Is normalization needed for prediction?

If a model is trained by image after normalization, is normalization needed for the prediction of one single image? My intuition tells me the result will not be affected without normalization, but I am not sure.

Yeah I think that you should apply normalization to your test image if your model is trained on normalized inputs because then your model may be experiencing some strange input at test time and which may not be so good for the task that you are upto.