| With the increase of image data on the Internet and the development of deep learning,the interpretation of the information contained in network images has become important.Face is one of the main appearance characteristics to distinguish each person.Gender recognition and age estimation can also be performed through the face.Face recognition has important application value in identity verification,human-computer interaction,and personalized product recommendation.Therefore,how to efficiently detect human faces from network images and accurately identify their gender and age has become an important research topic.Aiming at the problems of slow running speed and low accuracy in face detection,gender recognition and age estimation by the Centernet algorithm and single-feature weak classifier algorithm in the network avatar recognition task,1.This paper proposes an improved Centernet face The detection algorithm first uses the central area of the Gaussian heat map instead of the central point to increase the smoothness of the regression,introduces the NMS method to eliminate the redundant target frame,and then handles the data imbalance problem through focal loss,which effectively improves the detection speed and detection rate.2.Aiming at the classification method of traditional weak classifier using single feature,this paper proposes an improved multi-scale multi-feature fusion strong classification facial gender and age recognition algorithm.First,local feature extraction at different scales is performed on the face image And global feature extraction,and weighted fusion with the label text feature of the image,and finally use the multi-core enhanced classifier to perform gender recognition and age estimation on the features,which improves the accuracy of recognition and estimation.The paper selects IMDB-WIKI database and our custom dataset of 10,000 Weibo avatars containing human faces as the data source.Experimental verification results show that on the same test data set: 1.The improved Centernet algorithm proposed in this paper has improved ROC indicators and detection speed compared with other similar algorithms.2.The multi-scale and multi-feature fusion facial gender and age recognition algorithm proposed in this paper is more prominent than other similar algorithms in terms of recognition accuracy and 1-off accuracy in gender recognition and age estimation. |
