Can I use your RL technique on a model which take in input global feature image?

The model I'm training is fed with global region feature of each image, it doesn't use local feature based on regions. Do you think I can apply your technique to improve my CIDEr score's model?