- 最新
- 投票最多
- 评论最多
You have to train an AI with the same material it'll be used to classify, not something close. In this case you need to train it with "emoji on wallpaper" not just "emoji"
So you'll need a library of images that you'd normally classify with the AI, that a human has manually classified, and use those to train it.
Rotation is a good start, but you might want to go for more robust transformations and a more robust model based on fine-tuning. It really depends on what you plan to feed it later - how much variability there will be in the images you expect to do inference on.
This blog talks about transfer learning / fine-tuning and could be helpful. https://aws.amazon.com/blogs/machine-learning/run-image-classification-with-amazon-sagemaker-jumpstart/
Then take a look at the methods described here for data augmentation: https://www.tensorflow.org/tutorials/images/data_augmentation
Good luck!
What do you mean by more robust transformations? I plan to feed it with wallpapers, that are similar to this in question.
I have around 1200 different types of emojis. On each wallpaper can be any combination of any length of emoji. So it will be very hard to prepare these wallpapers. Do you have any idea? How can I solve this?