반응형
- Deep learning fundamentally requires a large amount of data for effective training.
- Additionally, to solve the chronic problem of overfitting in deep learning, it needs sufficient high-quality data.
- However, increasing the amount of data requires significant cost and time, and in some cases, it can be difficult to even collect or process the data.
- To address this issue, various Data Augmentation techniques have been developed to create new data using existing data.
- While there are various augmentation methods depending on the type of data, we will only cover those related to image data.
Basic image manipulation | Deep learning approach | Meta learning |
Geometric Transformation | Adversarial training | Neural Augmentation |
Color Space Transformation | GAN Data Augmentation | Auto augmentation |
Mixing Images | Neural Style Transfer | Smart augmentation |
Random Erasing | ||
Kernel Filters |
- While there are various methods of data augmentation, there are important principles to follow.
- Semantically Invariant Transformation means that augmentation should be performed while preserving the important aspects of the data.
#1 Basic Image Manipulation
- Geometric Transformation is a method of creating new images by applying Crop, Rotate, Contrast, Invert, and Flip operations to existing images.
- Color Space Transformation is a method of creating new images by adjusting the RGB values of existing images.
- Mixing Images is a technique that performs Weighted Linear Interpolation between two images using a λ value between 0 and 1, where the label is also assigned proportionally to the λ value.
- Random Erasing creates new images by erasing random areas of the image.
- We can also combine Basic image manipulation methods:
- Cutmix combines Mixing images technique and Random erasing technique.
- It works by drawing a box on image A and erasing it, then filling that empty area with a patch extracted from image B.
- PuzzleMix is an improved version of CutMix.
- It mixes two images while preserving important features from both images.
- Cutmix combines Mixing images technique and Random erasing technique.
#2 Deep learning approach
- Adversarial training:
- Adversarial attack refers to presenting intentionally manipulated input values (adversarial examples) to the training model to make the DNN produce incorrect results.
- Adversarial training is a learning method that creates multiple adversarial examples, presents them to the model, identifies under what circumstances the model makes misclassifications, and then modifies the model to improve overall performance.
- GAN Data Augmentation:
- Uses GAN (Generative Adversarial Networks) models to generate samples similar to existing data to increase the amount of data.
#3 Meta Learning
- Autoaugmentation
- Among numerous data augmentation methods, AutoAugmentation is a model that suggests techniques suitable for a given dataset.
- Google implements AutoAugmentation by training with PPO (Proximal Policy Optimization) to find the optimal combination among 16 commonly used data augmentation techniques.
- However, it requires extensive computational resources, has a large search space, takes a very long time, and can only be used in limited environments.
- Several methods have been proposed to improve the calculation speed of AutoAugmentation.
- Examples include Population Based Augmentation, Fast AutoAugment, and Faster AutoAugment.
- RandAugmentation
- While previous techniques focused on finding suitable augmentation methods, RandAugmentation doesn't seek a specific model but instead randomly selects and applies augmentation methods for each batch.
- While its performance is similar to other models, it has the advantage of having simpler code.
Reference
When I say something's gonna happen, it's gonna happen.
- Conor Mcgregor -
반응형