I J C T A, 9(2) 2016, pp.

547-555
International Science Press

Block Diagram Components Information

Extraction and Grading of Research Paper
based on Graphical Images
R. Siva1, G. S. Mahalakshmi2 and S. Sendhilkumar3

ABSTRACT
Block Diagrams claim large importance in the conceptual understanding of data. The Text Extraction from block
diagrams, the optimized Image processing techniques and the selective extraction of sentences which contribute to
the meaning and functionality of the component are key elements involved in the successful execution. The project
aims at presenting essential information of the block diagram components in a comprehensive manner, using the
Web-Crawling techniques and to grade a research paper based on the textual explanation of graphical images. The
images which are the object of interest for grading are graphical images like bar graphs and pie charts. The primary
aim is to detect images like bar graphs, pie charts. Next we extract text explaining the graphical image from the
research paper. After extracting text, we analyse it to see how good it is based on various metrics and hence grade
the paper.
Keyword: Research Paper Grading, Web Crawling, Definition Extraction, VLDB.

I.

INTRODUCTION

A Research paper is written after a strenuous research and it involves lot of work than just conveying the
idea to convince the audience who need the results. Thus, research paper grading is very important to judge
the standard of the research paper. A good research paper should follow the golden rules for research paper
and should convey all the required details.
The paper should contain a detailed summary, a crisp abstract, citations, work cited, a good writing
style and a logical, organised paperwork. Research paper grading Rubric, an existing and widely used
method grades the research papers based on: Organisation, Writing style, key points, Citations, Analysis,
Insight and Grammar or Language use. This project aims at grading a research paper based on the explanation
given for the graphical images present in it. The dataset is collected from VLDB. The Analysis of Block
Diagrams involves each constituent component to be visualized along with its functionality to understand
the broad picture. Research Papers range from having high resolution to moderate and sometimes, even
low resolution diagrams which affect the image processing stages and hence, result in inaccurate and
improper text extraction from the images present. The availability of commercial and open source software
which facilitate the image processing procedures provide a good foundation for implementing the necessary
subtasks.
Optical character recognition (OCR) is the mechanical or electronic conversion of images of typewritten
or printed text into machine-encoded text. It is a common method of digitizing printed texts so that it can be
electronically edited, searched, stored more compactly, displayed on-line, and used in machine processes
1

Department of Computer Science and Engineering, KCG College of Technology, Karappakam, Chennai 600097
Department of Computer Science & Engineering,College of Engineering Guindy, Anna University, Chennai 600025
3
Department of Information Science and Technology, College of Engineering Guindy, Anna University, Chennai 600025
Email: [email protected], [email protected], [email protected]
2

548

R. Siva, G. S. Mahalakshmi and S. Sendhilkumar

such as machine translation, text-to-speech, key data and text mining. Using the text processing algorithms
and the sentence tokenizing methods, the required text is extracted and used to focus on further analysis in
terms of functionality and definition.
The graphical images in the research paper are extracted and then subjected to histogram conversion.
Based on continuity of pixels we distinguish graphical images from non - graphical images. The figure
number of graphical images is quoted as keywords. The concerned research papers are converted to text
files and the required textual explanation of the graphical images is extracted based on keywords. The
extracted text is subjected to analysis and based on several metrics we grade the quality of textual explanation
for a graphical image.
II. RELATED WORK
(A) OCR Related Framework
Tesseract is an open-source OCR engine that was developed at HP between 1984 and 1994. Tesseract
therefore assumes that its input is a binary image with optional polygonal text regions defined. Tesseract
was probably the first OCR engine able to handle white-on-black text so trivially [4]. The key parts of the
Line-Finding Algorithm are blob filtering and line construction.
Baseline Fitting Algorithm works on the text lines that have been found, and hence, the baselines are
fitted more precisely using a quadratic spline. This was another first for an OCR system, and enabled
Tesseract to handle pages with curved baselines. Tesseract tests the text lines to determine whether they are
fixed pitch. Where it finds fixed pitch text, Tesseract chops the words into characters using the pitch, and
disables the chopper and associator on these words for the word recognition step.
(B) Document Summarization
There is scope to improve the NLP preprocessing. Integrating a named entity (NE) parser in pre-processing
stage for handling unknown names improve the quality of the mined patterns. The Acquisition of DomainSpecific Patterns [1] revolves around Data Collection, NLP and advanced text processing techniques, Pattern
Mining and Scenario Based Selection.
Summarization Frameworks rely on Conditional Random Fields and Contextual Modelling by Sentencelevel, Non-Textual and Textual Features [3]. An automatic general purpose text summarization tool
would be of immense utility in this age of information overload. Using the techniques used (by most
automatic hypertext link generation algorithms) for inter-document link generation, intra-document links
are generated between passages of a document. Based on the intra-document linkage pattern of a text, the
structure of the text is determined.
(C) Text Simplification
Text Simplification is enabled by Proper Corpus Analysis and Dynamic approach to Text Alignment. The
task of text simplification aims to reduce the complexity of text while maintaining the content. Text
compressions or simplification should be grammatical, and they should retain the most important pieces of
information. The dataset required for testing text Simplification techniques need to be of high quality. The
Proceedings of the VLDB Endowment, Vol 7 are utilized for the vast use of diagrammatic representations
[6].
(D) Web Crawling
A program which browses the World Wide Web in a methodical, automated manner is called a web crawler.
This process is called Web crawling or spidering. Many legitimate sites, in particular search engines, use

