Abstract: Feature vector construction of contourlet texture image retrieval algorithm generally was generated by cascading energy and variance of each sub-band coefficients, hence preventing the exact expression of detailed texture structure. To alleviate the defect of this reason, relative variances was constructed as a new descriptor for texture characters, and further combine the descriptor with the magnitude of sub-band coefficients and kurtosis to construct new feature vector of each texture image. Using this kind of novel feature vector, in the framework of contourlet and contourlet-2.3 texture image retrieval systems, experiments on texture images from Brodatz showed that the new algorithm is superior to the commonly used feature vector, average retrieval rates can be improved about 5%~7%