arXiv 2601.219572026-04-03cs.CV

PaddleOCR-VL-1.5：迈向一个鲁棒的多任务0.9B视觉语言模型用于野外文档解析

PaddleOCR-VL-1.5: Towards a Multi-Task 0.9B VLM for Robust In-the-Wild Document Parsing

PaddlePaddle Team, Baidu Inc.（百度公司PaddlePaddle团队）

Cheng Cui, Ting Sun, Suyin Liang, Tingquan Gao, Zelun Zhang, Jiaxuan Liu, Xueqing Wang, Changda Zhou, Hongen Liu, Manhui Lin, Yue Zhang, Yubo Zhang, Yi Liu, Dianhai Yu, Yanjun Ma

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

PaddleOCR-VL-1.5通过引入Real5-OmniDocBench基准测试，在文档解析任务中达到94.5%的SOTA准确率，并扩展了密封识别和文本定位任务，同时保持0.9B超紧凑模型的高效性。

摘要：

我们介绍了PaddleOCR-VL-1.5，一个升级模型，在OmniDocBench v1.5上实现了新的SOTA准确率94.5%。为严格评估真实世界物理失真（包括扫描、倾斜、变形、屏幕摄影和照明）的鲁棒性，我们提出了Real5-OmniDocBench基准。实验结果表明，该增强模型在新编纂的基准上实现了SOTA性能。此外，我们通过整合密封识别和文本定位任务扩展了模型的能力，同时仍保持0.9B超紧凑VLM的高效性。代码：https://github.com/PaddlePaddle/PaddleOCR

英文摘要：

We introduce PaddleOCR-VL-1.5, an upgraded model achieving a new state-of-the-art (SOTA) accuracy of 94.5% on OmniDocBench v1.5. To rigorously evaluate robustness against real-world physical distortions, including scanning, skew, warping, screen-photography, and illumination, we propose the Real5-OmniDocBench benchmark. Experimental results demonstrate that this enhanced model attains SOTA performance on the newly curated benchmark. Furthermore, we extend the model's capabilities by incorporating seal recognition and text spotting tasks, while remaining a 0.9B ultra-compact VLM with high efficiency. Code: https://github.com/PaddlePaddle/PaddleOCR

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。