Image inpainting plays an improtant role in application area. As an advanced signal processing method, sparse representation has been used in image inpainting. However given image bases are adopted in tradtional methods at the process of inpainting sparse images, it has no adaptive capability and sparse representation capability is limited. The best geometry direction is estimated according to reference images so that sparse transform is adaptive to geometry informations of images. And a representation method of sparser images is provided. Sparse inpainting is calculated by a minimized norm model such as l1. Experimental results show that edges and texture in images are preserved more perfectly comparing with traditional 2D wavelet transform methods. So higher peak signal to noise ratio is got.