Two new papers have been accepted at The 17th International Conference on Document Analysis and Recognition. The first paper, authored by Mohamed DHOUIB, Ghassen BETTAIEB, and Aymen SHABOU, presents an OCR-free approach for extracting information from visually rich scanned documents. The second paper, authored by Timothé FRONTEAU, Arnaud PARAN, and Aymen SHABOU, investigates adversarial attacks on document classification tasks, which is a relatively unexplored area of research. These papers will be presented at the conference on August 21st by Mohamed DHOUIB and Arnaud PARAN.

In the first paper, an OCR-free end-to-end information extraction model named DocParser is proposed. The proposed approach achieves state-of-the-art results on various datasets in terms of speed and accuracy, which makes it perfectly suitable for real-world applications. This is mainly due to the ability of the model to effectively extract rich features thanks to its innovative architecture.

In the second paper, the authors focus on adversarial attacks and defenses, a field that has gained increasing interest in computer vision systems. However, most of the previous research has been limited to natural images, which are very different from scanned documents. The authors applied several adversarial attacks to scanned documents and experimented with different strategies to protect models against such attacks.  This paper is the first study of its kind conducted by the community to evaluate the impact of these attacks on document image classification tasks and we hope that it will raise awareness of this important issue and inspire further research in this area.

Paper 1: DocParser: End-to-end OCR-free Information Extraction from Visually Rich Documents

Paper 2: Evaluating Adversarial Robustness on Document Image Classification