Research On Gender Recognition And Age Estimation Based On Multi-layer Feature Fusion

With the progress of science and technology,gender recognition and age estimation technology has become one of the hot technologies to solve this problem with many advantages such as non-contact authentication.Although there are many researchers,it is still a difficult problem.In the existing technology,how to extract rich and effective texture features,how to reduce resource waste,how to avoid parameter disasters and how to make full use of the extracted features To improve the accuracy of image classification,it is a problem that needs to be solved at present.Convolutional network-based recognition algorithms usually use only the output of the last layer as feature representation,but the information in this layer may be too rough in space,while the features acquired by earlier layers may be precise in space but do not contain semantics.In order to get richer hierarchical features,two methods of multi-layer feature fusion based on convolutional neural network are studied for gender recognition and age estimation tasks.The final results of these two tasks are verified by using Group data set,IMDB-WIKI data set and Morph data set.The specific work of this paper is as follows:(1)For gender recognition and age classification,gender recognition and age classification are regarded as classification problems.First,a simple framework is constructed to extract abundant spatial structure information using cascaded convolution kernels of 5×5,3×3 and 2×2,respectively.Then,different convolution layers of different numbers and sizes are connected to form different streams,so that the network framework can obtain high-level semantic information as well as obtain high-level semantic information.The low-level edge texture information is integrated with different streams to obtain different levels of deep convolution activation features to classify images by gender and age.In addition,when choosing the age loss function,considering the influence of gender on age,the accuracy of age classification can be improved by reducing the differences within the same gender.(2)Aiming at the problem of gender recognition and age estimation,the compact Soft Stagewise Regression Network(SSR-Net)framework is improved to make the features acquired by the network have more semantic information: residual module is introduced for SSR module in SSR-Net framework,which can get more advanced semantic information and alleviate the model regression caused by too many network layers.In order to improve the performance of the network,adaptive weights are introduced into the prediction module,and the enhancement of advanced features is deepened.The experimental results show that the accuracy of the classification framework based on multi-level feature fusion is higher than that of VGG16,and the average absolute error of RSSR-Net with residual and adaptive modules is 7.4357 on WIKI and Morph datasets,which is 0.36 lower than that of 7.7916 before improvement.