Research On Facial Attributes Recognition Based On Convolutional Neural Networks

With the rapid development of social information,there are a large number of facial images emerging on the Internet.Human face conveys many kinds of important information,such as gender,age,expression,ethnicity,etc.Facial attributes has widely application prospects in the fields of human-computer interaction,business recommendation,security systems,etc.Therefore,facial attributes recognition is of great significance and practical value,which is one of the research hotspots in computer vision.The method on facial attributes recognition is mainly divided into traditional method and deep learning method.Compared with the traditional method,facial attributes recognition based on deep learning can automatically learn facial representation from massive data,which can obtain higher layer semantic information and improve the generalization performance of the model.In recent years,researchers have proposed a large number of methods,which have significantly improved the accuracy of facial attributes recognition.However,many difficulties and challenges need to be solved.On the one hand,facial images are influenced by factors such as lighting conditions,occlusion,and pose variations,which interfere with feature extraction.On the other hand,facial attributes datasets are limited.In special,facial images with accurate age,expression,and ethnicity are scarce.In addition,there is an extreme class imbalance existing in dataset as the difficulty of data collection.As one of the difficulties on facial attributes recognition,the accuracy of age estimation needs to be further improved.Therefore,this paper mainly analyzes the method of age estimation and proposes a specific solution for the current difficulties.This paper mainly discusses some methods based on convolutional neural networks for age classification,and proposes a convolutional neural networks model based on the multi-class focal loss function to address the class imbalance of dataset.This method is based on focal loss.It can be extended to address the multi-classification task and overcome the impact of category imbalances for age classification.Specifically,our approach is designed to address the class imbalance via reshaping the standard cross entropy loss that it down-weights the loss assigned to well-classified examples,and focus on training hard examples to further improve the accuracy of age classification.Furthermore,to compare the effects of different depth models on age classification performance,we perform extensive experiments via using different depth convolutional neural networks with multi-class focal loss.Meanwhile,we select the appropriate normalization method and the value of dropout.Finally,we validate our method on well-known Adience benchmark.The experimental results demonstrate that the method based on multi-class Focal Loss can effectively address class imbalance and further improve the accuracy of age classification.In order to alleviate the limited problem of dataset,this paper is based on multi-task learning for gender and age classification.We conduct experiments with MORPH dataset for training and testing.The experimental results show that multi-tasking learning can effectively reduce training redundancy,and improve the performance of a single attribute.