SLIDE 1
LOC-DB Reference Extraction
DR. DR.-ING SHERAZ AHMED SYED TA TAHSEEN RAZA RIZVI
LOC-DB Architecture
2
LOC-DB Reference Extraction DR. DR.-ING SHERAZ AHMED SYED TA - - PowerPoint PPT Presentation
LOC-DB Reference Extraction DR. DR.-ING SHERAZ AHMED SYED TA - - PowerPoint PPT Presentation
LOC-DB Reference Extraction DR. DR.-ING SHERAZ AHMED SYED TA TAHSEEN RAZA RIZVI LOC-DB Architecture 2 LOC-DB OCR Component Types of Input files: Digital Born PDF Scanned Documents XML/HTML XML File Scanned Document Textual PDF
SLIDE 2
SLIDE 3
LOC-DB OCR Component
Types of Input files:
- Digital Born PDF
- Scanned Documents
- XML/HTML
3
XML File Textual PDF Scanned Document
SLIDE 4
Reference Extraction from: Scanned Documents
4
SLIDE 5
Scanned Documents: Reference Extraction
5
- Step 1: Binarization
- Greyscale(0-255)/color to
Binary (0-1)
RGB Image Binary Image
SLIDE 6
Scanned Documents : Reference Extraction
6
- Step 2: Image Classification
- Single/Double Column Documents
Single Column Document Double Column Document
Single Column Documents Double Column Documents
SLIDE 7
Scanned Documents : Reference Extraction
7
- Step 3: OCR (Optical Character Recognition)
OCR Result
SLIDE 8
Scanned Documents : Reference Extraction
8
- Step 4: Reference Segmentation
- Using ParsCit
SLIDE 9
Reference Extraction from: Textual / Digital Born PDFs
9
SLIDE 10
Digital Born PDFs : Reference Extraction
10
- Step 1: Text Extraction
Textual PDF Extracted Text
SLIDE 11
Digital Born PDFs : Reference Extraction
11
- Step 2: Reference Extraction
- Using ParsCit
SLIDE 12
Reference Extraction from: Structured XML
12
SLIDE 13
Structured XML : Reference Extraction
13
- Step 1: Preprocessing
SLIDE 14
Structured XML : Reference Extraction
14
- Step 2: Reference Extraction
- Using ParsCit
SLIDE 15
Scanned Documents Textual PDFs Structured XML
Reference Extraction Pipeline - Overview
15
Binarization Image Classification OCR Text Extraction Reference Segmentation Pre-Processing
SLIDE 16
DeepBibX: A Neural Network based approach
16
SLIDE 17
DeepBibX: Intuition
17
SLIDE 18
Neural Network Based Approach
18
SLIDE 19
Comparison with ParsCit
19
ParsCit Output DeepBibX Output
SLIDE 20
Comparison with ParsCit
- On a test set of 286 bibliographic document
images:
- Total: 5090 references
- ParsCit extracted: 3645 references
- Proposed approach: 4323 references
20
1000 2000 3000 4000 5000 6000 ParsCit FCN based approach
Extraction Comparison
Total References Total Detections
SLIDE 21
Thank you
21