Research on Feature Extraction and Transfer Learning of Convolutional Neural Network to the Style of Li Brocade

Boxiong Yang1,2*, Heng Xiao1,2, Tao Yang1,2, Ping Yin1,2

1School of Information & Intelligence Engineering, University of Sanya, Sanya, 572022, China

2Academician Workstation of Chen Guoliang, University of Sanya, Sanya, 572022, China

Abstract: Li Brocade is exquisitely made, brightly colored, and beautifully decorated, which embodies the national characteristics of the Li nationality in spinning, weaving, dyeing and embroidery. Every pattern and picture of Li Brocade has its own style. In deep learning, a convolutional neural network (CNN ｏｒ Conv Net) is a class of deep neural networks that are most commonly applied to analyzing visual imagery. VGGNet is a convolutional neural network developed by researchers from the Visual Geometry Group of Oxford University and Google Deep Mind. Both content and style can be extracted as features through the VGG-19 Deep Neural Network Model. VGG-19 is used to extract the pattern style of Li Brocade, and then the features of its style transferred to other pictures according to transfer learning. The experimental results show that the styles of Li Brocade are well extracted, migrated, and preserved. If more style features of the Li Brocade pattern are needed to be dug out, the parameter value of style layers should be set as large as possible.

Keywords: Deep learning; Feature extraction; Transfer learning; CNN; Style of Li Brocade