Cross-Modal Face Recognition Based On Deep Learning

Cross-modal face recognition is a research hotspot in the field of face recognition,and sketch face cross-modal recognition has important significance for real-world security and entertainment applications.Sketch cross-modal face recognition has been extensively studied in the early days.In general,the sketch cross-modal face recognition task is mainly two parts:synthesis/mapping/features extraction and face recognition.Recently,deep learning has already performed well on face recognition,and research on image synthesis problems with deep learning has increased.However,little work focuses on deep learning methods in sketch cross-modal face recognition.In addition,the existing sketch face datasets are small,so the solution of training the deep learning model using small data is not optimal.In order to apply deep learning better to sketch cross-modal face recognition tasks,this paper explores the effects of existing image synthesis and face recognition deep learning models on this task,and the basic solution is constructed by combining the existing images synthesis deep learning network and the deep face recognition model.Secondly,This paper proposes to embed the face recognition pre-training model in the image synthesis network to use the face recognition experience of the big data in the synthesis task to guide the synthesis of more recognizable images.In addition,in order to extract better deep learning features for sketch face images to improve recognition rate,and considering the problem of small number of sketch datasets,this paper proposes to introduce external big datasets to use existing sketch synthesis models for data augmentation.Then a sketch face recognition deep learning model is obtained,and the metric learning of triplet loss is used to adapt the data distribution of the training dataset.Finally,the proposed method achieves 100%and 98%recognition rates on the CUHK and CUSFS datasets,respectively.