Abstract: As the existing cell images segmentation algorithms are sensitive to noise and traditional SLIC(simple linear iterative clustering) algorithm can not divide boundary precisely, an improved SLIC method based on region merging is proposed to resolve the problems. First, the mean shift treatment is used to eliminate noise on the cervical cell images, then the two-dimensional otsu adaptive threshold processing is conducted to abtain the initial contour, then based on SLIC algorithm the superpixel region is obtained, and the superpixel regions are fused to the original image to complete the initial segmentation. Finally, in the initial segmentation map, the initial mark is used to obtain the mutual information, and the maximum similarity criterion is used to merge. In this way, no present threshold is needned. The non-marker background regions are merged with labeled automatically, while the non-marker object regions are identified and avoided from being merged with background. Several experiments of dividing the cytoplasm are conducted for cervical cells images. The proposed algorithm can extract cytoplasm from a single-cell cervical smear image more accurately in a relatively short time.

Key words: superpixel, image segmentation, initial contour, region merging, maximal similarity