A new Content-Based Image Retrieval (CBIR) method based on the high-order moments of multi-scale and multi-direction Gabor filtering coefficients was presented and the algorithm for measuring the features of smoothness and consistency of texture which was extracted by a Gabor filter bank was given by using the variance of filtering coefficients. The feature vector for CBIR is composed of the values of the energy, variance, kurtosis, smoothness and consistency of filtering coefficients in 4 scales and 6 directions. The typical images from Brodatz texture database and Corel images were utilized in the contrast experiments, which show that the proposed method gives better precision than traditional Gabor filtering method.