Research And Implementation Of Sketch Face Recognition Based On Deep Transfer Learning

Sketch face recognition technology plays an important role in the criminal investigation field and has become a research hot in the field of face recognition in recent years.The rapid development of deep learning has led to the maturity of optical face recognition technology.However,using deep learning to train a human face requires a large number of training samples.Due to the high cost of constructing a sketch face database,the samples of the sketch face database is small currently,it is impossible to train it by deep learning directly.In order to solve this problem,studied deep learning and transfer learning in this paper.The main idea is extracting optical face features from convolutional neural networks(CNN)and adapt it with sketch facial features by transfer learning to reduce sketch training samples.The main research work is as follows:Proposed adaptive scaled local binary pattern to extract optical face feature,the extracted features are classified by Gaussian process,the model is tested on the most authoritative optical face test set LFW,the accuracy is 98.7%.Finally,adapted the facial features which extracted by this method and the facial features which is extracted from the CNN respectively to the sketched faces by transfer learning,compared and analyzed the adaptation results.Trained the popular face recognition network models VGGFace,Caffe-face and Lightened CNN on the open source deep learning framework.The model's accuracy on LFW was: 97.41%,97.77%,99.03%.Adapted the optical face features extracted from each convolution layer of each net and sketched human face,compared and analyzed their results.Studied the rule of face features extracted from various layers of the CNN network.It is found that the layer is deeper,the extracted features is finer.Divided VGG16 into three parts,analyzed the extracted features from each part.According to the rule of CNN's feature,the optical face feature was extracted from the middle part of the VGGFace net,and it was matched with the sketch face by JDA,as a result of getting a complete VGG+JDA sketch face recognition model.Using face sketch datasets of Multimedia Experiments of the Chinese University of Hong Kong to test it,the test result is 97.4%,the result is compared with the test result of the traditional sketch face recognition method.