基于特征学习的图像超分辨率研究

发布时间：2018-02-15 06:02

本文关键词： 图像重建稀疏表示字典学习 K-SVD算法　出处：《山东师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着计算机以及互联网技术的日新月异的发展,网络上出现了大量的多媒体信息(如视频、图像、声音等),这些复杂多样的数据将人类带入到了计算机网络大数据时代。而这些信息数量庞大并且占用内存空间多又不便于存储和传输,如何从互联网上将这些繁冗的信息进行有效的检索、保存、挖掘其中蕴含的科技、商业和生活价值,成为如今需要人类急切处理的问题。人类是通过五官的感觉去获取信息的,而人类接收、感知周围环境事物的主要手段是视觉和语音信息。据最新研究统计,人类视觉所获得的信息占人类所有获取信息的接近百分之八十[1]。俗话说:眼睛是人类心灵的窗户,人眼看到的信息容量大、范围广。对于信息的采样及恢复,传统的Nyquist采样定理[2]给出了传统的框架,并确定了能够准确恢复原始信息的条件,即采样的频率应大于信号带宽的两倍以上。由于原理条件的限制及信息本身的特点,给信息的存储、传输等方面带来的困难在迅速增加。近年来,稀疏表示(Sparse representation)理论的出现引起了研究人员和学者的特殊关注。稀疏表示原理避免了传统的Nyquist采样原理的限制,将信号投影到特定的变换域,并根据信号在该域内的特有的稀疏特点及优化方法,达到恢复出原有的信号的目的。近几年来,机器学习、人工智能以及模式识别等技术引起人们的高度关注,本文在稀疏表示原理的基础之上,融合机器学习、人工智能等相关领域的知识,给出了一种基于特征学习的图像超分辨率重建算法。针对典型的特征提取算法提取的特征长度大,算法运行时所占用的空间多,导致算法运行时间长计算复杂度高的限制。在字典训练过程的特征提取阶段,本文通过提取图像的中频特征进行处理来减少运算时间,提高其运行的效率。并采用现今非常流行的效率较高的K-SVD方法进行字典的训练,并在字典训练之前采用PCA(Principal Component Analysis,PCA)[3]方法进行特征块的降维,来进一步降低算法的复杂。最后在重建阶段采用Batch-OMP算法进行稀疏稀疏的编码计算,提高了算法的精确度和重建图像的品质。
[Abstract]:With the rapid development of computer and Internet technology, a lot of multimedia information (such as video, image, etc.) appears on the network. Sound and so on, these complex and diverse data brought mankind into the era of computer network big data. And the amount of this information and the amount of memory space is not easy to store and transfer. How to effectively retrieve, preserve and excavate the technological, commercial and life values of these redundant information from the Internet has become a problem that human beings urgently need to deal with today. Human beings obtain information through the sense of five senses. And the main means for human beings to receive and perceive things in their surroundings are visual and phonetic information. According to the latest research statistics, The information obtained by human vision accounts for nearly 80% of all information obtained by human beings. As the saying goes: the eyes are the windows of the human mind, the information that the human eyes see has a large capacity and a wide range. The traditional Nyquist sampling theorem [2] gives the traditional frame, and determines the conditions under which the original information can be recovered accurately, that is, the sampling frequency should be more than twice the bandwidth of the signal. The difficulties in storage and transmission of information are increasing rapidly. In recent years, the emergence of sparse representation theory has attracted special attention of researchers and scholars. The principle of sparse representation avoids the limitation of traditional Nyquist sampling principle. The signal is projected to a specific transform domain, and the original signal is restored according to the special sparse characteristic and optimization method of the signal in this domain. In recent years, machine learning, Artificial intelligence and pattern recognition have attracted great attention. Based on sparse representation principle, this paper combines the knowledge of machine learning, artificial intelligence and other related fields. An image super-resolution reconstruction algorithm based on feature learning is presented. In the process of dictionary training, the if feature of the image is extracted and processed to reduce the computation time. The K-SVD method, which is very popular nowadays, is used to train dictionaries, and the PCA(Principal Component Analysis (PCA3) method is used to reduce the dimension of feature blocks before the dictionary training. In order to reduce the complexity of the algorithm, Batch-OMP algorithm is used for sparse and sparse coding in the reconstruction stage, which improves the accuracy of the algorithm and the quality of the reconstructed image.
【学位授予单位】：山东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】