Institutional Repository of Vilnius University
Lentelių struktūrų aptikimas ir informacijos nuskaitymas buhalteriniuose dokumentuose /
Abstract
dc:descriptionComputerization and technological advancements have made paper documents obsolete although many companies nowadays still use or have a lot of old paper documents. These paper documents slow down work efficiency significantly, because digital documents offer faster accessibility and easier data manipulation. One of the work fields with this problem is accounting. Accountants encounter paper documents almost daily and have to manipulate significant amounts of data in those paper documents. In order to alleviate this aforementioned problem paper documents must be digitalized. The hardest problem in document digitalization is table detection and data extraction. Therefore, this problem will be addressed in this thesis. A select few methods TableNet and DeepDeSRT were analyzed and presented. Out of the analyzed methods TableNet was chosen for implementation. Implementation process was thoroughly documented highlighting problems solved during implementation and aspects that were unclear after analyzing TableNet documentation. After implementation the neural network model was evaluated with accuracy, precision, recall and f1 score measures and compered to TableNet author documented numbers. The results implied that the model was correctly implemented. Three different model improvements were suggested and implemented. The first improvement was based on principals found in Single Shot Multibox Detector model. The second improvement was based on additional data pre-processing used for model training. The third was based on the results found in previous improvements. All three model improvement results showcased different improvements in the model performance. Each of the implemented improvements was evaluated with aforementioned metrics and compared to the base model implementation. Each improvement with the exception of the first showed improvements in several evaluation categories. Although, the first model improvement showcased better segmentation capabilities.gnition and Data Extraction from Accounting Documents.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2020
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Šmigelskis, Marijus,
- Contributors dc:contributor
-
- Valaitis, Vytautas
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD81704981&prefLang=en_US