Will doing image cropping excessively lead to poor performance of deep learning model?

Question

I am building a deep learning model to recognize images. From what I have read, data augmentation such as random cropping of images will lead to less overfitting of the model. However, I am not sure if doing it excessively will lead to a worse model. Of course, I can try one with more cropping and one with less cropping. But the problem is how can I know if the problem arises from the number of crops made.

Will making all possible crops of a size m x m from an image of size n x n lead to a better performance of the model?

I believe that it will. My reasoning is this: when we train a deep learning model, we look at the train loss and validation loss and train the model until it has a very low loss. Assume that initially we have a train set of 1000 images and the model takes 100 epochs to train. Now, we crop 10x extra images from the original train set. Each epoch can now be seen as equivalent to 10 epochs in the previous model which has less training data. However, each of training data in this 10 epochs is slightly different, compared with 10x duplicates in the previous model. Surely, this will lead to less overfitting. Is my reasoning correct?

In that case, is there any downsides of cropping all possible images of a smaller size, assuming that we have enough computational resources?

Currently I am looking at cropping all possible 64x64 images from an image of 72x72 which gives me a total of 64 new images per original image.

I have not seen any papers that deal with this. I will appreciate if someone can point me to one.

Answer 1

No, it will not harm performance, it will add, however, a few milliseconds to the general process. Maybe the best answer you could get is experiment with different approaches.

Answer 2

From my experiments, the performance gain is negligible and might lead to overfitting since similar images are repeatedly seen.