Face Recognition Algorithm Research Based On Sketch Image

Sketch based face recognition aims to determine the person's identity by retrieving in the photo database using simulated sketches automatically.It has important application in identification of the criminals and caricature-based image search.How to match sketch and photo image accurately becomes the key of this research.Firstly,in order to solve the problem of insufficient samples,a deep transfer learning method is adopted to recognize sketches and photos directly on the pre-training model.Secondly,in order to further reduce the modal difference,this paper adds the step of image synthesis before feature extraction.In addition,aiming at the more abstract cartoon sketch face image,this paper uses the method of ensemble learning to establish multiple weak classifiers on the extracted bottom features and facia semantic features,and finally realizes the differentiated selection and integration.The main work of this paper is as follows:(1)A face sketch-photo recognition method based on deep representation transfer learning is proposed.Firstly,the database containing a large number of face photos is used for pre-training in the deep convolution neural network system and the parameters of the network are obtained and migrated to the network for sketch image recognition.Then,the triplet image composed of the sketch image database is used as the input of the pre-training model,and train the model to minimize the defined loss function,so as to achieve the purpose of imposing restrictions on inter-class and intra-class distances.The experimental results show that the effective transfer learning strategy can improve the accuracy of face recognition in sketch images by 10%.(2)A face sketch-photo recognition method based on conditional generative adversarial network is proposed.Firstly,the paired sketches and photos of the horizontal stitching are input into the conditional generative adversarial network,and the more realistic natural images are obtained by training the model.Then,the generated natural image and the natural face photo in the database are input into the deep convolutional neural network to extract the depth feature,and then the similarity measure is performed on the feature to output the discrimination result.The experimental results show that the conditional generative adversarial network can generate realistic photos of the sketch.And adding the step of image synthesis before sketching the face matching,the face recognition rate is increased by 11%.(3)A face sketch-photo recognition method based on the combination of multi-layer features and multi-classifier ensemble is proposed for more exaggerated caricature sketch.In addition to using the traditional bottom features,the algorithm also uses facial semantic features that are not affected by modality.By combining the extracted multi-layer features,multiple weak classifiers are respectively established to obtain multiple classification results.Finally,the face matching recognition of caricature sketch is completed by voting method.The experimental results show that the facial semantic features play a very important role in caricature sketch recognition,and compared with the traditional bottom features,the recognition rate is increased by about 12%.The ensemble classifier also effectively improves the accuracy of face recognition.Using a variety of classifiers is better than using only a single classifier,which can basically improve by about 5%.