Spatial Coherency in Colourisation

Automatic colourisation is the function of inferring colour information from a grey-scale prior and then combining the colour with the grey-scale to form a colourised version of the image. We identify Spatial Coherence as a particular weakness in methods that use Convolutional Neural Networks for colourisation. Generated colours do not adhere to semantic edges and are not consistent within boundaries where we would expect uniform colour. Spatial Coherence, while often evident to the human eye, does not yet have an objective metric. We show, by segmentation of the combined ab channels of the CIEL*a*b* colour space, that a segmentation based on CNN colourisation is poor. We argue the need for the development of metrics to evaluate a colourisation’s performance on Spatial Coherence.