Segmentation can be facilitated by knowledge of vision task. In this paper, an approach of segmentation is proposed, which is guided by detection of salient features of the target, and makes the inter-classes variance based thresholding be effective. Rules for local segmentation are defined under consideration of the different distribution of the bulk and engine in IR ship images. The approach is applied to IR ship segmentation successfully, and can be adopted in real-time IR ship detection and recognition.