Among many features that can be observed from human skin, wrinkles are known to be very effective for assessing a subject's physical condition, surroundings, or lifestyle. In our previous work, we showed how to extract various wrinkle-related features such as total length, average width, depth, and size from magnified skin images and use them to estimate the degree of skin aging. To represent wrinkles on the skin images, we used a watershed algorithm and constructed its skeleton image, in which wrinkles are represented by 1-pixel lines. A skeleton image consists of polygons, which we call wrinkle cells. Since most wrinkle-related features are deduced from this skeleton image, accurate wrinkle representation is very critical. However, we found that the watershed algorithm produces over-segmentation for skin images; i.e., one wrinkle is represented by multiple smaller wrinkles in the skeleton image. To solve this problem, in this paper we propose an accurate skin wrinkle representation scheme that identifies and merges over-segmented cells in the skeleton image. Various experiments on our prototype system show that our scheme provides accurate skin wrinkle representation and thus improves the accuracy of skin age estimation.