Skin Colour Detection Based On An Adaptive Multi-Thresholding Technique

Today, human region detection in complex scenes has received a great attention due to the wide use of websites and the considerable progress of the still and video images processing tasks. Skin detection or segmentation is a very popular and useful technique for detecting and tracking of human body parts, especially faces and hands. It is employed in tasks like face or hand detection and tracking, filtering of objectionable web images, people retrieval in databases and the Internet. This thesis aims to build a skin detection system that will discriminate between the skin and non-skin pixels in still coloured images. This is done by introducing a metric, which measures the distances of the pixel colour to skin tone. The need for a compact skin model representation stimulates the development of parametric skin distribution models which is used in this research.An adaptive skin colour detection model has been proposed in this thesis. The model is based on the bivariate normal distribution of the skin chromatic subspace. The model uses the 2D Single Gaussian model (SGM), and the 2D Gaussian mixture model (GMM) to represent the skin colour distribution. The model also based on the image segmentation using an automatic and adaptive multi-thresholding technique. This thesis shows that the Gaussian mixture model alone or the Gaussian single model does not improve the performance of the skin detection model due to the number of false detections for high correct classification. For this reason, a combination of SGM and GMM in the same model is proposed in this research. The results show that when processing images of different people taken in different imaging conditions, the use of only one single threshold value is not adapted, and since the proposed method is capable of adaptively adjusting its threshold values and effectively separating skin colour regions from non skin ones, it is applicable to images with various conditions. The experiment shows that the suggested algorithm achieves a noticeable performance improvement and offers a robust solution for skin detection under varying illumination. The results show that the average of the correct rate “True Positive” rate for the test images is equal to 94.064% while the False Positive average is equal to 13.166%.

Preview PDF FK_2007_60.pdf Download (466kB)

Actions (login required)

View Item