딥러닝 기반 학습이미지 생성 모델과 시멘틱 세그멘테이션을 활용한 토지피복분류 효율성 평가에 관한 연구

Abstract

In classifying land covers based on deep learning, this study deals with the generation of generative adversarial network (GAN)-based images to quantitatively complement a small amount of image learning data sets and it also deals with the verification of the accuracy of classifying land covers using the said generated image data sets. In the field of image reading, in order to recognize and extract objects with high reading accuracy using artificial intelligence, a large amount of learning data sets to be applied to the algorithm are required. However, it is a reality that not only is there a shortage of image data sets that can be used jointly, but also a lot of time, manpower, and high cost are required to generate image data. Therefore, in this study, first, a small amount of initial aerial image learning data sets were used to create oversampled image learning data sets using CycleGAN's generator neural network. An experiment was conducted to utilize the oversampled data as additional learning data sets by evaluating the quality of the data using SSIM (Structural Similarity Index Map). In addition, in order to verify the utility of the generated oversampled learning data sets, a learning data sets scenario was constructed based on the SSIM index and applied to the deep learning algorithm for land cover classification. As a result, the classification accuracy was improved by up to about 9%. Thus, it was proved that a method of supplementing a small amount of initial learning data sets using additional learning data sets generated through GAN is effective. In addition, it was verified that the method can be used as a methodology that can additionally secure the amount of learning data sets that is evaluated to have a very large impact on deep learning performance