Image reversible data hiding is the process of embedding information in an image in a visually imperceptible way with the result that both the original image and the embedded information can be exactly recovered. In the traditional approaches of histogram shifting and causal window-based difference/prediction-error expansion, pixel value prediction methods have been rather crude and simple. In this paper we propose a novel multi-resolution reversible data hiding framework that makes it possible to utilize the state-of-the-art image interpolation methods for pixel value prediction. It turns out that less prediction error can be achieved in the proposed framework, which leads to larger embedding capacity. Experimental results confirm the performance gain of the proposed approach over the traditional ones.