Data Augmentation through Luminance Transformation
Sang-Geun Choi1, Chanil Park2, Sooyoung Cho3, Chae-Bong Sohn4
1Sang-Geun Choi, Department of Electronics and Communications Engineering, Kwangwoon University, Gwangwoon-Ro, Nowon-Gu, Seoul, Korea, East Asian.
2Chanil Park, Department of Electronics and Communications Engineering, Kwangwoon University, Gwangwoon-Ro, Nowon-Gu, Seoul, Korea, East Asian.
3Sooyoung Cho, Department of Electronics and Communications Engineering, Kwangwoon University, Gwangwoon-Ro, Nowon-Gu, Seoul, Korea, East Asian.
4Chae-Bong Sohn, Department of Electronics and Communications Engineering, Kwangwoon University, Gwangwoon-Ro, Nowon-Gu, Seoul, Korea, East Asian.
Manuscript received on 10 June 2019 | Revised Manuscript received on 17 June 2019 | Manuscript Published on 22 June 2019 | PP: 847-851 | Volume-8 Issue-8S2 June 2019 | Retrieval Number: H11430688S219/19©BEIESP
Open Access | Editorial and Publishing Policies | Cite | Mendeley | Indexing and Abstracting
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open-access article under the CC-BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Abstract: In deep learning, learning a model for data classification requires a large amount of data. Deep Learning has been applied to many areas over the years. In many areas, the amount of data is increasing for more accurate model learning. However, there are areas where data acquisition is difficult or limited. We will present a data augmentation method to solve this problem. In addition, it will lower the sensitivity to the light of the network via a data augmentation through a luminance transformation. We selected the YOLO v2 model to identify and compare the results through the proposed method. YOLO v2 is a version that improves both performance and speed over the previous YOLO model. The datasets for learning was PASCAL VOC 2011. For the comparison of the results, we will train the network by each dataset that performed data augmentation and those that were not. We will compare the performance of the two networks through three indicators. First, we will compare the change of loss according to epoch. Next, the accuracy of the network through the test set. Finally, we will see the results of the Class Activation Map (CAM)Data augmentation was performed by modifying luminance in this paper. The dataset used in the experiment was PASCAL VOC 2011 dataset, which consists of 28952 images and consists of 20 classes in total. The results obtained by training 50 epochs of network are compared. As a result of the test, the mAP of the general dataset is 0.6, and the proposed dataset is 0.7. In the test with the adjusted image of the light source, the mAP of the general dataset was 0.6, and the proposed method was 0.7. In the case of a network using a general dataset, the object is not recognized even when the image is detected through a different light source. In the CAM results, it was confirmed that the strength of catching features in the two networks was slightly different. In this paper, we confirmed that network accuracy improves when data augmentation is performed. Furthermore, the sensitivity of the network to light is also reduced. Based on these results, it is expected that the data augmentation will enable the more accurate network implementation in the area where the dataset is insufficient, and the performance when the data augmentation method is changed according to the experimental environment is expected to develop
Keywords: Deep Learning, YOLO v2, Data Augmentation, Luminance Transformation, Class Activation Map (CAM).
Scope of the Article: Communication