Domain adaptive learning for document layout analysis and object detection using classifier alignment mechanism

Abstract

Document structure analysis and object detection is the new research interest, which involves differentiation between distinctive semantic regions of textual and non-textual objects. Document layout analysis remains a major challenge as document elements pose diversified layout structures, shapes, and appearances. Lack of annotated training datasets and domain shifts between datasets further increases the intricacy. The paper proposes an adaptive detection model for cross-domain learning ‘XDOD’ for document structure recognition and object detection. The detection model overcomes the domain shift between documents using, (1) Document Object Attention (DA) module, which learns coarse-to-refined features, and (2) Classifier Alignment (CA) module, which reduces the object misclassification. The XDOD model has been evaluated on various publicly available document datasets belonging to different domains. Extensive experimental results portray that the proposed XDOD model performs significantly better than the existing benchmark model giving more than 97% of detection rate on all datasets.