A Bayesian approach to object detection in color images

In this paper, we describe a supervised parametric Bayesian approach to object detection problem in color images. The proposed method makes use of multivariate Gaussian approximation to color distributions of different image regions. First, a set of sample points with known classification is collected from the object and background areas of an input image. This labeled training set is used for maximum likelihood (ML) estimates of the color mean vectors and covariance matrices for the regions of interest. A decision function is then defined as the a posteriori class conditional probabilities of the object and background regions. Using the Bayes rule and Gaussian models, a quadratic decision function is obtained corresponding to a hyperquadratic decision surface in the (R,G,B)-measurement space. The underlying surface divides the space into two mutually exclusive volumes of the object and background color clusters, resulting in a pixel classification strategy in the spatial plane. The experimental result provided in the paper illustrates that the method is quite effective on noisy textured images.