In this paper, multiobjective optimization is applied to determine the optimal feature weights for a multi-resolution image quality metric. The optimization is conducted with respect to two aims; maximization of quality prediction accuracy and generalisation ability to unknown images. A thresholding of the optimal weights is applied to further reduce computational complexity of the quality metric. A detailed discussion of the tradeoff between prediction performance and complexity with respect to the threshold is provided. The metric design and evaluation is supported using mean opinion scores from subjective quality experiments that we conducted in two independent laboratories. Comparison to other state of the art quality metrics reveals the ability of the proposed metric to accurately predict subjective image quality perception.