Block Diagram Components Information Extraction and Grading of Research Paper based...

549

spidering as a means of providing up-to-date data. Many applications today, are dependent on the internet
for dynamic and latest reliable information. For extended meaning and information retrieval, the web is
utilized, without which, in-depth analysis of data is impossible. There are many commercial and opensource web-crawling techniques and platforms which ease the process of web-based searches like Scrapy
and the Wikipedia Python API.
(E) Text Extraction And Analysis
Gupta et al. [9] suggested Text Summarization Extractive Techniques. An extractive summarization to
method consists of selecting important sentences, paragraphs etc. from the original document and
concatenating them into shorter form. The importance of sentences is decided based on statistical and
linguistic features of sentences. An Abstractive summarization method consists of understanding the original
text and re-telling it in fewer words. It uses linguistic methods to examine and interpret the text and then to
find the new concepts and expressions to best describe it by generating a new shorter text that conveys the
most important information from the original text document.
Traditional scientific papers are unstructured documents, which are difficult to meet the requirements
of structured retrieval. Hence, how to extract and analyze the structured information of the papers becomes
a challenging problem. Jianguo et al. [9] proposed a structured information extraction algorithm for
unstructured and/or semi-structured machine readable documents.
Another approach given by Chuang et al. [7] is where text is extracted based on sentence segmentation.
The basic idea here is to separate out units (segmentation by cue phrases). Sentences are broken into
segments by special cue markers. Each segment is represented by a set of predefined features. This cue
phrase is used here because which it connects two segments and extracts text. The complexity of this
approach is O(n).
III. OUR IDEA
The contributions of our paper are: (1) Conversion of images to histogram and detecting graphical images
based on continuity of pixels. (2) Analysis of Extracted text based on several metrics and grading the
papers based on these metrics. (3) Effortlessly acquire and manage information about the block diagrams in
research documents. This project revolves around the modules encompassing Image Processing and Text
Processing.
Since OCR can handle only monochrome images, we pre-process the images. The High speed Tesseract
OCR engine uses optimized algorithms to perform excellent recognition in short time. Speed can be further
improved through multi-threading and optional GPU acceleration. Python Tesseract Leptonica Software is
the key to the text extraction from the block diagrams after the image is extracted from the portable document
format file.
For more accurate results, the Auto-Spellcheck libraries of Python, namely PyEnchant is used. The
Spelling Rectification module is essential due to the discrepancies that occur in the extractions of text from
images. Compilation of descriptive text phrases of the components involved in the block diagram uses
Pythons Natural Language Toolkit (NLTK) for processing the natural language which is available in raw
text format.
The input research document is converted to a text file to proceed with processing. Using the NLTK
Sentence Tokenizer, the insignificant sentences are eliminated. The grouping of data regarding each
component is done by means of the Phrase-Matching Algorithm. Local Search involves the processing of
research paper text for the descriptions that occur explicitly. For words that do not have necessary information
within the tokenized text, Web-Crawling broadens the horizons from which the meanings and information

