Research On Sketch Face Synthesis

Face sketch images are not only artistic portraits in life,but it can also become an important clue in criminal investigation to help the police find relevant personnel.With the development and progress of technology,face recognition technology has become more and more mature and widely used in various fields.However,due to the huge difference between sketch images and photo images,it is generally difficult to achieve satisfactory results by directly taking the sketch face image as the probe to retrieve the police mug-shot database.However,the face sketch synthesis technology can make the probe image and the library image be in the same image domain and help to obtain a higher recognition rate than they are in different image domains.In addition,by embedding the face sketch synthesis algorithm into the mobile phone or personal computer,it can help people to obtain sketch-style face images corresponding to photos face images faster,cheaper and more conveniently,which can enrich people's entertainment and life.In order to synthesize face sketch images that are both helpful for face recognition and have good visual effects,this paper mainly proposes the following three methods for sketch face synthesis:(1)Two-stage face sketch synthesis method based on FCN preprocessing.In this method,we divide the sketch face synthesis process into a preprocessing stage and a synthesis stage.In the preprocessing stage,an eight-layer fully convolutional network is trained to convert the photo images in the training dataset and test dataset into semi-sketch images.In the synthesis stage,the existing exemplar-based method is used to synthesize the final sketch images,but it takes the semi-sketches rather than photo images as input.Various experiments are conducted to verify the effectiveness of the proposed method in improving the sketch synthesis quality of the exemplar-based method.In addition,the experiments on cross-dataset indicate that the proposed method provides a new means for strengthening the generalization ability of the exemplar-based method.(2)Face sketch synthesis based on the generative adversarial network.In this method,a multi-adversarial network consisting of a generative network and three discriminant networks is used to complete the sketch face synthesis task.In the progress of training the generative network,we use the adversarial loss,the content loss in the feature space,the content loss in the pixel space,the style loss and the total variation loss to train the model.We design control experiments to verify the effectiveness of the multi-level adversarial structure and the style loss function.And for further development,the quality of synthesized results by a light generative network is discussed.Compared with the state-of-the-art face sketch synthesis algorithms,the results synthesized by the proposed method have better visual effects and higher accuracy of face recognition.(3)Dual-domain face sketch synthesis based on style transfer.A sketch-style image which has similar content with the test photo image is obtained by using the style transfer method.Just as the training sketch image is the counterpart of the training photo image in the sketch image domain,we take the sketch-style image as the counterpart of the test photo image in the sketch image domain.Then,the selection of neighbor patches and the calculation of reconstruction weight are performed in two image domains,the sketch domain and the photo domain,to fully mine the information in the photo images and sketch images in the training set to produce better results.Experiments on the public dataset demonstrate the proposed method can synthesize better results in terms of visual perception,objective image quality assessment and recognition accuracy.