In this paper the Optimized Vector and Marginal Median Self-Organizing Map (OVMMSOM) was proposed as a new method of train Self-Organizing Maps (SOM). This variant is based on order statistics, Marginal Median SOM (MMSOM) and Vector Median SOM (VMSOM). This training model combines MMSOM and VMSOM defining their particular importance through a λ participation index. To demonstrate the effectiveness of the proposal, images from the COIL100 data set was clusterized and the Compose Density between and within clusters (CDbw) validity index was used. The performed experiments show that the proposed model outperforms standard SOM network trained in batch and even results from MMSOM and VMSOM by separately.