Organizations: Unisound AI Technology Co., Ltd. China · School of Computer Science and Technology, Tianjin University China · MAIS, Institute of Automation, Chinese Academy of Sciences China
Accurate document parsing requires both robust content recognition and a stable parser interface. In explicit Document Layout Analysis (DLA) pipelines, downstream parsers do not consume the full detector output. Instead, they operate on a retained and serialized set of layout instances. However, on dense pages with overlapping regions and ambiguous boundaries, unstable layout hypotheses can make the retained instance set inconsistent with its parser input order, leading to severe downstream parsing errors. To address this issue, we introduce a lightweight structural refinement stage between a DETR-style detector and the parser to stabilize the parser interface. Treating raw detector outputs as a compact hypothesis pool, the proposed module performs set-level reasoning over query features, semantic cues, box geometry, and visual evidence. From a shared refined structural state, it jointly determines instance retention, refines box localization, and predicts parser input order before handoff. We further introduce retention-oriented supervision and a difficulty-aware ordering objective to better align the retained instance set and its order with the final parser input, especially on structurally complex pages. Extensive experiments on public benchmarks show that our method consistently improves page-level layout quality. When integrated into a standard end-to-end parsing pipeline, the stabilized parser interface also substantially reduces sequence mismatch, achieving a Reading Order Edit of 0.024 on OmniDocBench.
Figures & tables
Figure 1. Illustration of detector-to-parser handoff failures on a page. A dense page may produce multiple overlapping layout candidates around the same content region. Without suppression, duplicated candidates may be passed to the parser, causing content redundancy and an incorrect reading sequence. With heuristic NMS, an incomplete survivor may be retained while a better-localized alternative is removed, leading to content loss and an unstable parser input. A four-panel diagram comparing ground truth, overlapping layout candidates, and parser outputs with and without non-maximum suppression. Overlapping candidates cause duplicated content and reading-order errors without suppression, while heuristic suppression may keep an incomplete region and remove a better-localized one, leading to content loss.
Figure 2. Overview of the proposed parser-oriented structural refinement framework. Given a document image, a DETR-style detector first produces a compact pool of layout hypotheses. A lightweight refinement module then performs set-level reasoning over detector query features, semantic cues, box geometry, and image evidence, so that localization, instance retention, and parser input order are determined coherently from the same refined state. The downstream OCR-VL parser consumes a clean and well-ordered layout set, yielding more faithful Markdown/JSON output, particularly on structurally challenging pages. An overview diagram of the proposed parser-oriented structural refinement framework. A detector first produces layout hypotheses from a document image. Query features, box geometry, semantic cues, and visual evidence are fused into refinement tokens, which are updated by a multi-layer refinement decoder. The refined states are then used for instance retention and ordering to form a parser-ready layout, improving page-level layout quality, reading order, and final Markdown output.
Method
OmniDocBench
D4LA
DocLayNet
P
R
F1
P
R
F1
P
R
F1
DocLayout-YOLO
92.30
97.70
94.10
82.60
95.40
87.30
88.00
96.30
90.90
PP-StructureV3
94.80
96.20
94.60
85.70
91.00
86.00
92.40
95.70
93.00
MinerU2.5
95.80
97.00
95.90
90.40
92.50
90.20
92.80
97.70
95.25
PP-DocLayoutV3
95.81
96.26
96.03
89.04
90.40
89.71
91.62
95.46
93.50
dots.ocr v1.5
94.97
96.21
95.59
91.03
94.57
92.80
92.56
94.34
93.45
Table 1. PageIoU-based precision, recall, and F1 on OmniDocBench, D4LA, and DocLayNet. Best results are in bold and second-best results are underlined .
Method
Params
Overall
Scanning
Warping
Screen-Photography
Illumination
Skew
Score ↑
Reading Order ↓
Score ↑
Reading Order ↓
Score ↑
Reading Order ↓
Score ↑
Reading Order ↓
Score ↑
Reading Order ↓
Score ↑
Reading Order ↓
Pipeline Tools
Marker-1.8.2
-
60.10
0.367
70.27
0.238
58.98
0.390
63.65
0.325
66.31
0.337
41.27
0.543
PP-StructureV3
-
64.45
0.212
84.68
0.092
59.34
0.261
66.89
0.165
73.38
0.126
37.98
0.417
Specialized VLMs
DeepSeek-OCR 2
3B
73.01
0.153
89.59
0.056
66.53
0.209
71.65
0.157
76.02
0.122
61.28
0.221
Table 2. Results on Real5-OmniDocBench. Best results are in bold and second-best results are underlined .
