The basic reason of low retrieval rate of contourlet transform texture image retrieval system lies in the coefficients shift sensitive characters and poor local time-frequency character. In order to overcome the defects of contourlet transform, a Contourlet-2.3 transform based on mapping is proposed, and the shift invariant character is proved. A texture image retrieval system based on the new transform is proposed in which the feature vectors are formed by cascading the energy and standard deviation of each sub-band in the transform domain, and Canberra distance is used as similarity metric. Experimental results show that the image retrieval system is superior to that of the original contourlet transform, Contourlet-2.3 and contourlet transform based on mapping with the same length of feature vectors, retrieval time and memory.