Automatic Parsing and Extraction of Structured Data from Fax Images

This paper presents a sophisticated system designed to automatically convert fax images into structured, formatted data. The system employs a multi-stage approach to enhance accuracy and efficiency. Initially, faxes are classified into predefined categories to match their functional layouts, which allows the application of specialized data extraction models for different fax types. This pre-processing step enables the extraction models to focus on specific document styles, improving overall accuracy. The system utilizes GoogleNet for classification, achieving a high accuracy rate of 100%. For the data extraction process, the paper details two primary tasks: detecting text snippets and table cells, and recognizing these texts using Optical Character Recognition (OCR). The Light-Head R-CNN object detection model, combined with ResNet-101, is employed to effectively identify table cells and text regions. For text recognition, a Convolutional Recurrent Neural Network (CRNN) is used, incorporating VGG- 16 as the CNN backbone and a bi-directional LSTM with a CTC output layer for sequential character generation. The paper also introduces an incremental learning strategy utilizing author feedback to continuously refine the models. This approach, based on knowledge distillation, ensures the system adapts to new data without forgetting previously learned information. Future work will focus on improving the conversion of tabular data and addressing the variability in document layouts across different companies.