Efficiency evaluation of filter sizes on graph convolutional neural networks for information extraction from receipts

Abstract

Graph Neural Networks (GNNs) have attracted considerable attention due to their ability to analyze structured data represented as graphs. In invoice information extraction, GNNs have proven to be a powerful tool for automatically extracting relevant information from invoices, streamlining data entry processes, and improving efficiency. By modeling the invoice layout as a graph and exploiting the inherent structural dependencies, GNNs enable end-to-end extraction by encoding the graph structure and using deep learning techniques. This work proposes a Graph Convolution Network to extract information from invoices. Furthermore, an evaluation of the effect of filter sizes on the model’s accuracy was performed. We built an extraction model based on the filter size selected by the evaluation. We achieved the accuracy of the test set of 96.4% and the training set of 98.5% on the dataset of about 1.500 invoice images we collected.