Face Sketch Synthesis Via Multivariate Output Regression

This paper presents a multivariate output regression based method to synthesize face sketches from photos. The training photos and sketches are divided into small image patches. For each pairs of photo patch and its corresponding sketch patch in training data, a local regression model is built by multivariate output regression methods such as kernel ridge regression and relevance vector machine (RVM). Compared with commonly used single-output regression, multivariate output regression can enforce the synthesized sketch patches with structure constraints. Experiments are given to show the validity and effectiveness of the approach.