The Effect of Augmentation on Classification Algorithm to Determine Photo Angle

A technique for recognizing the angle in the suitable photo is required to ensure that the information is sent correctly. The image classification method is one approach for determining the angle of a photograph. Support Vector Machine (SVM) and K-Nearest Neighbour (K-NN) classification algorithms were employed in this study. The first total dataset collected from photographic websites contained 2,700 records and was split into three classes: high angle, eye level, and low angle. In addition, there is a second dataset with 679 datasets that was also directly downloaded from the same website. This dataset is separated into 5 classes: bird's-eye view, high angle, eye level, low angle, and frog's-eye view. An augmentation method is used with three strategies, including random rotation, random noise, and grayscale, to increase the number of datasets that have been collected. The entire dataset that will be classified contains 5700 data for the first dataset and 3179 data for the second dataset after the augmentation method adds data from the first and second datasets, totaling 3000 and 2500 data, respectively. The dataset is then divided into training and testing data using a ratio of 80:20. We used 2 scenarios in developing the model: the model with additional augmented data and the model with original data. The Support Vector Machine algorithm and the K-Nearest Neighbor technique in the 3-class and 5-class classification models provide low accuracy results, particularly in the range of 50%, for testing the first scenario. In the second scenario test, the KNN technique with K = 1 achieved an accuracy of 95.10% and an execution time of 3 seconds with a number of classes of 3, compared to 91.70% and 7 seconds with a number of classes of 5. In the SVM method, the RBF kernel was able to achieve an accuracy of 94.10% with an execution time of 6 minutes 32 seconds for the number of classes 3, while 90.60% with an execution time of 26 minutes was achieved for the number of classes 5 using the same parameters (C = 10, gamma = 0.001, and RBF kernel).