Training Bayesian classifier with scaling unique colors among image samples

摘要

To improve the quality of skin segmentation in images, the new training approach of the Bayesian classifier has been proposed. The segmentation problems arise in many applications of developing the intellectual systems based on video analysis. One of the important tasks is the human skin segmentation for further high-level processing. The most used approach is based on naive Bayesian classifier to separate human skin (foreground) pixels from the background ones. We propose the new approach for preparing training dataset which considers only unique colors from each image to avoid normalization influence related to the size of the foreground area. It provides stable improvement of the segmentation quality based on ROC-curves comparison. The testing was conducted on the union of 4 biggest public datasets containing labeled images with human skin. The results show that the average improvement about 3-4% of TPR and up to 10% of FPR metrics.