MixLoss: Table Structure Recognition in Industrial Documents via Collaborative Learning

Table Structure Recognition (TSR) is a fundamental challenge in document analysis, especially for industrial documents where tables often contain complex layouts and noisy formatting. Although recent image-to-markup approaches have made progress using end-to-end learning, they commonly suffer from misalignment between predicted cell bounding boxes and actual text regions, leading to structural parsing errors. To address this issue, we propose MixLoss, a collaborative learning framework designed to improve TSR performance by enforcing position-wise consistency between HTML structure prediction and bounding box detection. MixLoss combines HTML tokens with bounding box coordinates in a unified sequence, inserting coordinate tokens directly after filled cell tokens. This design ensures structural alignment while maintaining computational efficiency. Extensive experiments show that MixLoss delivers significant improvements on real-world industrial datasets, including a 2.2% gain on IX DocBench, while maintaining strong performance on standard benchmarks. These results demonstrate the effectiveness of collaborative learning in enhancing table structure recognition for practical industrial document parsing.