Document Image Layout Analysis via MASK Constraint

Document layout analysis plays an essential role in computer vision. With the development of deep learning, more and more deep learning methods are proposed to solve some challenges in document layout analysis. Semantic segmentation-based and object detection-based methods are two mainstream approaches for document layout analysis. Compared with methods based on semantic segmentation, methods based on target detection have certain advantages in ensuring the integrity of target objects, especially with the proposal of Mask R-CNN. However, since the document layout analysis task is different from the general target detection task, there is a particular semantic gap in the document layout analysis (i.e the image to be detected may contain text), and the Mask R-CNN cannot solve this problem well. Therefore, we design a hierarchical information augmentation module, which can fully utilize low-dimensional detail information and maintain high-dimensional semantic information. In addition, we propose a novel MASK-constrained module, which ensures that the global semantic information of the input module can be further mined by embedding MASK information in the input image. Furthermore, to combat the issue of overlapping bounding boxes arising from Mask R-CNN processing, we propose a Constrained Aggregation method. Finally, we validate our approach using benchmark datasets featuring complex layouts (such as DSSE-200 and FPD). The results underscore the significant performance gains achievable with our proposed method.