이미지 분류 성능 향상을 위한 무작위성 적용 방법론

초록

Securing various types of training data in image Classification is important for improving performance. However, increasing the amount of original data is cost-limited, so data diversity can be secured by transforming images through data augmentation. Recently, a new deep learning approach using Generative AI models like GAN or Diffusion Based models has emerged in the Data Augmentation task, and reinforcement learning-based methods such as AutoAugment and Deep AutoAugment using existing basic Augmentation techniques are also showing good performance. However, this method has the disadvantage of having a complicated optimization procedure and high computational cost. This paper conducted various experiments with existing methods Cutmix, Mixup, RandAugment. By combining these techniques appropriately, we obtained higher performance than existing method without much effort. Additionally, our ablation experiment shows that additional hyper-parameter adjustments are needed for the basic augmentation types when we use RandAugment. Our code is available at https://github.com/lliee1/Randomness_Analysis.

키워드