Is image sharpening a good idea for data augmentation?

Question

I'm training segmentation networks and while the dataset is somehow decent (~5k images) I wanted to augment it, so far I'm trying:

• RandomFlip
• RandomRotate
• RandomBrightness changes
• RandomShadows

Due to constraints of the problem I can't do random crops or shifts. Other than those augmentations I was looking into image sharpening, and was wondering if it could be a good candidate for dataset augmentation. I could find it in some web articles and many augmentation projects on Github, but I can't find any solid papers that refer to it as a possible augmentation technique. Does anyone have some experience/tips on the matter?

Answer 1

At first, I have to mention that $5k$ cannot be considered as a large dataset for training a deep neural network. Anyway, about the question. In general, yes you can, but you have to be aware of some points. Data augmentation can be helpful or it can damage your entire predictions. The reason for each is that whenever you utilize data augmentation, you are somehow changing and manipulating the distribution of your data in hand. Because neural networks are considered to deal with random processes which their behaviour is iid, your data samples should not be dependent on each other. There is also another perspective. Your training data should have the same distribution as your test data. If you are sure that while testing what you are going to face can be sharpened images, so go ahead and carry out that, but if you're sure that your sensors are placed in a noisy situation, something you can encounter in self-driving cars, you can be pretty much sure that your raw data is blurry, and it is almost impossible to have sharp images due to velocity unless you take a preprocessing step and after that you feed it to your network.