基于视觉感知的超分辨率图像重建及其质量评价

发布时间：2018-12-10 22:07

【摘要】：在实际应用中,由于成像系统内在硬件设备的限制,高分辨率图像通常难以获得。单纯的通过改善硬件设备来提高图像的分辨率不仅代价高昂,而且面临着成像系统短期内难以克服的一些技术难题,因此以软件方式来提高图像的分辨率意义重大。超分辨率图像重建技术就是采用信号处理技术,从单幅或多幅低分辨率图像中重构出质量较好的高分辨率图像。本文主要研究了单幅低分辨率图像重建及其质量评价问题,在对现有方法改进的基础上,结合人类的视觉感知特性,得到与人类主观偏好一致的高分辨率图像和图像质量评价标准。人类视觉系统是一个高度复杂的智能信息处理系统,能够在极短的时间内完成图像内容的处理。因此,将人类视觉系统的视觉感知特性与计算机图像处理算法结合,能有效地提升后者的处理效率。在视觉信息感知初期,人类视觉系统并非对所有图像区域平等地进行处理,而是通过视觉关注机制筛选出感兴趣区域优先解读。基于视觉关注的图像显著性检测算法能有效减少待处理图像内容,从而提升图像处理效率。由于视觉系统的分辨能力有限,人眼无法察觉出处于一定阈值以下的信号内容变化。我们可以利用这一特点,去除对人类视觉系统无影响的变化信息,提升客观图像质量评价指标与主观评价的一致性。论文围绕基于视觉感知的超分辨率图像重建及其质量评价方法展开研究,主要内容与贡献包括:1、传统的插值算法容易引起边缘模糊,而人类视觉系统通常容易注意到图像中的物体边缘。传统的插值算法也无法很好的处理噪声,而由于视觉掩蔽效应,平坦区域的噪声更容易引起人们的注意。针对上述问题,提出一种基于边缘聚焦与自适应滤波的超分辨率重建方法。该方法首先采用传统的插值方法获得初始高分辨率图像,再通过边缘聚焦找出图像中的边缘像素,针对不同像素采用自适应滤波方法,自动调整滤波器参数,以获得最佳的滤波效果,获得最终的高分辨率重建图像。同时,我们还提出了一种快速的图像块搜索算法,用以加速滤波方法。实验表明,该方法无论在视觉效果还是客观评价上都获得较好的重建性能。2、人类视觉系统的视觉关注机制决定了人们在观察一副图像时,总是优先处理感兴趣区域的图像内容。而在计算资源有限,实时性较高的情况下,提升感兴趣区域的优先级,能够在保证感兴趣区域重建质量的情况下,加快超分辨率图像重建算法的速度。因此,我们提出了一种基于显著性模型的超分辨率图像重建框架,该框架具有较好的适应性和可扩展性。此外,传统的插值方法直接将低分辨率图像中的像素作为高分辨率图像中对应位置的像素处理,并没有考虑到实际的成像过程中模糊、下采样和噪声干扰等一系列降质因素的影响。针对这个问题,我们还提出了一种基于图像局部结构相似度的块插值方法。3、由于超分辨率重建图像与输入的参考低分辨率图像之间存在着图像大小不一致的问题,目前广泛使用的全参考图像质量评价标准并不适用于超分辨率重建图像的质量评价,而学者们对于超分辨率重建图像的质量评价也鲜有研究。针对上述问题,提出一种新的半参考的超分辨率图像质量评价标准。一方面,从结构相似度反映出重建图像与参考低分辨率图像之间一致性;另一方面,从边缘的显著度和模糊程度反映出人类视觉系统对重建图像质量的客观评价,将两者融合获得最终的评价指标。实验表明,该图像质量评价标准能够较好地适用于超分辨率重建图像的质量评价任务。上述研究成果从人类视觉感知的特性出发对现有的超分辨率图像重建和质量评价方法进行了补充和改进,具有一定的前瞻性和挑战性,具备一定的理论意义及实用价值。
[Abstract]:In practical applications, high-resolution images are often difficult to obtain due to the limitations of the hardware devices within the imaging system. It is not only expensive to improve the resolution of the image by improving the hardware equipment but also some technical problems which are difficult to be overcome in the short time of the imaging system. The super-resolution image reconstruction technique is to reconstruct a high-resolution image with better quality from one or more low-resolution images by using signal processing technology. In this paper, the reconstruction of single-frame low-resolution image and its quality evaluation are mainly studied. On the basis of the improvement of the existing method, the high-resolution image and the image quality evaluation standard which are consistent with the human's subjective preference are obtained. The human vision system is a highly complex intelligent information processing system, which can complete the processing of the image content in a very short time. Therefore, the visual perception characteristic of the human vision system is combined with the computer image processing algorithm, and the processing efficiency of the latter can be effectively improved. in that early stage of visual information perception, the human vision system is not proces equally to all the image regions, but rather the region of interest is selected by the visual attention mechanism. The image saliency detection algorithm based on visual attention can effectively reduce the image content to be processed, thereby improving the image processing efficiency. due to the limited resolution capability of the vision system, the human eye cannot detect the change of the signal content below a certain threshold. We can use this feature to remove the change information that has no effect on the human vision system, and improve the consistency of the objective image quality evaluation index and the subjective evaluation. This paper studies the reconstruction of super-resolution image based on visual perception