Research On Heterogeneous Images Based Face Recognition

Face recognition,as one of the most simple and accurate human biometric recognition technologies,has been widely used in our daily life.At present,the main challenges of face recognition include pose,expression,illumination,occlusion,age and so on.However,when recognizing heterogeneous face images,the accuracy of human is much higher than that of machine.So,this paper focus on face recognition based on heterogeneous face images(sketches,caricatures).The main contents can be divided into the following three aspects:1)we propose a representation based on facial components for caricature face recognition.The fusion representation of the photo,caricature and generated face images are selected in the caricature face recognition.The generated face is to extract and reconstitute the facial components(eyes,nose and mouth)in caricature.We adopt ResNet for extracting the representations and Softmax loss and CenterLoss as the loss function.Restricted by the scarcity of caricature databases,we build a caricature dataset which includes a total of 14633 pictures of 259 people and each image is annotated.The methods we proposed achieves 70.84% accuracy on the caricature dataset we proposed experimental,which is the best performance.2)we propose the method of generating heterogeneous face images which is based on Cycle Generative Adversarial Networks.We generate sketch and caricature separately with the usage of Cycle Generative Adversarial Networks.The training data we used is the most common sketch and caricature datasets.At last,the generated sketch can retain more detailed information and the generated caricatures are exaggerated on some degree,while some of the caricature are distorted.3)we propose a representation based on face and heterogeneous faces for face recognition.Before recognition,face frontalization is performed on all the images with huge pose.After frontalization,all images in VGGFace2 dataset are generated into sketches and caricatures.Then,four experiments are done which includes photo,photo and sketch,photo and caricature,photo,sketch and caricature.To verify the robustness of the model,three trained model are used for face verification on LFW face dataset.The experimental results show that the features which is fused with heterogeneous face images are more discriminative and robust.