Optimal feature space for semantic image segmentation
Sergey Anishchenko, Mikhail V. Petrushan · Pattern Recognition and Image Analysis · 2014
A new method for semantic segmentation with object adaptation is proposed. Segmentation is performed on a feature map obtained as a weighted sum of HSV-space components: hue, saturation, and- value. Homogeneity criterion for grouping pixels into clusters and weighted sum coefficients is adjusted using the particle swarm optimization (PSO) algorithm. The method is tested using images wherein a face is a semantically meaningful object. The accuracy of the segmentation is shown to be higher when using a feature map obtained as a weighted sum of color components than in cases when a single color component is used. The accuracy of the proposed method is estimated as the correspondence of a fragmented segment to a face region, detected using the Viola-Jones method, within an image.