Line fitting based segmentation and its application to skin image analysis

As mobile devices are recently equipped with advanced image sensors, it is expected that they would make it possible to diagnose one's skin condition conveniently anywhere and anytime. For this application it is necessary to classify the skin features from a skin image accurately. Although two conventional thresholding methods Otsu and Kapur perform well in a range of applications, we found that they are inappropriate for the specific skin analysis application of our interest. Otsu poorly works for non-bi-modal distributions while Kapur shows poor performance for long-tailed distributions. So, we propose a new line-fitting based thresholding method to detect pigments and pores from a skin image where the image histogram is simply modeled by two lines. Experimental results show that the proposed method finds better threshold which leads to the most visually plausible detection, compared to Otsu and Kapur.