Hairstyle pattern recognition based on CNNs
Sun Chao, Won‐Sook Lee · 2017
Hairstyle recognition is a challenging task since hairstyles span a diverse range of appearances in real-world. However, it is possible to start from recognizing the most basic hairstyles then dealing with more complex hairstyles. In this paper, we present a novel hairstyle pattern recognition system based on CNNs. We first give the definitions of four basic hairstyles: straight hairstyle, curly hairstyle, kinky hairstyle, and braid hairstyle. Then We leverage the power of the pre-trained CNN model to learn the distinct features of those basic hair patterns from representative hairstyle patch dataset. The CNN model can classify different hairstyles from patches and perform hairstyle recognition in full hair images. Patch-based recognition makes our system very flexible to be applied to hair images that captured from multi-views. The experiment results show that our system can perform recognition for both simple hairstyle images and complex hairstyle images.