Measurement of image quality is of fundamental importance to numerous image and video processing applications. Objective image quality assessment (IQA) is a two-stage process comprising of the following: (a) extraction of important information and discarding the redundant one, (b) pooling the detected features using appropriate weights. These two stages are not easy to tackle due to the complex nature of the human visual system (HVS). In this paper, we first investigate image features based on two-dimensional (2D) mel-cepstrum for the purpose of IQA. It is shown that these features are effective since they can represent the structural information, which is crucial for IQA. Moreover, they are also beneficial in a reduced-reference scenario where only partial reference image information is used for quality assessment. We address the second issue by exploiting machine learning. In our opinion, the well established methodology of machine learning/pattern recognition has not been adequately used for IQA so far; we believe that it will be an effective tool for feature pooling since the required weights/parameters can be determined in a more convincing way via training with the ground truth obtained according to subjective scores. This helps to overcome the limitations of the existing pooling methods, which tend to be over simplistic and lack theoretical justification. Therefore, we propose a new metric by formulating IQA as a pattern recognition problem. Extensive experiments conducted using six publicly available image databases (totally 3211 images with diverse distortions) and one video database (with 78 video sequences) demonstrate the effectiveness and efficiency of the proposed metric, in comparison with seven relevant existing metrics.