一种改进的基于样本块的图像修复算法

发布时间：2018-04-14 07:16

本文选题：图像修复 + 样本块　；参考：《山东大学》2017年硕士论文

【摘要】：随着时代的发展,人工智能等技术的推广越来越普遍,数字图像智能处理技术作为其中的一员,自然也得到普遍的运用。其中,数字图像修复(Digital Image Inpainting)又是数字图像处理学科中的一个重要分支方向,数字图像修复的过程是对一幅图画中的被污染或破坏的区域进行像素修补填充的过程,该技术的主要目标为复原被污染或损坏的数字图像,最终使得人们在正常的观看中察觉不到该图像有曾经被污损过的痕迹。无论是在数字照片处理应用场景,如损坏的数字人工制品壁画,历史照片修复等领域,还是在数字图像处理领域高层次的智能理解领域的研究过程中,如图像识别之前需要将图中对象删除或者对内容进行完整性修复,该技术都有非常广泛的需求。数字图像修复问题在数学领域上是一个不适定逆向的问题,该问题需要人为地设定图像的先验条件以定向地引导图像的自动修复过程,从而制造出一幅在感官上上没有遭受破坏、让人思维上可接受的处理图像。从而可知,怎样选择适用的数字图像先验条件模型就成了数字图像修复技术的关键条件。领域发展至今,修复技术根据破损污染区域的面积可以分为两种:处理面积较小的偏微分技术(Partial Differential Equations);处理面积较大的基于纹理重构的技术。Criminisi算法[19]是该论文的重点研究内容,该文中的算法是在Criminisi的基础上进行改进,对优先权的计算公式、寻找最优匹配块等内容进行了改进,并引入了平滑处理,用于克服在修复过程中产生的断层现象。在寻找优先权最大的样本块时,保持数据、自信度等不变,再加入梯度项,使得寻找到的优先权最大的样本块更加可靠可信,避免了信息量较少带来的缺陷;锁定最大优先权样本块后,论文在寻找最相似样本块的过程中引入了结构相似性信息,从整体上衡量了块之间的联系,而不是单单利用每个像素点之间的差异,使得匹配到的样本块与待修复块之间的可靠信息更多;通过研究过程中发现,基于样本块的图像修复算法经常会出现断层的现象,在视觉上很违和,因而,我们在所有的修复过程结束后又引入了 FMM平滑处理,使得最后的结果更加的理想。为验证算法的有效性,本文对不同类别的图像进行仿真实验,其中包含纹理复杂,结构丰富以及曲线多的图像,所得结果与Criminisi等算法的结果图进行比较,实验数据和实验效果图论证了该文中对Criminisi改进的算法的合理性。
[Abstract]:With the development of the times, the spread of artificial intelligence and other technologies is becoming more and more common. As one of the members, digital image intelligent processing technology has been widely used.Digital Image painting is also an important branch of digital image processing. The process of digital image restoration is the process of pixel repair and filling of contaminated or damaged areas in a picture.The main goal of the technique is to restore the contaminated or damaged digital image, which makes the image undetectable in the normal viewing.Whether in the field of digital photo processing applications, such as damaged digital artifact murals, historical photo restoration, or in the field of digital image processing in the field of high-level intelligent understanding of the research process,Such as the need to remove objects or repair the contents before image recognition, this technology has a very wide range of requirements.The problem of digital image restoration is an ill-posed reverse problem in the field of mathematics. This problem needs to set a priori condition of the image artificially to guide the automatic restoration process of the image.To create a sensual, ideologically acceptable processing image.Therefore, how to select a suitable prior condition model of digital image is the key condition of digital image restoration technology.Up to now, according to the area of damaged contaminated area, the restoration technology can be divided into two kinds: partial Differential equationsof small area, and texture reconstruction based on texture reconstruction. Criminisi algorithm [19] is the focus of this paper.In this paper, the algorithm is improved on the basis of Criminisi, the calculation formula of priority is improved, the optimal matching block is found, and the smoothing process is introduced to overcome the fault phenomenon in the repair process.When looking for the sample block with the largest priority, keeping the data, confidence and so on unchanged, adding the gradient term, the sample block with the largest priority is more reliable and avoids the defect caused by less information.After locking the maximum priority sample block, the paper introduces the structural similarity information in the process of searching for the most similar sample block, and measures the relationship between the blocks as a whole, instead of using the difference between each pixel point.Through the research, it is found that the image restoration algorithm based on the sample block often appears the phenomenon of fault, which is very contrary to the harmony in the vision, so,We introduced the FMM smoothing process after all the repair process, making the final result more ideal.In order to verify the validity of the algorithm, this paper carries out simulation experiments on different kinds of images, including images with complex texture, rich structure and many curves. The results obtained are compared with the results of Criminisi and other algorithms.Experimental data and experimental effect diagram demonstrate the rationality of the improved Criminisi algorithm in this paper.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】