Data augmentation adds more variety to what the CNN sees during training. This augmentation could be a single pixel shit that introduces “new” data sample to the CNN. Of course more variety in training data increases the robustness of your model. But remember to use a realistic data augmentation that can represent expected conditions on the testing images. Check thisArticle that discusses this issue (it is mainly about color augmentation, but we can generalize this issue to other augmentation techniques).