Aiming at the problem that collaborative sparse image Compressive Sensing (CS) reconstruction based on fixed transform bases can not adequately exploit the self similarity of images, an improved reconstruction algorithm combining the Total Variation (TV) with adaptive low-rank regularization was proposed in this paper. Firstly, the similar patches were found by using image block matching method and formed into nonlocal similar patch groups. Then, the weighted low-rank approximation for nonlocal similar patch groups was adopted to replace the 3D wavelet transform filtering used in collaborative sparse representation. Finally, the regularization term of combining the gradient sparsity with low-rank prior of nonlocal similarity patch groups was embedded to reconstruction model, which is solved by Alternating Direction Multiplier Method (ADMM) to obtain the reconstructed image. The experimental results show that, in comparison with the Collaborative Sparse Recovery (RCoS) algorithm, the proposed method can increase the Peak Signal-to-Noise Ratio (PSNR) of reconstructed images about 2 dB on average, and significantly improve the quality of reconstructed image with keeping texture details better for nonlocal self-similar structure is precisely described.