Adaptive Stochastic Resonance in Color Object Segmentation

Object segmentation is one of the most important tasks in image analysis and computer vision. One of its applications is to obtain an object location from the segmented image. Color thresholding provides a fast and simple scheme for object segmentation. But it is sensitive to lighting conditions and other noise effect in images taken from real-world environments. This paper shows that addition of a small amount of noise can improve the accuracy of such color object segmentation. It shows that this "stochastic resonance" or SR effect does occur for various performance indices that measure how well an object is segmented from the background. These measures include mutual information, error pixels count, and position error. Then the paper shows that a simple learning algorithm can learn the approximation of the optimal standard deviation of the additive Gaussian noise that we add to the system. The algorithm requires the size of an object but does not use the correct object location.