Associative memory based on ratio learning for real time skin color detection

Abstract

A novel approach for skin color modeling using ratio rule learning algorithm is proposed in this paper. The learning algorithm is applied to a real time skin color detection application. The neural network learn, based on the degree of similarity between the relative magnitudes of the output of each neuron with respect to that of all other neurons. The activation/threshold function of the network is determined by the statistical characteristic of the input patterns. Theoretical analysis has shown that the network is able to learn and recall the trained patterns without much problem. It is shown mathematically that the network system is stable and converges in all circumstances for the trained patterns. The network utilizes the ratio-learning algorithm for modeling the characteristic of skin color in the RGB space as a linear attractor. The skin color will converge to a line of attraction. The new technique is applied to images captured by a surveillance camera and it is observed that the skin color model is capable of processing 420/spl times/315 resolution images of 24-bit color at 30 frames per second in a dual Xeon 2.2 GHz CPU workstation running Windows 2000.