Preprocessing with Skin Segmentation to Improve Monk Skin Tone (MST) Classification

Skin tones come in a diverse range of shades and are often necessary for various computer vision tasks. While skin detection is a well-studied focus, skin tone classification is not. Most works also use the Fitzpatrick Skin Type (FST) scale despite it being developed for skin phototype classification purposes. Due to the lack of a proper skin tone scale, the Monk Skin Tone (MST) was created, which was used to perform skin tone classification using Convolutional Neural Network (CNN) model as well as the Individual Topology Angle (ITA) value threshold method with skin segmentation as a preprocessing for both classification methods. To investigate if skin segmentation can improve classification accuracy, the MST classification was performed using both segmented and unsegmented images involving the IT A value threshold method as well as two CNN models that were trained using different image types. The experimental results showed that the best way to perform MST classification was using a CNN model trained and tested with segmented images, which yielded the highest accuracy score being 0.91, with no signs of over or underfitting. The results also supported that segmentation is necessary, and that it can improve a classification model's accuracy.