A novel method for segmenting bright objects from dark background for grayscale image is proposed. The concept of this method can be stated simply as: to pick out the local-thinnest bands on the grayscale grade-map. It turns out to be a threshold-based method with local adaptive thresholds, where each local threshold is determined by requiring the average normal-direction gradient on the object boundary to be local minimal. The method is highly automatic and the segmentation mimics a man's natural expectation even the object boundaries are fuzzy.