Stacked Ensemble of Convolutional Neural Networks for Follicles Detection on Scalp Images

An average person's head is covered with up to 100000 hairs growing out of follicular openings on the scalp's skin. Automated hair therapy requires precise detection and localization of follicles on the scalp and still poses a significant challenge for the computer vision and pattern recognition systems. We have proposed an automated vision system for follicles detection based on the classification of digitized microscopic scalp images using an ensemble of convolutional neural networks (CNN). A pool of adapted state-of-the-art CNNs have been transfer-trained on over 700k microscopic skin image regions of 120x120 pixels and their outputs further fed to the final stacked ensemble learning layer to capture a wider context of the connected neighboring regions of the original FullHD scalp images. A high validated f1 score (0.7) of detecting regions with follicles beats the industry's benchmark and brings this technology a step closer towards automated hair treatment as well as other emerging applications such as personal identification based on follicular scalp map.