graph image due to low accuracy. The wavelet coefficients could identify the domi-nant characteristics from the graphs outperforming than the Hough transformation, especially Coiflet 1. However, to make my dataset more separable, I decided to as-semble the result data from wavelet coefficients and Hough transformation together.
ANNSVM provided a remarkable accuracy, up to 0.91. Note that this system can classify the types of graphs, i.e. bar graph and 2Dchart, effectively. During I im-plemented the search engine system (the final system), I could distinguish the types manually because a number of data were small. Moreover, to show a concept of the classification system, I omitted this process to increase a precision of the overall system. However, if the size of data volume was increased, this system was surely necessary.
I presented the graph components extraction and identification system pre-sented in Chapter 4. To accomplish the problem, I adapted DBSCAN to automat-ically estimate a proper Epsilon value for each data. I evaluated the system by benchmarking the performance between the proposed system with flexible Epsilon and another with fixed Epsilon. The proposed method provided very high identi-fication rate, 0.93; whereas the other was only 0.31. This was because it returned the results as original images that contained many irrelevant parts. In other words, DBSCAN with fixed Epsilon could not detect the legend properly.
The OCR-error correction had been proposed in Chapter 5. I used a con-structed ontology and other online ontology to suggest correct words to error ones.
The results of correction were validated by comparing the performance between the proposed method, which used an ontology, and a traditional one, which used edit distance. The proposed method provided the highest F-measure and accuracy, 0.86 and 0.84 respectively.
Not only the graph components are important features to create the ontology for my semantic search engine, but also graph information resided in a data section of the graph is necessary. I proposed a method of graph information extraction in Chapter 6. It used to extract tendency of data plots, bar height, and significant relationship of graphs. I made some simulations to prove the validation of the system.
I found that I could obtain a number of tokens unavailable in the captions of the
graphs but were instead taken from other graphs sharing the same concepts, such as quantification and plasma. This proved that my proposed method could introduce new knowledge by utilizing ontology concept.
I integrated entire proposed systems to a semantic search engine system in Chapter 7. I implemented a web-based application connected to my constructed ontology. The ontology recorded all information extracted from the previous meth-ods. The F-measure from my system was much higher than the ES-based system, which represented a better quality of the search engine system. This summary also agreeable to the result from questionnaires gathered from the participants.
From the results analysis of the whole study, essential findings in this disser-tation had been found based on the facts of the study. They were listed as below.
• CNN was unsuitable for classifying images if the image edge was only a trivial characteristics for classification.
• The most suitable wavelet coefficient applied to my dataset was Coiflet 1.
• The wavelet coefficients could identify the dominant characteristics from the graphs outperforming than the Hough transformation.
• An order of algorithms was a trivial matter; in other words, SVMANN and ANNSVM were similar.
• The 5-layer ANNs is applicable to my target data.
• Based on my observation, the graph component extraction system can excel-lently handle the graph image who legend quite separates from other surround-ings.
• The graph component extraction can support the OCR-error correction system because it helps to reduce the noise ratio from the input images.
• The correct word suggested from the ontology is more accurate than edit dis-tance.
• In my opinion, a scope of the question to query the information from graphs was more expanded than a traditional search engine that uses keywords and obtains document relevant to the keywords.
• Ontology offers concise knowledge that cannot be done by the traditional search engine.
• Ontology-based search engine provides extended information outperforming than the traditional one.
• Users are satisfied to use the ontology-based search system because most users feel comfortable to use it.
During the studies, not only essential findings, which are corresponding to the facts from the experiments, but the byproduct findings should also be described here. They are the findings derived from the experiments but do not relate to the core of the studies. As the following list, I showed the byproduct findings found in this research.
• A low image quality highly affects to the performance to the systems.
• Irrelevant parts included in the extractable legend also negatively affects clas-sification performance.
• A query time in SPARQL was longer than querying in a database.
• The input image should be cleaned as much as possible to prevent OCR mis-understanding because they interfere the recognition process.
• To realize the characteristics of the frequency domain in an image containing texts, the high-frequency area is denser than other areas in the DFT images.
Similar to the image not containing texts, a difference is that its density in the high-frequency area is smaller than the image containing texts. To sum up, the image with texts has a higher frequency than the one without texts.
• Based on my ontology structure, it is possible to indicate a category that a graph belongs to by using NER class.
• Regards the OCR process, a space between alphabet characters in a word also provides a confusion to the OCR because the OCRCR may recognize the word separately even it should be one word.
I discovered several interesting findings that I had never expected before. I used the ontology-based search engine system and selected a question about graph relationships. I inputted some keywords and obtained the results that I required for.
I obtained several graphs corresponding to the keywords. After I collected the results of graph relationships, I found some similarities. Surprisingly, the obtained graph axis relationships provided around 10% to 20% similarity when compared to the graphs corresponding to the keywords. For example, the keywords “accuracy, per-formance” were inputted to my search engine system and obtain five graphs relating to the keywords, including their graph axis relationships. I examined the similarity among the graphs. The proportion of the graph axis relationship’s similarity was 20%. This represented that the graphs related to the same keywords should have corresponding relationships partially. This finding can be used to discover other graphs that contain the similar relationships in other documents.
Moreover, I examined the results of token relationships by inquiring the sys-tem. The ontology-based search system can cope a variant expression of words. I simulated some keywords that represented a similar expression and inputted them to the system to find their token relationships. The results showed that some same relationships were acquired. The rate of identical relationships appeared was around 30% to 40%. Therefore, the similar keyword expression can obtain the identical rela-tionships. For example, I inputted two set of keywords: “method, performance” for the first iteration and “algorithm, performance” for the second iteration. I collected the relationships from both iteration and compared them to find the identical rela-tionships. I obtained an identical ratio about 36%. With this finding, it is possible to find other relevant description contents, e.g., paragraphs, in multiple documents.
For example, I obtain a token relationship in a sentence, and I find other paragraphs that contain the same token relationship. Since, the relevant description contents should be expanded.