梯度直方图约束的多尺度块先验模型

On multi-scale patch prior model with gradient histogram constraints

块先验模型在图像复原领域取得了较大的成功,但其整体模型强制局部性的缺点,易出现局部伪影、视觉观感较差的问题,提出一种新的集成多尺度块先验和梯度直方图先验的图像复原方法。对原始图像实施滤波和下采样以保持尺度不变性,在多尺度上施加同一局部块模型,即保持块低维模型的简单性,又在图像较大区域实施非局部性;将梯度直方图全局统计先验加入正则约束中,利用Wasserstein距离对复原图像与参考直方图的相似性进行度量。借助半二次分裂和最优传递理论,求解所提出的模型。通过在图像去噪和去模糊实验,相比传统方法无论在客观质量评价还是视觉观感上都更有优势,验证了方法的有效性。

Patch prior model has achieved great success in image restoration,but due to the enforced locality of the overall model,it tends to exhibit local artifacts and poor visual perception.A new image restoration method integrating multi-scale patch prior and histogram prior is proposed in this paper.The original image was filtered and down-sampled to maintain the scale invariance,and the same patch local model was applied on multiple-scale images,by which the simplicity of the low-dimensional patch model was maintained and the non-locality implemented in a large area of the image.The histogram global statistical prior was added to the regular constraint,and the similarity between the restored image and the reference histogram was measured by the Wasserstein distance.And the proposed model was solved by the theory of half quadratic splitting and optimal transfer.The effectiveness of the proposed model is verified by experiments of image denoising and deblurring in that our model has advantage over traditional methods in terms of both objective quality assessments and visual perception.