Research On Few-shot Facial Attributes Recognition Based On Meta-learning

In recent years,with the development of artificial intelligence and deep learning technology,facial attributes recognition has become one of the research hotspots in the field of computer vision.It has been widely concerned by researchers.Because it is difficult to collect facial attribute images with a large number of labels,how to learn the information about facial attributes from a single or several samples is an urgent problem to be solved.In order to solve this problem,the research of facial attributes recognition with few-shot is carried out in this thesis.The traditional methods of facial attributes recognition based on deep learning mostly rely on large-scale data sets containing complete label information.The traditional methods for large-scale data sets have some problems such as under fitting,poor generalization ability and poor recognition effect in few-shot data sets.At the same time,the quality of samples will also have some impact on the recognition effect.The idea of meta-learning "learning how to learn based on previous experience" has the characteristics of fast fitting,improving model generalization ability and few-shot learning.Therefore,this thesis uses the method based on meta-learning to study the problem of few-shot facial attributes recognition.At the same time,in order to improve the performance of the meta-learning method in the facial attributes recognition task of few-shot,this thesis also uses the face preprocessing method to preprocess the samples.The research content of this thesis mainly includes the following two aspects:(1)In order to reduce the negative impact of wild face images on few-shot facial attributes recognition tasks,a combination of Deep Alignment Network(DAN)and a priori embedded network structure GAN Prior Embedded Network(GPEN)is used to preprocess face images according to the characteristics of face images.First,use the DAN model to get the key points of the face image,then align and clip the face according to the key points,and finally use the GPEN model to enhance the face.(2)A model-independent meta-learning algorithm(CA-MAML)with cosine annealing learning rate optimization is proposed for few-shot facial attributes recognition.CA-MAML adds convolution layer to the basic network of Model-Agnostic Meta-Learning(MAML)and introduces Leaky Relu activation function to solve the problem of poor performance of model-independent meta-learning(MAML)in facial attributes recognition.At the same time,in order to make the CA-MAML algorithm more stable and generalized,the learning rate of cosine annealing is used to replace the fixed learning rate of the outer ring in the MAML algorithm,and the multi-step loss optimization is used in the inner loop.The experimental results show that the proposed CA-MAML meta-learning algorithm has good performance.In this thesis,a large number of experiments have been carried out to verify that the face image preprocessing method based on DAN and GPEN is effective for few-shot facial attributes recognition based on meta-learning.At the same time,the CA-MAML algorithm is verified to have good recognition effect and generalization in few-shot facial attributes recognition tasks.