Ear biometrics has been found to be a good and reliable technique for human recognition. Due to significant advantages, ear biometric has gained momentum. In this direction, we propose a method to cluster ears based on similarity measures and to make use of this for quick retrieval of the image followed by the personal identification. This work involves elicitation of shape based features like distribution of planner area, moment of inertia with respect to minor and major axis and radius of gyration with respect to minor and major axis from ear images. We use four similarity measures for clustering 605 ear images. The method involves two phases, i. Determining centroids of predetermined groups by k-means clustering and ii. Using so obtained centroids to refine the clusters using similarity measures. From the computational experiments carried out on 605 ear images, it is revealed that Cosine, dice and Jaccard similarity were able to effectively cluster the image database into three groups. However, Overlapping similarity measures ended up only in two groups. The cluster analysis showed comparatively high values of entropy, purity, specificity, precision, recall and F-measure respectively for Jaccard, Dice and Cosine similarity function. The image retrieval rate became faster by by an average of 12.33% when database was organised in cluster groups when compared with image retrieval time with unorganised data followed by recognition accuracy of 92.5%.