Deep Multi-Task Multi-Label CNN for Effective Facial Attribute Classification

Facial Attribute Classification (FAC) has attracted increasing attention in computer vision and pattern recognition. However, state-of-the-art FAC methods perform face detection/alignment independently. The inherent dependencies between these tasks are not fully exploited. In addition, most predict all facial attributes using the same CNN network architecture, which ignores different learning complexities of attributes. To address above problems, we propose a novel deep multi-task multi-label CNN, termed DMM-CNN, for effective FAC. Specifically, DMM-CNN jointly optimizes two closely-related (i.e., landmark detection FAC) to improve performance by taking advantage learning. deal with diverse attributes, divide into groups: objective subjective Two architectures respectively designed extract features groups dynamic weighting scheme is proposed automatically assign loss weight each attribute during training. Furthermore, an adaptive thresholding strategy developed effectively alleviate problem class imbalance Experimental results on challenging CelebA LFWA datasets show superiority method compared several methods.