Facial expression recognition (FER) is an area of active research, both in computer science and in behavioural science. Across these domains there is evidence to suggest that humans and machines find it easier to recognise certain emotions, for example happiness, in comparison to others. Recent behavioural studies have explored human perceptions of emotion further, by evaluating the relative contribution of features in the face when evaluating human sensitivity to emotion. It has been identified that certain facial regions have more salient features for certain expressions of emotion, especially when emotions are subtle in nature. For example, it is easier to detect fearful expressions when the eyes are expressive. Using this observation as a starting point for analysis, we similarly examine the effectiveness with which knowledge of facial feature saliency may be integrated into current approaches to automated FER. Specifically, we compare and evaluate the accuracy of ‘full-face’ versus upper and lower facial area convolutional neural network (CNN) modelling for emotion recognition in static images, and propose a human centric CNN hierarchy which uses regional image inputs to leverage current understanding of how humans recognise emotions across the face. Evaluations using the CK+ dataset demonstrate that our hierarchy can enhance classification accuracy in comparison to individual CNN architectures, achieving overall true positive classification in 93.3% of cases.