Abstract:

Blind debluring is a long-standing and challenging inverse problem. The key of restoring the high-quality image is the development of the image prior. So, in order to recover the original clear image from its blurred version, this paper develops a new and effective image prior：the dark pixel image prior, and proposes a dark-pixel-based blind deblurring method. The work is based on the inherent property of the blurred image：the dark pixel of blurred images is less sparse. In the blur process, the value of the dark pixel of the blurred image will increase when averaged with neighboring high-intensity pixels, making the dark pixel of blurred images less sparse. Therefore, making use of this less sparse characteristic, this paper can distinguish the blurred image and the clear image effectively, and make the blind image deblurring come true. However, the sparsity of the dark pixel introduces a non-convex non-linear optimization problem. In order to solve the proposed model effectively, this paper introduces a linear approximation of the minimum operator to solve the proposed model effectively. Extensive experiments indicate that in a comparison with several recent representative blind deblurring methods, the proposed method can get superior performance.

Key words: blurred image, blind deblurring, dark pixel, less sparse, linear approximation

摘要：

图像的盲去模糊问题是一个长期的且具有挑战性的逆问题。能否找到正确的图像先验是能否复原出高质量清晰图像的关键。因此，为了能够复原出高质量的清晰图像，找到了一种新的且有效的图像先验——图像中的暗像素先验，并提出了一种基于暗像素先验的模糊图像盲复原方法。该方法是基于模糊图像的内在本质特性所提出的，模糊图像中的暗像素是非稀疏的。在模糊过程中，清晰图像中的暗像素会因为与它周围的亮像素进行加权平衡，而导致模糊图像中暗像素的像素值增加，从而导致模糊图像中暗像素的稀疏性降低。因此，利用模糊图像中暗像素的这种非稀疏特性，能够有效区分模糊图像和清晰图像，从而实现模糊图像的盲复原。但是，基于暗像素的先验会导致一种非凸和非线性的最优化问题，为了能够有效地求解提出的模型，引入了一种最小化操作的线性近似来实现提出模型的最优化求解。大量的实验证明了该方法与近几年一些极具代表性的模糊图像盲复原方法相比，具有更好的性能。

关键词: 模糊图像, 盲去模糊, 暗像素, 非稀疏, 线性近似