Multi-Output Convolutional Neural Network for Automatic Human Head Attributes Classification

Abstract

Human head attribute classification (HHAC) is a fundamental and substantial research domain in pattern recognition and computer vision. However, recent HHAC networks do not consider the correlation of common characteristics among the attributes over different regions of the human head. To address the above problem, this study proposes a multi-output convolutional neural network to jointly learn the features of human head attributes with common characteristics. The proposed network contains two convolutional blocks and five output layers, where each output layer learns to predict a specific group of human head attributes. In order to properly learn the correlation among the human head attributes, this study divides these attributes into five groups: hair, face, style, accessories, and appearance. Extensive experiments showed that the proposed network obtained an average classification accuracy of 95.29% and 97.93% on the challenging CelebA and LFWA datasets, respectively. Thus, the proposed network is approximately 2% and 10% superior to the closest competitor (i.e., PS-MCNN) on both datasets. In addition, the proposed network achieved higher classification accuracy compared to the existing networks almost in all human head attributes. That findings demonstrate the effectiveness of the proposed network and the attributes grouping method in learning the correlations among human head attributes correctly.