End-to-End Information Extraction by Character-Level Embedding and Multi-Stage Attentional U-Net

Information extraction from document images has received a lot of attention\nrecently, due to the need for digitizing a large volume of unstructured\ndocuments such as invoices, receipts, bank transfers, etc. In this paper, we\npropose a novel deep learning architecture for end-to-end information\nextraction on the 2D character-grid embedding of the document, namely the\n\\textit{Multi-Stage Attentional U-Net}. To effectively capture the textual and\nspatial relations between 2D elements, our model leverages a specialized\nmulti-stage encoder-decoders design, in conjunction with efficient uses of the\nself-attention mechanism and the box convolution. Experimental results on\ndifferent datasets show that our model outperforms the baseline U-Net\narchitecture by a large margin while using 40\\% fewer parameters. Moreover, it\nalso significantly improved the baseline in erroneous OCR and limited training\ndata scenario, thus becomes practical for real-world applications.\n