说明:本报告以 RapidOCR 对 PDF 每页截图的识别结果为页码锚点,将 MinerU 的图文块按语义重新对齐到真实 PDF 页,解决此前按块序号简单分配导致的错位问题。四列分别为 PDF 原页、MinerU(OCR+vlm)、Unstructured(hi_res)、RapidOCR(本地方便,基于 ONNX)。
图内OCR原文] i S & =3 Oxford S5 £y < Reading B S Tree X of easure — i 19 . ) < = > / = = /7 = N , \) Fo o , -9 R R < = 3 & B o> \ 3, < 7 R e < ¢ A < U —— Q) N L A S> SN 3. = 2. = y » & X a ies created by Roderick Hunt and Alex Brycht V
RapidOCR 识别结果
ox of
Oxford
Reading
Tree
reasure
Treasuse
Ten
eries created by Roderick Hunt and Alex Brychta
图内OCR原文] S . 8! \ = — —— t T Y y —= | S N £ — _ ——— — — ~ T e It S — y /i ~ I N2 % by A M £ 4 ¢ \ ) = 3 2 13 & v — =~ — C S g @D | = . | o W s o
“Dig here,” said Biff.
Chip dug in the sand.
图内OCR原文] /) _ — = — = ¢ J [ \ ! S S A X / 17 5 ~ £ =\
图内OCR原文] & AN 1 ] \ =\ S5 M <5 = =2 )/ 1l \\/ \ N , 4 - — 4 o\ N __ S NN | &
We drill.
We Build Level aa
We paint
图内OCR原文] I ' " NN Al ) — / = — N S oc, NN
We glue.
We Build Level aa
图内OCR原文] N ™ < @2 4 = ) 7] \\ il | it Ve \ \ A\ _ ] 2 % \ A ENN t | \ —
We hang
10