Face Sketch Synthesis And Augmentation

Since human face can provide considerable information about one's traits and can convey details of his/her characteristic to the others. Thus, in recent years, more and more computer scientists have begun to do research about the human face related tasks for its widely practical applications. In this paper, we complete a very interesting project that how to generate a face sketch from a face photo.The approach used here to generate a face sketch from a face photo can be described as patch-based method. In the patch-based method, the whole process can be divided into three steps. Firstly, the face photo is divided into a series of face photo sketches. Then, for every face photo path, a face sketch is generated. At last, the generated face sketches are stitched together to be a whole face sketch.In the process of generating a face sketch patch from a face photo patch, the Markov Random Field model is built and the algorithm, which is called Loop Belief Propagation, used to solve this model is described. In order to illuminate how this algorithm works, some preliminary knowledge is detailed first, which includes the basic information of undirected graph model and the factor graph.Then, some different stitching methods are described, which are averaging pixel stitching method, minimum error cut stitching method and Multiresolution Spline algorithm. The advantages and disadvantages of these stitching methods are analyzed. And a new blend method which is based on Multiresolution Spline algorithm and a blend trick which is called full-coverage blend trick are designed to stitch the generated face sketch patches into a whole face sketch. The details of Multiresolution Spline algorithm, which include Gaussian pyramid and Laplacian pyramid, are also given in Character three for the sake of a better description of the novel blend method and the full-cover blend trick.In addition, we discuss the role of the size of the divided patches in the whole patches-based method. After analyzing, it can be proved that the smaller the patches are, the more mottled the generated face sketch will be. But the detail information of one face sketch, such as the special texture or shadow information, can only be learned by using small patches. In addition, in order to make the stitched whole face sketch be smooth, the size of the patch should be large. However, if the size of the patch is so big, the blended face sketch will be blurry. Thus, the size of the divided patches should neither be too big or too small. In this research, we propose Non-negative Matrix Factorization retraining to solve this problem and the larger patches are used in this step.At last, we generalize proposed methods, which are Markov Random Field and Non-negative Matrix Factorization retraining, to global optimization methods. Also, we analyze how they are connected.