The traditional blind image deblurring algorithm based on the statistical prior models has the disadvantages of sensitivity to noise and limited detail recovery, while the learning-based image deblurring algorithm has poor adaptability for blurring kernel and noise level. To address the above problems, we propose a simple and effective low pixel sparse prior based on the statistical differences between the histograms of original and blurred images first. Then, in order to remove the noises and artifacts in restored image, a deep convolution neural network is designed to learn image denoising prior, which combines low pixel sparse prior and gradient sparse prior to form a new image deblurring model. Meanwhile, we estimate the blurring kernel in the structure layer so as to get a more accurate one, and the structure layer can be obtained by the image decomposition method. Numerous experimental results show that the proposed algorithm can restore more image details, and show more robustness to image type, blurring kernel type and noise level. The proposed method outperforms other recent state-of-the-art related approaches.