【作者】 Rukundo Olivier；

【导师】 曹汉强；

【作者基本信息】 华中科技大学 ， 通信与信息系统， 2012， 博士

【摘要】 图像插值处理需要一幅参考图像或源图像以构造一幅新图像,其大小由设置的插值比例控制。该处理在许多领域有很重要的应用,包括卫星成像,医学成像,尤其在军用和民用电子产品领域。插值一幅数字图像,需要在源图像创建新空间并填入估计的像素值。如果估计技术精确度不高,则将得到一幅有可见痕迹且丢失细节的低分辨率图像。在超大图像数据库下,从低分辨率图像获得高分辨率是非常耗时的。但是,直接从传感器插值图像数据会导致细节不足,而在一些情况下,显示那些细节极为重要。而且,在多媒体系统中,用户浏览快速简洁,可以快速检索数据。论文中,提出了图像插值技术,并对算法在速度和PSNR方面作了改进。首先,提出了2种基于ACA的高分辨率图像插值算法。ACA最初由文献[118]提出,用于模拟电路缺陷的检测和标记,本文则用于处理图像插值问题。本论文提出的OBACA算法使用ACA确定有助于改善插值后输出图像分辨率的潜在像素。实验结果表明,本方案结果优于传统的双线性插值。本论文提出的第2个算法AACA,把图像像素的权重看作蚁群的信息素轨迹的函数。实验结果表明,AACA比OBACA算法要快。PSNR也比其他算法更优。第二,论文提出了3种图像插值算法,均基于高分辨率图像插值的最小绝对值差对应的像素值,减少了计算量。第一种算法SAD,对插值点的4邻域点选其一作了重新处理,计算该点的值和双线性插值的均值,并乘上一个控制因子k,k值由实验分析确定。实验表明,SAD有效且和最近邻插值,双线性插值和双三次插值相比具有更高的PSNR.对SAD改进后称为ASAD算法,直接令插值点的像素值为对应最小绝对值差的像素值。实验表明,ASAD在MET和PSNR上都优于SAD。最后,提出的NNV算法使用了模计算方法,实验表明,比SAD和ASAD算法获得了更优的结果。第三,论文提出了两种快速图像插值算法,一种基于勾股定理,另一种通过计算双线性加权平均,获得数字图像的高质量尺度变化,减少了像素组数量。第2种算法的理论分析表明PSNR和MET将优于传统的双线性插值(BI)算法,减少了BI加权平均操作次数,把直角三角形中的斜边和短边用图像像素绝对值差代替。两种算法和传统的插值算法相比,在处理速度和PSNR上都有更优的表现。最后,论文基于ACO提出了图像边缘细化方法并用于数字图像插值。图像边缘细化方法应用于线性插值后ACO边缘检测之后。细化过程建模类似于旅行商问题(TSP),在应用ACO找到最细的边缘前,根据某些规则集,并行的更改或移除不需要的像素信息。实验表明,基于插值的边缘细化能获得更高的分辨率/质量,且视觉上优于其他插值算法,但计算量较大。本论文提出的插值方法可以获得高分辨率,且插值快速。实验表明,这些方法可以提高PSNR和速度。下一步的工作可将之应用于彩色图像和实时低分辨率的图像插值。更多还原

【Abstract】 Image interpolation process requires a reference or source image to construct a new image whose size is controlled by the interpolation ratio selected or set. This process has been a problem of prime importance in many fields due to its wide application in satellite imagery, biomedical imaging, and particularly in military and consumer electronics domains. Interpolating a digital image amounts to creating empty spaces in the source image and filling in with the guessed pixel values. If the guessing technique is not of high accuracy, it would result in a low resolution image with visible artefacts and missing details. With huge image databases, creating high resolution from the low resolution image is time-consuming. However, image data interpreted directly from sensors often come up with insufficient details whereas in some cases it is of high importance to visualize sufficiently those image details. In addition, with the advances on multimedia systems, users can retrieve data quickly with a relative ease for neat browsing and rapid availability of such data.In this dissertation, I present the results of a 3 years research study on image interpolation techniques and algorithms-improving the performance in speed and PSNR.Firstly, I proposed two algorithms for high resolution image interpolation based on ACA. The idea to use the ACA, to deal with image interpolation problem, originated from the work presented in [118] about analog circuits faults detection and isolation. The OBACA algorithm used ACA to determine the potent pixels that would contribute in improving the output image resolution after interpolation. Experimentally, the proposed scheme yielded better results than conventional bilinear interpolation. The second AACA algorithm introduced the weights of image pixels as function of pheromone trails dropped by ants according to exploitation and exploration mechanism. Experimentally, AACA was found to be faster than OBACA algorithm. The PSNR also remained superior to all other algorithms compared with.Secondly, I proposed three image interpolation algorithms based on the pixel value corresponding to the smallest absolute difference for high resolution image interpolation using less computational efforts. The first proposed SAD algorithm was based on reprocessing one of the four pixels surrounding the unknown-value location and calculating the mean between that pixel and the value created by the conventional bilinear algorithm and by multiplying the mean the control factor k whose value was selected according to experimental analysis. Experimentally, SAD showed the effectiveness and higher performances only in terms of the PSNR when compared to the conventional nearest, bilinear and bicubic interpolation algorithms. The SAD improvement resulted in ASAD algorithm which replicated directly the pixel value corresponding to the smallest absolute difference to the unknown-value location. Experimentally, ASAD demonstrated higher in terms of MET and relatively PSNR than SAD. Finally, I proposed the NNV algorithm which achieved higher results than both SAD and ASAD using the mode calculation scheme. Experimentally, the results obtained were superior to those provided by SAD and ASAD.Thirdly, I proposed two speedy image interpolation algorithms, one based on the pythagorean theorem and another on reducing the number of pixel groups undergoing the bilinear weighted average operations have been introduced for high quality scaling of digital images. The latter showed that a simple theoretical analysis can lead to an improvement in PSNR and MET when compared to conventional bilinear (BI) interpolation by reducing the number of the BI weighted average operations while the pythagorean based algorithms replaced the hypotenuse and catheti with image pixels absolute difference. The two proposed algorithms demonstrated higher performances in terms of the processing speed as well as PSNR in some cases, when compared to the conventional interpolation algorithms mentioned.Finally, I proposed an image edge thinning scheme based on ACO and its application to digital image interpolation. The proposed edge thinning scheme was applied, after two processes, that is, the linear interpolation and the ACO edge detection have been completed. The thinning process was modeled like the Traveling Salesman Problem (TSP) before applying the ACO to find the thinnest ridge according to some rules set and the process of modifying/removing the unwanted pixels information was conducted in parallel. Experimentally, the edge thinning based interpolation demonstrated higher performances in terms of high resolution/quality and visibly outperformed other interpolation algorithms mentioned but at increased computational efforts. The methods presented in this dissertation permit to achieve High resolution (H.R) and either speedy image interpolation. Experiments showed that the proposed methods are able to increase the interpolation PSNR and speed. Future developments of the proposed methods may be adapted to color and real-time low resolution (L.R) images interpolation.更多还原