基于深度学习的用户头像多标签分类

Multi-label Classification of User Avatars Based on Deep Learning

在众多的社交图像数据中,用户头像能直观地代表一个用户,是用户个性和偏好的集中体现.如何快速准确地提取用户头像信息,对用户群体进行筛选和分类,成为一个重要的研究问题.在收集到的自定义用户头像数据集上分别用调整后的ResNet50V2、InceptionV3和DenseNet121对其进行多标签分类实验,对比其性能,发现各网络收敛速度较慢,学习指标较低;为提高模型表现,融合3个网络的特征层,构造集成网络并将其训练.测试结果表明,集成网络在精准度P、召回率R和F0.5上相较于子网络均有显著提高.

Among numerous social image data,a user avatar can intuitively represent a user,which is a concentrated expression of the user’s personality and preferences.How to quickly and accurately extract user avatar information,filter and classify user groups,has become an important research problem.We used the adjusted ResNet50V2,InceptionV3 and DenseNet121 to perform multi-label classification experiments on the collected custom user avatar dataset,and compared their performance.We found that the convergence speed of each network was slow and the learning index was low;in order to improve the performance of the model,we fused the feature layers of the three networks,constructed an integrated network and trained it.The test results show that the integrated network has significant improvement in accuracy P,recall rate R and F0.5 compared with the sub-networks.