多参数梯度稀疏正则化图像非盲去模糊

Non-blind Image Deblur with Multi-parameter Gradient Sparse Regularization

自然图像在不同纹理区域具有不同的梯度特性,通过对图像梯度进行合理分层规划将图像纹理划分为5个区域,对各纹理区域梯度进行l_p范数约束,且每个区域对应1个合适的p指数值,建立多参数正则化模型,有效避免了全局单一p指数的缺陷.最后结合GISA稀疏编码框架,得到更加稳固的复原结果.通过实验对比,发现提出的多参数梯度稀疏正则化方法可以有效地提升图像纹理细节部分视觉质量,同时也能取得更高的信噪比.

Natural images have various gradient distributions in different textures, by taking advan- tage of this characteristics we partition image into five different texture regions. We use lp-norm to constraint the gradient of each texture region, and every region corresponds to a suitable p value. Such that we build the multi-parameter regularization model which overcomes the drawback in global image of single p value. Finally, combine with GISA sparse coding framework to get a more solid re- sult. By comparing the experimental result, it proves that our algorithIn can effectively improve the visual quality in details and achieve higher signal to noise ratio.