Face Age Recognition In Social Network

As one of the attributes of human faces,age has a very wide range of applications,especially for the age estimation of faces in social networks,which has great potential commercial value.Face images in social networking are more changeable than those in laboratory environment.They often have many characteristics such as changeable individual posture,serious occlusion,large illumination change and uneven image quality.Traditional feature description is difficult to express well for images in such a non-restricted environment,while deep learning method needs lots of storage and computing resources and is not easy to land.In this paper,we will explore how to use depth learning to estimate face age in social networks more efficiently.Our works explore efficient deployment of deep learning model in two aspects.First,it changes the activating function,deepens network layers as well as introduces short-circuit path to overcome the weakness of lightweight network.Comparing with the networks used by others,the improved network structure has a significant advantage on model size and computational complexity,and receives good results on the public age-evaluating set.Second,this paper introduces quantitative method to further compress trained models,which gains a balance between model efficiency and model accuracy more flexibly.The quantitative method fully exploits the expressive redundancy of general float number and limitation of model parameter range and uses less bits to represent trained model parameters.A more efficient model storage method is obtained on the premise that the accuracy is sacrificed as little as possible.This paper compares the experimental results under different quantitative bits,the size of the model is effectively compressed on the premise that the accuracy of the model is almost no loss,which proves the practicality of the quantitative method.