Scene table structure recognition with segmentation collaboration and alignment *

This study addresses the challenge of scene table structure recognition (S-TSR). In contrast to the well-aligned tables in PDF documents or screenshots, natural scene tables tend to be inclined, rotated, and curved, thereby damaging the structure of the table. Consequently, many state-of-the-art methods cannot operate effectively owing to a lack of well-aligned priors. We propose a novel segmentation collaboration and alignment network (SCAN) to address the problem of table structure recognition in natural scenar-ios. Our SCAN combines the location and logical information of table cells through the segmentation collaboration module to segment the cell region accurately. The proposed cell alignment module aligns the segmentation result to restore the distorted table structure. The experimental results demonstrate that our method is robust to different challenging S-TSR sub-scenarios and achieves new state-of-the-art performance, with TEDS scores of 90.7 and 98.5 on the S-TSR benchmarks WTW and TAL_OCR_TABLE, respectively. (c) 2022 Elsevier B.V. All rights reserved.