Novel color space representation extracted by NMF to segment a color image

This paper considers the task of separating pixels in color image into background and foreground classes. Using the machine learning technique known as Nonnegative Matrix Factorization, data pertaining to different color channels – selected by color spaces – are combined, and a novel space representation is extracted. The novel representation of the image includes additional information, namely “metacolor”, which could be related to foreground and background and adopted to improve binary segmentation of the investigated image. In both qualitative and quantitative experiments, the use of novel color space representation produces some improvements in the binary segmentation results when it compared to those obtained applying common simpler thresholding algorithms directly to the original image.