Abstract: Aiming at the problem that the recognition rate of gender and age using facial features was low, we proposed a new method based on improved high-resoultion network (IHRNet). Firstly, we integrated the MobileNetV3 structure with a small number of parameters and a high recognition rate in IHRNet, and combined with the advantages of multi-scale feature extraction of the high-resolution network, which effectively improved the accuracy of facial feature recognition. Secondly, in order to reduce the risk of overfitting, the network first used the IMDB-WIKI face data set for pre-training, and then loaded the pre-training model for training and testing on the Adience face data set. Finally, compared with three similar algorithms such as ResNet50, HRNet and MobileNetV3. The experimental results show that the accuracy of IHRNet in age and gender recognition is as high as 82% and 95%, respectively, which are an average of 9% and 3% higher than similar algorithms, respectively, and the number of parameters is reduced by 36% compared with the previous one. The effectiveness of the improved algorithm is verified.

Key words: high-resoultion network, mobile network, gender recognition, age recognition, pre-training