Title: Graph-type Classification Based on Artificial Neural Networks and Wavelet Coefficients

Year of Publication: Dec - 2015
Page Numbers: 77-85
Authors: Sarunya Kanjanawattana, Masaomi Kimura
Conference Name: The Second International Conference on Digital Information Processing, Data Mining, and Wireless Communications (DIPDMWC2015)
- United Arab Emirates


Typically, in many documents, there are graphs that authors use to illustrate their points. To utilize the information in such graphs via an automated process, we must extract semantic information from the graphs depending on the type of graph used; however, it is difficult to identify the types of graphs using an automated computer program, even though it is easy for humans to do. Several studies have developed image classification methods to solve a variety of problems; however, applying these studies to graphs is not straightforward. Moreover, characteristics of graphs have not been fully clarified and can be different even for graphs of the same type, which causes traditional systems to be further inapplicable. In this study, we propose a method to classify graph images based on their types. Our approach consists of two major steps. First, a preprocessing step applied wavelet, Fourier and Hough transformations to create one-dimensional images. Second, a neural network step classifies the constructed one-dimensional data into a specific type of graph. Given our work, the key objectives of this study are to propose an innovative method for correctly classifying types of graphs and to extract graph characteristics suitable for identifying these types of graphs. We evaluated our method by comparing the performance among of our method to convolutional neural network and support vector machines. From our results, we determined that our approach was able to obtain an accuracy of over 80%.