Direction Based Graphical Model For Face Sketch Synthesis

With the constantly development of the science technology and the improvement of the extent of information,people's demand for intelligence life has grown rapidly.The face recognition technology has been widely studied and applied.As a branch of face recognition technology,face-sketch synthesis also has gained growing attention.In public security area,due to the certain constraint of the environment or other conditions,police may not be able to acquire suspect's high resolution photos.The feasible solution is to ask professional artists to draw the facial sketch of the suspect according to the description of the witnesses as well as the help of the poor quality security camera videos.Then we can transform the photos from the police system face database to sketches,and use the sketch of the suspect for retrieval and recognition,then further confirm his identity.In digital entertainment,as people's desire for pursuing character and self-expressing becomes stronger,using stylistic pictures as personalized appearances has become a trend,which makes sketch one of their best choices.Thus,studying face sketch synthesis is of great significance.In this thesis,the main research object is data-driven face sketch synthesis.By analyzing and comparing the existing face sketch synthesis methods,we make innovations to the methods,aiming at the drawbacks of the current approaches.The main work of this thesis can be outlined as follows:1.We proposed a directional graph model based face sketch synthesis method.Most of the current face sketch synthesis methods have neglected the direction information of the sketch texture.They only use Euclidean distance to measure the similarity between different image patches,which would cause the inconsistency of the texture direction between neighbor patches.Aiming at the aforementioned problem,we proposed a directional graph models based face sketch synthesis method: First,we use Gabor function to extract the direction information of the image patch.Then we synthesize face sketch via combining the image direction information with image texture through Markov network.Experimental results demonstrate that our method can synthesize details like hair very well,which has greatly improved the subjective visual effect of the synthesized face sketch.2.We realized multi-view transformation for face sketch.This method aims at certain circumstances in law enforcement that the surveillance video can only acquire side faces of the suspect,or the witnesses only spot local facial feature of the suspect.If we use the side face image directly to recognize or match in the police face database,the recognition accuracy would drop dramatically,and there may be even unrecognized situation.The best strategy is to use the frontal images from the current face database to complete multi-view transformation.The series of virtual multi-view face images can make up a more comprehensive database,which could further improve the accuracy of face recognition.As for the application in digital entertainment,the multi-view transformation for face sketch can provide users with more choices to meet their need for pursuing personal character.Based on the proposed directional graph model based face sketch synthesis method,we make a further step and realize the transformation of multi-view face.The experimental results show that our method can realize multi-view transformation from frontal face sketch with simple and fast properties,which would be conductive to the promotion in practical applications.