Computationally efficient algorithms for predicting the file size of JPEG images subject to changes of quality factor and scaling

To enable the delivery of multimedia content to mobile devices with limited capabilities, high volume transcoding servers must rely on efficient adaptation algorithms. Our objective in addressing the case of JPEG image adaptation was to find computationally efficient algorithms to accurately predict the compressed file size of images subject to simultaneous changes in quality factor (QF) and resolution. In this paper, we present two new prediction algorithms which use only information readily available from the file header. The first algorithm, QF Scaling-Aware Prediction, predicts file size based on the QF of the original picture, as well as a target QF and scaling. The second algorithm, Clustered QF Scaling-Aware Prediction, also takes into account the resolution of the original picture for improved prediction accuracy. As both algorithms rely on machine-learning strategies, a large corpus of representative JPEG images was assembled. We show that both prediction algorithms lead to acceptably small relative prediction errors in adaptation scenarios of interest.