating synthetic data for KTP, SIM, and KK; text detection for KK documents; and text recognition for KK documents were developed. Synthetic data generation was accomplished using synthetic composites techniques, involving original images that were modified by adding synthetic elements that were previously absent. The original images (images from each KTP, SIM, and KK document) were modified by removing specific information in the images. False information was then added to the modified images, followed by introducing noise and tilting. The text detection and text recognition modules were built in three stages. The first stage involved selecting the most suitable model through benchmarking. The second stage consisted of training the chosen model using a KK dataset. The final stage was model evaluation to ensure that the model exhibited good performance and improvement compared to the previous state. Based on the conducted benchmarking, the selected model for text detection was DB with the MobileNetV3 backbone. Meanwhile, the chosen model for text recognition was SVTR with the SVTR-Tiny backbone. The selected models also demonstrated improved performance after training with the KK dataset, with the text detection model achieving precision of 97.80%, recall of 97.29%, and an F1 score of 97.54%, while the text recognition model achieved an accuracy of 99.99%.