Hair Recognition Based on Multi-task Convolution Network

Abstract

Aiming at the problem that the prior art is difficult to utilize the mutual relationship between labels to achieve efficient classification of hair style multi-labels, a hair styling classification method based on multi-task convolutional neural network is proposed., a multi-task joint learning model was constructed to try to realize the simultaneous recognition of hair shape and color. First, all the labels are jointly learned through the shared network layer, and the correlation between labels is automatically mined and utilized from the perspective of feature extraction. Then complete specific categories of learning tasks at different sub-network layers, thereby eliminating ambiguity in multi-label classifications. Finally, multiple classifiers are trained to achieve parallel prediction of all labels. The research shows that the proposed method can extract multiple features of the hair at the same time and directly classify and identify it. Compared with the single task network, the precision has obvious advantages.