Generating Facial Line-drawing With Convolutional Neural Networks

The vigorous development of artificial intelligence has made robots move into our lives.Robots can not only meet the needs of industry,but can also be applied to all aspects of the art life.At present,robots can carry on the artistic creation of comics,sketches,paintings and so on.Line-drawings use points,lines and other simple elements to express the basic features of objects,which is more realistic,more vivid and more suitable for robots.As researches' study of Line-drawings are not being investigated so much,and there are no existing algorithms.We have learnt the algorithm advantages of making comics,sketches,paintings,and analysis an algorithm to generate facial line-drawings based on image processing and two algorithms based on convolution neural network.This paper present a accurate,fast and robust algorithm for making facial line-drawings.At the beginning of this paper,we use the image processing algorithm to make facial line-drawings.This algorithm extracts the hair and face positions based on the skin color detection method,analyzes the connected regions and obtains the feature domain which is closer to the real face.In this area,outline is extracted as the facial line-drawing.This method can accurately extract the outer edge of the face,but the details of the facial features can not be expressed accurately,and the facial contours of the human face are not clean and tidy,which can not generate a accurate facial line-drawings.Then,in order to get more neat and accurate line-drawings,we generate facial line-drawings based on convolution neural network,which is the most advanced technology of image processing.Combined with the characteristics of facial line-drawings,there are two requirements for the network: 1)End to end to generate line-drawings by training and predicting the whole image;2)No requirement for the size of input images.In this paper,an algorithm of Sketchy Facial Line-drawing Generation(SFLG)is proposed according to the requirements,the design method of network structure and the definition method of loss function are analyzed in detail.However,This method get a rough facial line-drawings with low accuracy.Next,In order to get more precise and meticulous facial line-drawings,we improve the SFLG network,and design Facial Line-drawing Generation(FLG).Improving the network pooling method,activation function;Improving loss function,adding fusion loss which is more suitable for FLG network;And improving the network structure,the new network fuse feature map of five stages at pixel level,and then having complete convolution,deconvolution,side output operation to output five side output maps,fusing those five maps at channel level and post processing to get the facial line-drawing.Finally,the paper makes comparison in the facial drawings results and algorithms of three methods,and analysis of the advantages and disadvantages.Experiments show that,facial drawings generation based on convolutional neural network,suits a wider range of different faces,has a certain universality;The algorithm is stable,line-drawings that generated by this method can accurately represent the face;The network uses complete convolution,deconvolution,not only can end to end make a image,but also has no no requirement about the size of input images.The FLG network is an accurate,robust and fast method for generating facial ling-drawings.