Method
Overall ↑
Text Edit ↓
Formula CDM ↑
Table TEDS ↑
Table TEDS-S ↑
Reading Order Edit ↓
Pipeline Tools
Marker-1.8.2
71.30
0.206
76.66
57.88
71.17
0.250
PP-StructureV3
86.73
0.073
85.79
81.68
89.48
0.073
Specialized VLMs
Dolphin-v2
89.78
0.054
87.63
87.02
90.48
0.054
OCRFlux-3B
74.82
0.193
68.03
75.75
80.23
0.202
Table 3. Parsing results on OmniDocBench. Best results are in bold and second-best results are underlined .
Method
Latency (ms/page) ↓
Params ↓
PP-DocLayoutV3
154.45
32.48M
PP-DocLayoutV3 + NMS
154.52
32.48M
D-FINE
119.55
30.68M
D-FINE + NMS
119.70
30.68M
D-FINE + LayoutReader + NMS
199.63
400M
Ours ( N=300 )
148.34
61.57M
Table 4. Efficiency comparison of layout analysis under our implementation setting. Latency is measured in milliseconds per page. N denotes the size of the input hypothesis pool.
Method
D4LA
DocLayNet
OmniDocBench
P
R
F1
P
R
F1
P
R
F1
PP-DocLayoutV3
82.91
93.96
88.09
87.01
95.50
91.06
90.25
96.68
93.35
PP-DocLayoutV3 + NMS
89.04
90.40
89.71
91.62
95.46
93.50
95.81
96.26
96.03
D-FINE
87.41
97.62
92.23
80.91
97.92
88.61
93.70
97.04
95.34
D-FINE + NMS
91.76
95.92
93.79
86.97
97.68
92.01
95.52
95.97
95.74
Ours
92.04
95.90
93.93
91.49
97.76
94.52
95.97
96.50
96.23
Table 5. Ablation on NMS-free structural refinement. Best results are in bold and second-best results are underlined .
Method
Textual
Image
Table
Equation
Page_margins
D-FINE
0.967
0.947
0.968
0.885
0.857
D-FINE + NMS
0.971
0.956
0.972
0.901
0.873
PP-DocLayoutV3 + NMS
0.966
0.961
0.966
0.904
0.902
Ours
0.972
0.956
0.975
0.932
0.933
Table 6. Class-wise F1 on OmniDocBench. Best results are in bold and second-best results are underlined .
Variant
OmniDocBench F1 ↑
Overall ↑
Reading Order Edit ↓
D-FINE
95.34
89.97
N/A
D-FINE + NMS
95.74
93.24
N/A
D-FINE + Soft-NMS
95.79
93.29
N/A
D-FINE + LayoutReader + NMS
95.74
93.24
0.175
D-FINE + LayoutReader + Soft-NMS
95.79
93.29
0.171
Ours w/o Lret
95.41
90.84
0.084
Table 7. Compact core comparison and objective ablation on OmniDocBench. N/A indicates no reading-order prediction.
Figure 3. Qualitative comparison on two challenging layouts (Case A: equations, Case B: chemical schemes). Our method avoids the severe redundancy of raw outputs and the content loss caused by traditional NMS, yielding clean and complete structures. A two-row qualitative comparison of layout predictions on an equation page and a chemical-scheme page. Without NMS, multiple overlapping boxes are produced. With heuristic NMS, some valid content is removed and the retained layout becomes incomplete. The proposed method produces clean, complete, and well-aligned layout regions in both cases.
Pipeline
Variant
Reading Order Edit ↓
Overall ↑
MinerU2.5
Original
0.044
90.67
+ Ours
0.032 ( ↓ 27.3%)
94.03 ( ↑ 3.7%)
DeepSeek-OCR 2
Original
0.057
91.09
+ Ours
0.034 ( ↓ 40.4%)
92.48 ( ↑ 1.5%)
FireRed-OCR-2B
Original
0.041
92.94
+ Ours
0.029 ( ↓ 29.3%)
93.98 ( ↑ 1.1%)
Table 8. Exploratory compatibility study on OmniDocBench. The first row reports the original result, and the second row reports the result after incorporating our module.