Figure 3.4: Processes of all experiments: (a) applying the CNNs to 1Dimg and 2Dimg in CNN 1Dimg and CNN 2Dimg respectively, (b) applying the SVMs to WLHT in SVM WLHT, (c) applying the ANNs to WLHT in ANN WLHT, (d) applying the SVMANN to WLHT in SVMANN WLHT, and (e) applying the ANNSVM to all datasets in ANNSVM 1Dimg, ANNSVM 2Dimg, ANNSVM WL,
ANNSVM HT, and ANNSVM WLHT
for this method because it often provides good results when performing with compli-cated information, and my datasets are nonlinear data. After I combined the outputs of the SVMs with a class label, I applied the ANNs to the outputs and obtained classification results. The last experiment was ANNSVM ALL (i.e., Figure 3.4e) in which I used my proposed method, i.e., the ANNSVM. Note that ANNSVM ALL represents the experiments conducted by ANNSVM with all datasets used in this study. In ANNSVM 1Dimg and ANNSVM 2Dimg, the ANNSVM was applied to 1Dimg and 2Dimg to compare results with those of CNN 1Dimg and CNN 2Dimg.
In ANNSVM WL and ANNSVM HT, I also applied the ANNSVM to WL and HT because I needed to evaluate how data affected the system. Further, the most signifi-cant experiment was ANNSVM WLHT, in which I presented the performance of my main method applied to WLHT to indicate the effectiveness of my proposed idea.
To evaluate my approach, I compared results to other experiments that also used WLHT. Regarding SVMANN WLHT and ANNSVM WLHT, I conducted experi-ments to show the difference of performance in a case that I switched their orders.
A motivation for rearranging the order of algorithms was to examine whether the results had been influenced by algorithms switched.
In this study, accuracy values of each dataset showed the performance of each method. These values represent are the proportion of the total number of predictions that were correctly classified.
Initially, I classified training instances into three classes that are bar graph, 2Dchart and pie chart, with 303, 322 and 297 images for each class respectively. The number of images is 922 images in total. The graphs had been selectively gathered from the Internet because an amount number of graph images could be collected for training process comfortably. Moreover, I needed the data to find the suitable classification model for the graph images; hence, it should be not matter wherever graph images originally came from. I manually normalized the collected images by eliminating unused areas, such as unnecessary text. Moreover, I evaluated the experiments with tenfolds cross-validation because such an approach can mitigate the problem of overfitting.
Note that I trained the ANNSVM models individually. Each model used independent parameters estimated by SVMs and ANNs parameter estimations. In practice, I applied ANNs to my data. Then, I obtained a new dataset from ANNs outputs, it will be an input for SVMs for classification.
3.3.2 Results
Figure 3.5: Results from CNNs and ANNSVM that used 1Dimg and 2Dimg:
(a) table statistically showing summarized results and (b) bar graph graphically illustrating results from these experiments
Here, I checked the obtained results by myself. I compared the results of CNN 1Dimg, CNN 2Dimg, ANNSVM 1Dimg, and ANNSVM 2Dimg to confirm the validity of ANNSVM when applied to images. I compared my classification system to CNNs because it is a powerful and popular image classifier. The 1Dimg repre-sented the dataset of one-dimensional images, while 2Dimg reprerepre-sented the dataset of two-dimensional images. Results are shown in Figure 3.5. The CNN 1Dimg and CNN 2Dimg provided similar accuracies, approximately 0.33, which were close to the
Figure 3.6: Results from ANNSVN that used WL and HT a) table statistically showing summarized results and (b) bar graph graphically illustrating results from
these experiments
results of ANNSVM 1Dimg and ANNSVM 2Dimg with the linear kernel, i.e., ap-proximately 0.35; however, the experiment of my proposed method (i.e., ANNSVM with the RBF kernel) presented largely different results. In 1Dimg, the accuracy increased to 0.79. Comparing this results to those of 2Dimg applied to my proposed method, the accuracy was approximately 0.56. Thus, compared to two-dimensional images, the one-dimensional images were a better candidate for graph-type classifi-cation using ANNSVM with the RBF kernel.
To identify which features of data influentially impacted data separability, I conducted experiments for ANNSVM with WL and HT (i.e., Figure 3.6). The WL contained only wavelet coefficients, whereas HT included only results of the Hough transformation. I found that, again, results obtained via the linear kernel were not significant; however, using the RBF kernel, accuracy for WL was higher than that of HT, indicating that wavelet coefficients provide influential features that make data separable.
Figure 3.7: Results from SVM, ANN, SVMANN, and ANNSVM that used WLHT: (a) table statistically presenting summarized results, (b) bar graph graph-ically illustrating results from SVM WLHT and ANN WLHT, and (c) bar graph
graphically showing results from SVMANN WLHT and ANNSVM WLHT
I conducted SVMs, ANNs, SVMANN, and ANNSVM with WLHT constructed from my preprocessing method. Results are shown in Figure 3.7.
The results of SVM WLHT showed accuracy for a RBF kernel as slightly better than that of a linear kernel as indicated in Figure 3.7b. They were on average 0.85 for the linear kernel and 0.86 for the RBF kernel. As for the outcome from ANN WLHT, it was moderately 0.83. Apparently, accuracy in SVM WLHT which used the SVMs was slightly more appropriate.
In SVMANN WLHT, I found accuracies for WLHT were stably high, with an average of 0.9, independent of wavelet families. The highest accuracy in SV-MANN WLHT was 0.905 in the case of Symlet 2. As for ANNSVM WLHT, the average accuracy was 0.87 for the RBF kernel (i.e., my proposed method) and 0.35 for the linear kernel. Though the average accuracy for my proposed method was slightly lower than that of the SVMANN in SVMANN WLHT, the highest accuracy among all experiments was 0.91 in the case of ANNSVM WLHT in which ANNSVM was applied to data obtained by Coiflet 1 (i.e., Figure 3.7c).