550

R. Siva, G. S. Mahalakshmi and S. Sendhilkumar

can be obtained. Web-Crawling is very essential in the extraction of dynamic and more refined information
from the Internet and aids the description extraction process for uncommon terms present in the input set of
documents presented. Finally, the processed text as a result of local and global search is presented in a
consolidated manner which contains the components and their corresponding description for clarity in
understanding the importance of each block diagram entity.
It also grades the quality of research paper based on the description that is given for graphical images
present in it. The grade is given based on the justification given for explanation of the graphical images.
Since graphical images explain difficult contexts with ease, a good explanation for such figures would
improve the quality of the research paper. Analysis of such a text which explains the images would give
great scope for criticizing the amount of work the researcher has put into the research paper. Thus based on
text analysis metrics, we aim to grade the research paper with justification. The dataset is collected from
VLDB.
IV. BLOCK DIAGRAM COMPONENTS INFORMATION EXTRACTION
(A) Text Extraction from Image
Using the Ghostscript software based on an interpreter for Adobes PostScript and Portable Document
Format page description languages, the PDF Image is extracted and stored separately as a png file. Extraction
of text involves text region detection, text localization, tracking, character extraction, enhancement, and
recognition of the text from a given image. Variations in text may occur because of differences in size,
style, orientation, alignment of text. Low image contrast and composite backgrounds serve as a problem
during the extraction of text. Using the Python Tesseract OCR Engine, we aim at processing the image with
dilation, monochrome filter and improving the image quality. Accurate Text Recognition is required and
with enhanced image processing and text detection algorithms, Tesseract OCR can easily recognize difficult
documents of poor image quality. Parameters can be used to hint favouring accuracy over speed and training
the engine improves the results. The Extraction of the text from the block diagram is shown in Fig.3 which
is done using Tesseract and enhanced with the Ghost Script Software. Figure 3 shows the topology of
Block diagram components extraction.
(B) Spelling Rectification
The text that is extracted from the image is processed and prior to Natural Language Processing, the spelling
of each component are checked and corrected for appropriate results. Using the Python API for Spell
Checking, namely the PyEnchant, the suggestions help in rectifying the errors in the text extracted. The
Spelling Rectification Process is improved by adding a domain- based dictionary to increase the efficiency
and accuracy of correcting the components extracted, as the additional dictionary expands the collection of
possible words that can occur in technical documents and research papers.
(C) Sentence Tokenization
The Research Papers contain text pertaining to the components involved in the block diagram that are
required for analysis. The document is searched and the sentences that are related to the components are
tokenized for obtaining the object, relation and the subject. Using the Tokenizer package supplied by the
NLTK Library, namely the Punkt Sentence Tokenization model of NLTK tokenizer tokenizes the paragraphs
into sentences.
(D) Web- Based Refining
For the components without any description in the paper, web-crawling is used for obtaining an appropriate
description to widen the scope of search. We restrict the search within the top few links and try to retrieve

Block Diagram Components Information Extraction and Grading of Research Paper based...

551

a suitable meaning of the component. With the help of the keywords obtained in the document, the most
appropriate context- based description is coined to complete the consolidation of all the components with
their corresponding description. The Wiki API enhances the definitions that are obtained from the known
web encyclopaedia, Wikipedia.org.
(E) Sentence Consolidation
Using the phrase-matching algorithm, the essential sentences required for the formulation of the component
definition are extracted and presented in a concise form for validation. The local and global searches
produce results which are consolidated for the final presentation of the technical terms with their information.
V. GRADING OF RESEARCH PAPERS
The block diagram of the proposed system is given in Figure 1. Each block is explained with the methodology
used.
(A) Pre-Processing
Research paper is given as input in the form of PDF. The PDF contains graphical images and no other
images like logos, tables etc. The preprocessing step to be done is Image Extraction from the PDF files.
These images are extracted using A PDF Image Extractor tool and then the extracted images are stored in a
separate folder. This will evaluate to 100% if there are no logos or tables in the PDF format of the research
paper.
(B) Histogram Conversion and Image Detection
Inputs given are the images which are extracted in the preprocessing step. An image is converted
to histogram using matlab functionality. Here logos and tables are converted to histogram as the
functionality considers them as images. This will reduce the accuracy. The histogram files are stored in
a separate folder. A histogram is non-continuous for graphical image and it is continuous for the rest of
the images.

Figure 1: The diagrammatic representation of functional blocks of grading based on graphical images explanation and the
Block diagram component information extraction topology

