Key Information Extraction and Recognition from Rich Text Images

Abstract

Key information extraction and recognition from rich text images are crucial for various applications. There are two main tasks involved in this process: Line Item Recognition (LIR) and Key Information Localization and Extraction (KILE). LIR aims at identifying and interpreting data line items in a document. The essential information in each line item is then classified or extracted, a task known as KILE. A widely used approach for this problem is sequence based, which relies on the generalization of a language model and requires a significant amount of training time. We present an effective and reliable solution to the problem by using RoBERTa, a transformer model trained on a large corpus, along with the LION optimizer to improve the training process. A comprehensive evaluation was conducted on two different benchmarks, emphasizing two different languages, English and Vietnamese. Experimental results on DocILE indicate that the proposed framework significantly improves the KILE task with a 7.24% increase in accuracy compared to the baseline and also enhances the correct recognition rate at the LIR stage. On MCOCR, the method achieved a Character Error Rate (CER) of 28.6%, which is competitive with the state-of-the-art on this dataset.