Fusing label distribution and on-hot encoded labels

A while ago I came across a paper for image classification that utilized both label distribution and one-hot encoded labels to classify images. An image has a label distribution for all classes (4 classes) if the label is like:

t=[0.1, 0.6, 0.1, 0.2]

and it's corresponding one-hot encoded label is:

t_one = [0, 1, 0, 0]

Since the label distribution was available, the authors were using both labels in the process of training to take advantage of latent information in the label distribution.

I'm not finding the research anymore and it had all these processes in a nice diagram. Unfortunately, I forgot to save it. Do you know of such research or other related research?