Fiber cross-sectioning often creates fiber clusters in microscopic images, in which fibers touch or overlap each other. Prior to any geometrical analysis, it is critical to separate touching/ overlapping fibers so that the features of individual fibers, not fiber clusters, can be identified. Automatic separation of irregular, complex fiber cross-sections remains challenging in image analysis for fiber characterization and measurements. This paper introduces an algorithm based on the image set theory to separate clustered fibers in cross- section images. An image is partitioned into three subsets, fiber edges, fiber interiors, and background. The Euclidean distances between edge pixels and interior pixels are used to assign the edge pixels to specific interiors. The assignment leads to the divisions among the merged edge pixels. The experimental results demonstrated that the new algorithm can optimally separate clustered fibers of various cross-sectional shapes, including W-shaped and cross-shaped fibers.