Multiclass prediction gives high confidence score for one class when none of the classes are present in the image

I have trained a model with 2 classes using transfer learning from resnet101. The prediction works quite well when there is at least one of the classes present on the image being predicted. However, when none of the classes are present on the image, I always get high confidence score ~99% on the first class. I have struggled for a while with this and couldn’t eliminate the problem.

One thing that I will try out is to include an extra ‘empty’ class and retrain the model with the new dataset, but what I am afraid in this approach is that it might just ran into the same issue, or the model might focus on something about the empty images and create new problems for me.

Any other suggestions that I might try out? Appreciate your help.