Image Preprocessing
Resizing, normalization and data augmentation.
What you will learn
- Listing the preprocessing steps
- Understanding the benefit of augmentation
Common steps: resizing to one size, normalizing the values to [0,1], converting to grayscale when colour does not matter, and data augmentation (Augmentation: flipping, rotation, changing lighting) to diversify training.
Transfer learning requires the size and normalization the original network was trained with, and the visual tools do this automatically.
Implementation steps with the tool
- Teachable Machine: photograph the objects from different angles and in different lighting (manual data augmentation).
Quick review question
Why might flipping an image horizontally do harm when recognising letters?
