Multi-task Attribute Joint Feature Learning

摘要

Recognizing face attributes can improve face recognition as well as provides useful information in face image retrieval. Usually the attributes are studied separately. Considering that the attributes are inter-related, they can be regarded as sharing common data structure. In this paper, we propose to take advantage of Multi-task learning (MTL) framework to learn attribute feature simultaneously. Specifically, the attributes are divided into several tasks. The attribute feature information can be better shared across the tasks with MTL. According to the value of weight vectors of all features learnt by MTL, we can select much lower number of feature dimension for attribute recognition without losing the prediction precision. The experiments are conducted on LFW database with nine face attributes from three tasks to verify our method. The experiment results compared with Single Task Learning (STL) show the effectiveness of the proposed method.