Abstract: To further compress the 3D mesh data and improve the network transmission speed, meanwhile not reduce the model rendering visual appearance, this paper presents a texture compression algorithm based on 3D geometric information. Firstly, remesh the 3D mesh to get semi-regular topology structure. Secondly, extract the 3D feature points on the semi-regular mesh and the corresponding features in texture image are clustered with k-means method to form some regions of interest（ROI）. Finally, the encoding and decoding operations are realized on the texture image by combining the embedded zero-tree wavelet based on the ROIs. The experimental results show that the proposed algorithm can acquire more accurate ROI in texture image with high compression rate. After being decoded, reconstructed, and texture mapping, the texture mapped to 3D model has a better visual effect.