Document dewarping has been researched for many years but remains unresolved, particularly due to multi-scale document issues (where the background occupies a large proportion). To enhance the multi-scale awareness of document dewarping algorithms, we propose a two-stage framework with explicit scale-aware capabilities, named Scale- Doc, which consists of a scale-aware stage and a dewarping stage. The scale-aware stage, a convolutional network based on self-attention mechanisms, is proposed to explicitly remove the document background. The dewarping stage introduces a lightweight method that dewarps warped documents by predicting document edges using sparse control points. Furthermore, to train the scale-aware stage network and validate the effectiveness of ScaleDoc, a new document dataset, DocW, has been cre- ated. DocW comprises 1k images of varying scales and warping levels, all of which are authentically captured rather than generated. Comparative and ablation studies are conducted on the newly created DocW dataset and DocUnet benchmark dataset. Dewarping results, measured by the MS-SSIM and LD metrics, and OCR results, measured by the CER and ED indicate that the proposed model outperforms the current state-of- the-art (SOTA) models in dealing with multi-scale document challenges.
More information about dataset DocW could be found in dataset.md
