Two Similarity Measure Methods Based on Human Vision Properties for Image Segmentation Based on Affinity Propagation Clustering

We firstly present an image segmentation method based on affinity propagation clustering which needs not to initialize cluster centers and is more reliable than traditional clustering methods such as K-Means clustering and so on. However, it is very difficult to get good image segmentation results through adjusting the only parameter ¿preference¿ of affinity propagation clustering, and sometimes the segmentation results don´t accord with human vision properties. To tackle the two problems, we propose two similarity measure methods based on human vision properties for measuring the similarities between pairs of data points of an image. The experiment results show that compared with the traditional Euclidean distance, the two similarities proposed can lower the level of difficulty of selecting parameters and make the segmentation results more according with human vision properties.