552

R. Siva, G. S. Mahalakshmi and S. Sendhilkumar

Based on this, we detect the graphical images. Both normal images and pie-charts have structural
similarity due to continuous pixels. So it will assume even a pie chart as a normal image which reduces
accuracy. The rest of the graphical images are detected and their figure number is stored.
(C) Conversion of PDF to Text
Research paper is given in the form PDF. It is assumed to be free from captions. PDF file is converted
to text files using pdf to text conversion java code. The itext jar file of java library facilitates the
conversion. Output is the converted Text file .The converted text files do not contain the captions given to
images.
(D) Text Extraction
The research paper is subjected to a search based on keywords (the word figure), then the related text will
be extracted using java code. Input is the Text file along with keywords. Lines of text containing the
keywords are obtained as output. This will leave out the explanations which doesnt contain the keywords
explicitly .For example in some exceptional cases, the textual explanation will not contain keywords (eg.
as shown above), and explanation of the figure would be followed. The system might fail to extract these
texts (since no keywords like fig/figure were found). This might increase errors. Even a single-line textual
explanation will be extracted, which reduces the efficiency of the system.
(E) Text Analysis and Grading
Several metrics like key-word count, stopword count, line count measure, spelling check, and reference
word check and noun format check measure have been implemented to grade the research papers extracted
textual explanation. Graphical images are pictorial explanation. They replace large textual explanation.
Hence the line count measure should produce a low value as result.
The number of lines should not be too less or too more. If the number of lines explaining the figure is
bulky, it reduces the quality. So we use this as one metric for grading. Usually the explanation containing
the keywords can refer to the keyword once. Repeated reference annoys the reader. Hence the measure of
keyword count should be as low as possible. The next metric used is spell check. We check if the extracted
text is free from spelling mistakes. If there are more errors in spelling, the grading decreases eventually.
Graphical images are best explained when there is sufficient information provided about the peaks,
slopes, axes, coordinates, min and max values. Such reference words are indexed and checked. If the
extracted text contains such words then the explanation is of good quality. The stop word frequency is
expected to be lower. The extracted text is subjected to stopword count. Textual explanation of the graphical
images is themselves not more than a few lines. Out of these lines, if the stop words count is more, then not
much explanation is provided. The stop words are detected and a count is maintained. If the count exceeds
threshold the grading decreases as quality faces a fall. More the number of stop words, Lesser the grade is.
VI. EXPERIMENTAL RESULTS
The dataset is collected from VLDB. The Results are calculated based on the block diagram component
extraction accuracy and the average number of words in each component present in the dataset, are three on
an average. This is the determining factor for the proper sentence extraction and consolidation. To evaluate
the text extraction process from the image, it is required to be in high resolution. To do so, the processing
of the image must be done with care as it influences the accuracy of the text extraction. Spelling Rectifications
require the complete dictionary containing technical terms because the domains of the research papers are
scientific and uncommon.

Block Diagram Components Information Extraction and Grading of Research Paper based...

553

Text Extraction Accuracy (%) = (Number of correctly identified components/Total number of components)
X 100 (1)

Figure 2: Sample image and text Extraction using Tesseract

To evaluate the text extraction process, the image is required to be in high resolution. The occurrence of
certain symbols like arrows, colons, bounding boxes around the components hinders the image processing
phase. Pre-processing the image includes the usage of Ghost script and changing the properties of the
image. The text-extraction accuracy is improved with the help of Pythons Spell-checking library called
PyEnchant. The related contents with respect to each component is extracted and presented. This poses the
challenge of identifying only the significant points and eliminating the lines which do not define the
component. Web Crawling from Wikipedia is done with ease but there are difficulties faced while trying to
crawl from google directly. The approach to enhance the significant information extraction is done by
training the dataset with necessary keywords. The presence of definitions in sentences, involving the
components in the research papers aids in faster consolidation of components and web-crawling determines
the time complexity of the presentation of components that do not have local definitions.
The evaluation metrics used for grading part are: Structural similarity measure, pattern matching defect,
Accuracy, error rate, precision, recall, and efficiency. A table with all the evaluation metrics and the percentage
accuracy is shown in Fig.2. The preprocessing step involves Extraction of all images from the research
paper, which is given as input in the form of a PDF. Image Extraction is done using A PDF Image Extractor
tool. This will extract all images present in the paper. It considers even tables and logos as images and will
extract them. This reduces the precision.
A recall of 76.92% shows that most of the extracted text and images are relevant. Lower value of error
rate and pattern matching defect shows that the accuracy of the project is high. Even a single line textual
explanation is extracted which makes the efficiency of the system to be 75%. Structural similarity measure
is 60% since even pie-charts are considered as normal images and might be ignored during image detection.
VII. CONCLUSIONS AND FUTURE WORK
The conclusion of our experimental analysis suggests that 51% accuracy is obtained with the dataset, being
the Proceedings of the VLDB Endowment, Vol 7, 2013-2014. Tesseract [5] is now behind the leading
commercial engines in terms of its accuracy. Its key strength is probably its unusual choice of features.
Its key weakness is its use of a polygonal approximation as input to the classifier instead of the raw
outlines. With internationalization done, accuracy could probably be improved significantly with the judicious
addition of a Hidden-Markov-Model-based character n-gram model, and possibly an improved chopper.
Also we have successfully detected graphical images from the dataset and extracted their textual explanation.

