Detecting Faces in Colored Images Using Multi-skin Color Models and Neural Network with Texture Analysis

This paper presents an efficient hybrid system to detect frontal faces in colored images regardless of scale, location, illumination, race, number of faces, and complex background. The general architecture of the proposed system encompasses three methods: skin color segmentation, rule-based geometric knowledge, and neural network-based classifier. In the proposed system, multi-skin color clustering models are applied to segment human skin regions, iterative merge stage creates a set of candidate face regions, then, the facial feature segmentation removes false alarms, caused by objects with the color that is similar to skin color. Furthermore, the rule-based geometric knowledge is employed to describe the human face in order to estimate the location of the “face center”. Then the Artificial Neural Network (ANN) face detector is applied only to the regions of the image, which are marked as candidate face regions. The ANN face detector must decide whether a given sub-window of an image contains a face or not. Partial face template is used, instead of the whole face, to reduce face variability, in order to make the training phase easier and to reduce misrecognition. To increase the accuracy of the system, twelve-texture descriptors are calculated and then attached as input data with each face image to train neural network to describe the content of sub-image window such as X-Y-Relieves, smoothness, ratio of darkness, etc. Training neural network is designed to be general with minimum customization. Comparisons with other face detection systems have revealed that our system shows better performance in terms of positive and negative detection rates.