and its quality evaluation method. The main content and contribution include: 1. The traditional interpolation algorithm is easy to cause the edge blur, and the human vision system is usually easy to notice the edge of the object in the image. The traditional interpolation algorithm can not process the noise well, and due to the visual masking effect, the noise of the flat region is more likely to cause the attention of the people. In view of the above problems, a super-resolution reconstruction method based on edge focusing and adaptive filtering is proposed. The method comprises the following steps of: firstly, obtaining an initial high-resolution image by adopting a traditional interpolation method, and finding the edge pixels in the image through the edge focus, adopting an adaptive filtering method for different pixels, and automatically adjusting the filter parameters to obtain the best filtering effect; a final high resolution reconstruction image is obtained. At the same time, we propose a fast image block search algorithm to speed up the filtering method. the experimental results show that the visual attention mechanism of the human vision system determines that the image content of the region of interest is always preferentially processed when a pair of images is observed. In the case of limited computing resources and high real-time performance, the priority of the region of interest can be improved, and the speed of the super-resolution image reconstruction algorithm can be accelerated under the condition of ensuring the reconstruction quality of the region of interest. Therefore, we propose a super-resolution image reconstruction framework based on the saliency model, which has better adaptability and expandability. In addition, the traditional interpolation method directly processes the pixels in the low-resolution image as the pixel of the corresponding position in the high-resolution image, and does not take into account the influence of a series of quality-reducing factors such as fuzzy, down-sampling and noise interference in the actual imaging process. in ord to solve that problem, we also propose a block interpolation method based on the similarity of the local structure of the image. The full-reference image quality evaluation standard which is widely used at present is not applicable to the quality evaluation of super-resolution reconstruction images, and scholars have little research on the quality evaluation of super-resolution reconstruction images. In view of the above problems, a new semi-reference super-resolution image quality evaluation criterion is proposed. On the one hand, the consistency between the reconstructed image and the reference low resolution image is reflected from the degree of similarity of the structure; on the other hand, the objective evaluation of the quality of the reconstructed image by the human vision system is reflected from the degree of saliency and the degree of blurring of the edge, and the final evaluation index is obtained by fusing the two. The experiment shows that the image quality evaluation standard can be well applied to the quality evaluation task of the super-resolution reconstruction image. The research results, from the characteristics of human visual perception, complement and improve the existing super-resolution image reconstruction and quality evaluation method, have a certain forward-looking and challenging, have certain theoretical and practical value.
【学位授予单位】：江西财经大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】