Domain Adaptation Method For Document Object Detection

Documents are essential information carriers, and obtaining digital information from document images cannot be overemphasized. Document images can be considered objects consisting of multiple semantic elements or regions. Correctly detecting different semantic objects in a document has been of considerable research significance and relevance. As a critical step and method for digitizing document image information, document object detection is also the basis of many downstream document tasks. In this paper, based on the classical target detection framework Faster RCNN model, we apply "strong alignment" at the image level and "weak alignment" at the instance level to mitigate the domain differences between source and target domains by training domain classifiers in the scenario where there are domain differences between different language documents and documents with varying styles of layout. A new loss function is designed to obtain a reliable relative comparison between the most possible positive case and the most difficult negative case, and to distinguish the more difficult categories. This scheme effectively migrates the detection knowledge learned by the model in the source domain data to the target domain data, which significantly improves the generalization performance and robustness of the document-oriented object detection model.