Geometry image is a kind of normalized model representation method to present three-dimensional geometric model. Because of its array representation, geometry image has become one of the important tools to effectively deal with complex three-dimensional models. Unfortunately, part of the input model is stretched greatly in the parameterization process of geometry images generation. And the regular resampling fails to capture the features on the surface which will result in reconstruction error. To solve this problem, the paper proposes a new geometry image method based on area stretching. We find the area with larger distortion by calculating the area stretching metric for each triangle, and determine the region to be processed. After dealing with the mesh locally, we can get a more accurate resampling result, which will produce higher quality reconstructions. Experiments show that, our method produces better quality of the reconstruction models and achieves obviously PSNR gain.