554

R. Siva, G. S. Mahalakshmi and S. Sendhilkumar

Table I
Block Diagram Components Information Extraction Accuracy

No. of
Components

No. of
Papers

Paper
Count

No. of components
Extracted from
block diagram

3
4

1
3

10

11

12

13

15

18
25
TOTAL
51

1
1

1
1
2
1
1
1
1
1
2
2
1
1
3
3
1
2
1
2
1
1
1
1
2
1
1
2
2
3
1
1
2
1
2
1
1

2
2
4
3
4
5
2
3
5
5
6
7
3
5
6
8
4
5
6
7
9
7
9
10
5
11
6
9
6
7
10
6
10
12
13

Local
No. of
Definition Components
Present
to be Web
Crawled
2
2
3
2
3
3
2
3
4
4
3
6
3
4
3
4
3
3
4
4
6
5
6
4
4
8
4
9
6
6
9
5
6
9
10

Successfully
WebCrawled
Components

Accuracy
(%)

0
0
0
1
1
2
0
0
1
1
1
1
0
1
1
3
1
2
1
1
2
2
2
4
0
2
1
0
0
1
1
1
2
2
2

66.67
50
75
60
80
100
33.33
50
83.33
71.43
57.14
100
37.5
62.5
50
87.5
44.44
55.56
55.56
55.56
88.89
70
80
80
36.37
90.91
41.67
75
46.15
53.85
76.92
40
53.33
61.11
48

0
0
1
1
1
2
0
0
1
1
3
1
0
1
3
4
1
2
2
3
3
2
3
6
1
3
2
0
0
1
1
1
4
3
3

51.54%
Table II
Evaluation metrics with accuracy in percentage for grading of papers

Evaluation Metrics

Percentage (%)

Structural Similarity Measure

Pattern Matching Defect
Accuracy
Precision
Recall
Error Rate
Efficiency

60
30
80
83.33
76.92
20
75

Block Diagram Components Information Extraction and Grading of Research Paper based...

555

The extraction takes place post the conversion of the research paper from pdf to text format. The analysis
of the extracted text is done based on the above suggested metrics and a grading is assigned to the research
paper as output. Future work could include better search optimization strategies for combining points for
the components with the use of web-crawler so that technical terms can be extracted from authentic sources.
To increase the success rate of the text extraction, the image must be processed further before passing it to
the OCR Engine.
REFERENCE
[1] Du, M. and Yangarber, R., (2015), April. Acquisition of Domain-specific Patterns for Single Document Summarization
and Information Extraction. InThe Second International Conference on Artificial Intelligence and Pattern Recognition
(AIPR2015) (p. 30).
[2] De Nart, D. and Tasso, C., (2013), December. A Keyphrase Generation Technique Based upon Keyphrase Extraction and
Reasoning on Loosely Structured Ontologies. In DART@ AI* IA (pp. 49-60).
[3] Chan, W., Zhou, X., Wang, W. and Chua, T.S., (2012), July. Community answers summarization for multi-sentence question
with group L 1 regularization. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics:
Long Papers-Volume 1 (pp. 582-591).
[4] Patel, C., Patel, A. and Patel, D., (2012), Optical character recognition by open source OCR tool tesseract: A case
study. International Journal of Computer Applications, 55(10), pp.50-56.
[5] Proceedings of the VLDB Endowment, Vol 7, 2013-2014 http://www.vldb.org/pvldb/vol7.html
[6] Salton, G., Singhal, A., Mitra, M. and Buckley, C., (1997), Automatic text structuring and summarization. Information
Processing & Management,33(2), pp. 193-207.
[7] Chuang, W.T. and Yang, J., (2000), July. Extracting sentence segments for text summarization: a machine learning approach.
In Proceedings of the 23rd annual international ACM SIGIR conference on Research and development in information
retrieval (pp. 152-159). ACM.
[8] Chen, J. and Chen, H., (2013), A structured information extraction algorithm for scientific papers based on feature rules
learning. Journal of Software, 8(1), pp.55-62.
[9] Gupta, V. and Lehal, G.S., (2010), A survey of text summarization extractive techniques. Journal of Emerging Technologies
in Web Intelligence, 2(3), pp. 258-268.
[10] Prabhakar, M. and Chandra, N., (2012), Automatic Text Summarization Based On Pragmatic Analysis. International
Journal of Scientific and Research Publications, 2(5).
[11] Edmundson, H.P., (1969), New methods in automatic extracting. Journal of the ACM (JACM), 16(2), pp. 264-285.