Learning colours from textures by sparse manifold embedding

The capability of inferring colours from the texture (grayscale contents) of an image is useful in many application areas, when the imaging device/environment is limited. Traditional manual or limited automatic colour assignment involves intensive human effort. In this paper, we have developed a user-friendly colourisation technique, where the algorithm learns the relation between textures and colours in a user-provided example image and applies the relation to predict the colours in the target image.The key contribution of the proposed technique is trifold. First, we have explicitly built a linear model for the texture–colour relation. Second, we have considered the global non-linear structure of the data distribution by applying the linear model locally; and the local area is determined automatically by sparsity constraints. Third, we have introduced semantic information to further improve the colourisation. Examples demonstrate the effectiveness of the proposed techniques. Moreover, we have conducted a subjective study, where user experience supports the superiority of our method over existing techniques.

► To learn colours from examples, we build a linear model of the texture–colour relation.
► We have considered the global non-linear structure of the data distribution by locally applying the linear model.
► The local area is determined automatically by sparsity constraints.
► We have introduced semantic information to further improve the colourisation.
► We have conducted subjective study to assess the colourisation.