A brief review of state-of-the-art object detectors on benchmark document images datasets

Document image analysis (DIA) has become a challenging brand in computer vision, which is the foundation of document understanding applications. Page object detection is one of the crucial tasks in DIA, locating instances of semantic objects and graphical objects on a page of documents. Based on the development of deep learning algorithms, the performance of detectors is greatly improved. Therefore, this paper reviews the development of page object detection and publicly available document image datasets in recent years. The research also empirically evaluates the state-of-the-art object detection methods on page object detection problems. In particular, we experiment with fourteen methods on two benchmark datasets: IIIT-AR-13k and UIT-DODV-Ext. Finally, we discuss the architecture and performance of these methods to demonstrate their effectiveness and point out a set of development trends. The analysis provides a comprehensive evaluation, which follows state-of-the-art algorithms and future research.