Most of existing two-dimensional compressed sensing and reconstruction methods for images are complemented by utilizing one-dimensional signals compressed sensing and reconstruction algorithms, which is inefficient and increases memory requirements. Two-dimensional random projection theory and two-dimensional projected gradient algorithm can overcome these disadvantages. However, fixed threshold parameter used in two-dimensional projected gradient algorithm for different images at different sampling rate may lead to poor reconstruction quality. Here, we propose an adaptive two-dimensional projected gradient algorithm based on image texture property. The parameter η of bivariate shrinkage is calculated according to image texture information during the iterative reconstruction process. Experimental results show that compared with two-dimensional projected gradient algorithm, the proposed adaptive twodimension projected gradient algorithm provides superior performance on both the image reconstruction quality and visual effect.