Label smoothing is a regularization method, which smooths the real one-hot label such as [0,1] and moves closer to the "middle", such as becoming [0.05, 0.95], so that the trained model will not be excessive Confidence, improve generalization ability.
reference:
Detailed Explanation of Label Smoothing
Label smoothing label smoothing_Braylon1002's Blog-CSDN Blog_label smoothing