Hair counting with deep learning
Alessio Gallucci, Dmitry Znamenskiy, Nicola Pezzotti, Milan Petković
Abstract
We present a set of deep learning models aimed at solving the hair counting problem in human skin images. All the models are end-to-end, providing a mapping from the input image to a single scalar corresponding to the number of hair. The list of models corresponds to the most common deep learning architectures that worked over-time in various applications, where some of the networks were adapted to output the hair count. Results show that autoencoder architectures with skip connections work best for such end-to-end counting task, hinting at increased performance when multi-task learning is used. With the results presented, we speculate on the possibility to remove human annotator from the tedious task of manual counting of skin hair.